YT YouTube2Text

YouTube2Text → Transcripts → How to Build & Sell AI Agents: Ultimate Beginner’s Guide

How to Build & Sell AI Agents: Ultimate Beginner’s Guide — Transcript

by Liam Ottley · 50,742 words · 7,072 segments · language en · Watch on YouTube

Open in transcriber Summarize with AI Download SRT/VTT

Full transcript

0:00Two years ago, I taught myself how to
0:01build AI agents without any prior
0:03experience in AI. And since then, I've
0:05started multiple AI businesses that have
0:07generated over $5 million in revenue.
0:10I've grown from 0 to 450,000 subscribers
0:12here on YouTube. And I've built AI
0:15agents for some of the biggest companies
0:16in the world. It's pretty safe to say
0:18that learning how to build AI agents has
0:20completely changed my life. So, in this
0:22full course, I'm going to teach you
0:23everything, and I mean everything that I
0:25have learned over the past 2 years about
0:27building and more importantly making
0:29money with AI agents, even if you don't
0:31know how to code. My hope is that you
0:32too can learn and use this incredibly
0:34powerful skill in order to build the
0:36life of your dreams before these AI
0:38agents start taking our jobs. As you can
0:40see, by the length of this video, I'm
0:41not going to be holding anything back.
0:43So, in order to make it a bit easier to
0:44consume, we're going to be splitting it
0:46into three different chapters. First,
0:47we'll build your foundational
0:49understanding of AI agents, covering
0:51what they are, how they work under the
0:52hood, and the key concepts that you need
0:54to know before you actually start
0:56building them. And this is without any
0:58technical background required. Secondly,
0:59we will be diving into four different
1:01endto-end AI agent tutorials, taking you
1:04over my shoulder every step of the way
1:06as we build some of the most popular AI
1:07agent use cases on the market today. And
1:10I have carefully planned these builds
1:11out to give you a taste of multiple
1:12different noode platforms and different
1:14AI agent types. and you'll learn how to
1:16build each of these step by step
1:18watching over my shoulder. And finally,
1:19I will give you my proven blueprint for
1:21monetizing your AI agent building skills
1:23over the coming years while this
1:25technology continues to explode in
1:26adoption and popularity. I'll be sharing
1:28the exact strategies that I've used to
1:30generate millions of dollars with this
1:31skill set over the past 2 years. So,
1:33let's get into it. Now, if you're new to
1:35the channel, let me quickly share why
1:36I'm qualified to teach you about AI
1:38agents. So, if you are new to the
1:40channel and don't know who I am, my name
1:41is Liam Mley and 2 years ago, I started
1:43learning about AI with no prior
1:45experience in the field. I was teaching
1:46myself how to build AI agents and chat
1:48bots all through my own self-study,
1:50which I documented right here on this
1:52YouTube channel from day one. So, you
1:53can go back and watch all my previous
1:54videos of how I got from there to here.
1:56This led me to starting Morningside AI,
1:58which is my AI automation agency where
2:00we build AI systems and agents for
2:02businesses from basic customer support
2:04systems to full AI SAS platforms and
2:06even we've built my own AI agent SAS
2:08platform, Agent, which is now over
2:1045,000 users. At Morningside, we've
2:12worked with publicly traded companies
2:13and even recently an MBA team. I also
2:15run the world's largest AI business
2:17community with over 120,000 members on
2:19school as of this recording. and through
2:21my community and through this YouTube
2:23channel here, I've taught hundreds of
2:24thousands of people from all backgrounds
2:26really how to build and make money from
2:28AI agents. So, everything that I'm about
2:30to teach you today is exactly what
2:32helped me to achieve all of this. So,
2:34let's dive
2:39in. Now, if you look at how long this
2:41video is, you'll realize that there is a
2:42lot to cover here. And now, I don't want
2:44you to give up halfway through the
2:46video. So, let's quickly take a moment
2:47just to get clear on why AI agents are
2:49quite literally the next big thing. And
2:51I know that sounds cliche and you hear
2:52it all the time, but seriously, they
2:54are. And why learning to build them is
2:57by far one of the most valuable skills
2:58that anyone can have over the coming
3:00decades. And if if my experience over
3:02the past 2 years is anything to go by,
3:04that should be enough proof for you guys
3:05to believe me. So, stick with me. But
3:07here's the hard truth about AI and jobs
3:08right now. According to the latest
3:10research, McKenzie predicts that AI and
3:12agents could automate up to 50% of
3:14current work by 2030. And the World
3:16Economic Forum reports that 41% of
3:18companies plan to reduce their workforce
3:20due to AI. Now, a lot of this sounds
3:22doom and gloom, and of course, many
3:23people are naturally worried about their
3:25career in future based off seeing this
3:26kind of data. But it's not all bad if
3:29you know where to look. And this video
3:30isn't about making you feel all sad.
3:32It's about uplifting you. And if you
3:33look on the flip side of the same data,
3:35these same reports reveal an enormous
3:37opportunity for those willing to seize
3:38it. So the World Economic Forum's future
3:40of job report states that 50% of
3:42employees plan to reorient their
3:44business in response to artificial
3:46intelligence. And due to this
3:47reorientation, 66% of employees plan to
3:50hire talent with specific AI skills such
3:53as prompt engineering. So on one hand,
3:55we have the expectation of massive
3:57layoffs and automation of work over the
3:58next 5 to 10 years. But on the other, we
4:01have the majority of employers searching
4:03for people who have these AI skills or
4:05really just basic AI literacy. Why?
4:08Because AI literate employees who can
4:10automate parts of their work can have 5
4:12to 10x the output of someone who doesn't
4:14have any AI literacy. And I promise you
4:15that brushing up on your AI and reaching
4:17this point of AI literacy so that you
4:19can be on the winning side of this next
4:215 to 10 years is actually so much easier
4:23than you think. I mean, it's easy as
4:25watching this entire video in order to
4:26build your AI skills base to make a big
4:28step towards AI literacy. And if you
4:30don't believe me when I say that a
4:31little bit of self-study like this video
4:33goes a long way, here's a great clip
4:35that I've seen recently on the All-In
4:37podcast from one of the most respected
4:38investors and technologists in the
4:40world, Navar Raakan, alongside a whole
4:42bunch of other very smart billionaires.
4:44Again, I would say the easiest way to
4:46see that AI is not taking jobs or
4:48creating opportunities is go brush up on
4:50your AI. Learn a little bit, watch a few
4:52videos, use the AI, tinker with it, and
4:55then go reapply for that job that
4:57rejected you and watch how they pull you
4:59in. And so this video is exactly what
5:01Naval is talking about. So whether
5:02you're an aspiring entrepreneur wanting
5:03to learn valuable AI skills and launch
5:05an AI business like I have, or you're a
5:07business owner wanting to just
5:09understand agents so you can use them to
5:11grow your business, or maybe you're just
5:13wanting to make sure that you are the
5:14last person that your boss thinks of
5:16firing because he or she's an AI wiz and
5:18I can't afford to lose them. Then I have
5:20made this video for you guys. Now, what
5:22I want you to do is close out all of
5:23your other tabs. Go get a notebook and a
5:25pen and a beverage of choice and make
5:27sure you make a commitment to yourself
5:29right now in order to finish this
5:31training and ensure that you are going
5:32to be empowered by AI and not replaced
5:35by it. Now, if you've done all that,
5:36let's get stuck into
5:39it. All right, so step one in building
5:41AI agents is knowing what the hell an
5:43agent actually is. So, 2 years ago, when
5:45I first started learning about AI
5:47agents, I had no idea what they actually
5:49were. The term AI agent gets thrown
5:51around a lot in almost like everywhere
5:52these days. You got AI agents this, AI
5:54agents that. But what actually is an AI
5:56agent? Well, the clearest definition
5:58that I found that helps beginners to
5:59really wrap their head around what they
6:01are is this. An AI agent is a digital
6:03worker that can understand instructions
6:06and take actions in order to complete
6:08tasks. So, in a very simple way, just
6:10like businesses have employees who
6:12handle different tasks, an AI agent is
6:14like having a digital employee. But the
6:17cool thing is that you can build them
6:18and you can make them do whatever you
6:20want. You're like literally building an
6:22employee that you can put to work to do
6:23things for you. And of course, they cost
6:25much less to run than a human and they
6:27don't need sick days and they don't
6:29start beef with Mike over in the sales
6:30department because of his comment at the
6:32coffee machine. So, I'm sure you can see
6:34the appeal of this kind of digital work
6:36and AI agents to businesses who are
6:38looking to adopt them.
6:41In order to really understand why these
6:43AI agents are such a big deal, we need
6:45to look at where we are coming from. So
6:47most of you have probably encountered
6:48those chat bots on websites before. You
6:50know those little little chat widgets
6:52that pop up saying like, "Hey, how can I
6:54help you?" So these kinds of chat bots
6:55are pretty basic, right? They a lot of
6:57the time they're useless and they're
6:58they're kind of like a waiter who can
7:00only really recite the menu but can't
7:02actually take your order or or bring
7:03your food. They can't do anything. They
7:05just respond with some kind of
7:06pre-written answers. Well, nowadays it's
7:08a simple AI generated answer. But AI
7:10agents are different, right? So, here's
7:13an example. If you ask a regular chatbot
7:15about booking an appointment, it might
7:17say, "Oh, our business hours are 9 to5.
7:19Please call to book." And that's it.
7:21They just give you some information
7:22back. But with an AI agent, it could
7:25actually go and check the calendar, find
7:27some available slots, go back and forth
7:29with the person that they're chatting to
7:30in order to book an appointment, send
7:32you a confirmation email, then update
7:34the business's scheduling system and CRM
7:36automatically in seconds. This ability
7:39to take action is what makes agents so
7:41powerful. They're not just fancy chat
7:42bots. They're actually digital workers
7:44who can search through databases, update
7:47spreadsheets, send emails, book
7:49appointments, generate hold documents,
7:52and much much more. And so building and
7:54deploying an AI agent is a bit like
7:56hiring a new employee because when you
7:58bring someone into a business, you need
8:00to firstly explain their roles and the
8:02responsibilities to them. You need to
8:04give them access to your system so they
8:06can use them. And you need to trust them
8:08to handle those tasks independently. And
8:10now when we are building agents, as we
8:11see later, it's exactly the same, except
8:14these agents are going to be working
8:1524/7. They're never going to get tired.
8:17They can be duplicated and modified
8:19instantly. And they cost a fraction of
8:21what a human employee does. And this is
8:23exactly why understanding how to build
8:25and sell AI agents is becoming such a
8:27crucial and valuable skill these days.
8:29Because whether you're an entrepreneur
8:31looking to scale your business or you're
8:33an employee wanting to become
8:34irreplaceable and and make more money at
8:36work, knowing how to create and deploy
8:38these digital workers is like the
8:41biggest cheat code in the whole world
8:42right
8:44now. Now that you understand what AI
8:46agents actually are, let's look under
8:48the hood and see how they actually work.
8:50Just like humans need a brain, memory,
8:52and tools in order to do their job, AI
8:55agents need specific components in order
8:57to function correctly. An AI agent needs
8:59five key parts in order to work.
9:01Firstly, every AI agent needs a brain.
9:03In the AI world, we call this a large
9:05language model or an LLM for short. And
9:07you've probably heard of some of these.
9:08You've got GPT from OpenAI, Claude from
9:11Anthropic, Gemini from Google, etc. You
9:13can think of the LLM as having a super
9:15smart intern who can understand your
9:17instructions in plain English, and then
9:19figure out how to get things done from
9:20those instructions. So, without this
9:22brain, all of the other parts would be
9:23useless, right? It's like having a whole
9:25desk full of office supplies but having
9:27no one sitting there in order to use
9:28them. Secondly, the brain needs
9:30instructions on how to behave. And this
9:32is prompting. So writing a prompt for an
9:34agent is how you program a lot of the
9:36behavior of it rather than having to
9:37code it manually. And this is really
9:39what makes building AI agents so much
9:41more accessible to non-coders as the way
9:43of actually programming the
9:44functionality and how they work is done
9:46through clearly written instructions
9:47rather than having to actually code it.
9:49Thirdly, agents need memory. Imagine
9:52trying to have a conversation with
9:53someone who forgets everything you said
9:5430 seconds ago, right? So, memory is
9:56really important because it allows your
9:58agent to remember what you talked about
10:00just a few messages ago, keep track of
10:02the tasks that it's been working on,
10:04build on previous conversations, and
10:06even in more advanced ones, it can learn
10:08from your past interactions. And the
10:09good news about memory is that most AI
10:11agent platforms completely handle this
10:13memory component automatically. So, you
10:14don't need to worry too much about it.
10:16But just know that it is an important
10:17part of a functioning AI agent. The
10:19fourth component of an agent, and this
10:20one is optional, but it is external
10:22knowledge. AI models like GPT and Gemini
10:25are pre-trained on a huge amount of
10:27data, but that data is basically cut off
10:29at a certain point, eg 2024. It's kind
10:32of like having a new employee who only
10:34really knows what they learned in
10:35school. But just like you can train an
10:37employee like that with your company's
10:39specific materials, you can also give an
10:41AI agent additional knowledge on top of
10:44the information it was trained on
10:45through providing things like PDFs of
10:47your company documents, spreadsheets
10:48with product information, customer
10:50service transcript, or basically any
10:52other textbased information. Without
10:54this added knowledge, agents will be
10:56limited to general information and
10:57couldn't handle specific business tasks.
10:59But as I said, knowledge is optional and
11:01you will only need it in some builds.
11:02Finally, and this is the most important
11:04part, we have tools. So tools are what
11:07transform an AI agent from just being
11:08able to chat to being able to actually
11:11get things done. So you can think of
11:12tools like giving your digital employee
11:14access to different softwares. Just like
11:16you might give a new hire access to your
11:18email or your calendar or your CRM
11:20system, you can give an AI agent access
11:22to digital tools that let it take
11:24actions when needed. These tools let
11:26your agent do things like checking
11:27real-time data, updating databases,
11:29sending messages and notifications,
11:30creating documents, all the stuff we
11:32went over just before and much, much
11:34more. The really powerful part, which
11:35we're going to cover later, is when
11:36agents use multiple tools together in
11:38order to solve complex problems, just
11:41like us humans would use multiple
11:42different websites and softwares when
11:44doing our tasks. Now, let me show you
11:46how all of these parts work together in
11:47a real example. So, say you want an
11:50agent to handle customer support. When
11:52the agent is sent a message, the brain
11:54immediately understands the prompt that
11:55it has been given and also understands
11:57what the customer is asking. It checks
11:59its recent memory before replying each
12:01time to understand the full context of
12:03their conversation. And if the brand
12:04detects that the customer wants a
12:06specific question answered from the
12:08knowledge base, it will use its external
12:10knowledge in order to deliver the right
12:12information to them. And finally, it may
12:14use tools to update a customer's account
12:16or to process a refund whenever required
12:18during the conversation. So all of these
12:20things are happening in seconds as the
12:21conversation is going on. Which is why
12:23AI agents are such a game changer. They
12:25can combine all of these components in
12:27order to create a fully capable digital
12:29worker that very very closely replicates
12:31how humans
12:34work. Now that you know the anatomy of
12:36an agent and the five parts of it, a
12:38more practical framework for
12:40understanding how we actually plan and
12:41build AI agents is what I call the three
12:44ingredients. Basically, you only have
12:45three elements to plan when creating an
12:47AI agent, which when mixed in various
12:49ways can create millions and different
12:51types of agents for different use cases.
12:53This is because the AI model or brain
12:55can be easily swapped in and out and
12:57isn't really a major factor in the
12:59performance of the agent as any of the
13:01top models that you pick from any of the
13:02different providers at any given time,
13:04they're all pretty good. And also, the
13:05recent chat memory is handled by default
13:07in almost all cases when you're building
13:08on these platforms that you're going to
13:09see later. What this leaves us with is
13:11what really matters when building and
13:12planning AI agents. Firstly, the
13:14knowledge, the external data that you
13:15want the agent to be able to use when
13:17answering. Secondly, the tools, the
13:19different actions that you want the
13:20agent to be able to take, eg saving the
13:22contact info to the CRM or getting some
13:24live data on stocks or sending an email.
13:27And then finally, prompting, which is
13:29the glue that ties everything together
13:31and determines how the agent behaves.
13:33So, write these down. While the agent
13:35has five components, the brain, the
13:36prompt, the memory, the knowledge, and
13:38tools, your main focus as an AI agent
13:40builder is in the three ingredients of
13:42prompting, knowledge, and tools. In the
13:44next chapter, we'll be looking at how
13:45you actually build an agent using the
13:47different combination of these three
13:48ingredients. But first, we need to dive
13:50deeper into the keystone of
13:52understanding how to build your own
13:53valuable digital workers. And it all
13:55comes down to
13:58tools. Now, we need to dig a lot deeper
14:00on tools as they are by far the most
14:02powerful part of AI. agents. But in
14:05order to understand deeply and be able
14:06to build powerful agents with them, we
14:08need to take a few steps back and
14:10actually cover the basics of how
14:12software and the web and internet as a
14:14whole works. Now, this is as techy as
14:16it's going to get in this video, but I
14:17promise once you understand this, it's
14:19so important and it's literally like
14:20having a superpower. So, please stick
14:22with me through this. So, remember how
14:24we said that tools are what allow agents
14:25to take an action to actually do things
14:28rather than just chat? Well, the way
14:30agents use tools and do work online is
14:33just how we do it as well, but with one
14:35key difference. Instead of clicking
14:37buttons and typing into forms, agents
14:39use what we call APIs. And every time
14:42you use the internet, you're actually
14:43making dozens of requests to APIs as
14:46well and getting responses back, but you
14:48just don't realize it. So, let me show
14:49you what I mean. So, when you click on
14:51this video, here's what actually
14:53happened. Firstly, your browser sent a
14:56request to YouTube servers saying, "Hey,
14:59I want to watch this video." And then
15:00YouTube servers sent back all of the
15:02data needed. And thirdly, your browser
15:05unpacked that data and started playing
15:07the video on your screen. So this
15:08request and response pattern happens
15:10with almost everything that you do
15:11online. When you open up Instagram, you
15:13are requesting your feed from Instagram
15:15service. When you send a tweet, you are
15:17sending your data through Twitter's
15:19service. And when you check your email,
15:20you are requesting from Google the
15:22latest messages in your inbox and
15:24they're sending it back and your browser
15:25is loading it. Thankfully, we get pretty
15:27websites and apps that make it very easy
15:29for us to do this and use software via
15:31APIs through a nice application. But
15:33under the hood, it is still two
15:34computers talking back and forth,
15:36requesting, sending, and displaying new
15:38information for us on our screen. These
15:40request and response happen through what
15:41we call APIs, which are application
15:43programming interfaces. So, you can
15:45think of APIs like waiters in a
15:47restaurant. Basically, they're going to
15:48take your order or your request to the
15:50kitchen, which are the servers of the
15:52business, and then they bring back your
15:54food, which is the response. So, you
15:56have request and response, and you have
15:57you as the client, and them as the
15:59server. There are two main types of
16:01requests that you can make. Firstly,
16:02either a get request. This is basically
16:05just like asking for information like
16:07checking the weather or looking up the
16:08price or loading this video. You're
16:10requesting to get the information to do
16:12something. Secondly, we have post
16:14requests, which is when you're sending
16:15some kind of information like posting a
16:17tweet, sending an email, or uploading a
16:20photo. So, go back and write both those
16:21down because we're going to be using
16:22them extensively in the building section
16:24of this video. Now, here's where it gets
16:26interesting. So, AI agents use these
16:28same APIs as their buttons to do things.
16:30So, each tool an agent has access to use
16:32is essentially an API that it is able to
16:35call. So, these kinds of tools come in
16:36two different flavors. We have pre-made
16:38integrations like Google Calendar or
16:40Gmail where it kind of comes out of the
16:42box ready for you to use and just plug
16:43straight into your agent. And then we
16:45have custommade tools that we can build
16:47ourselves. So you can think of pre-made
16:49integrations like buying a readymade
16:50meal where they've done a lot of the
16:52hard work versus custom tools where we
16:54are like cooking from scratch. And both
16:56work, but custom tools give you a lot
16:58more control. And this is a skill that
17:00I'm going to be teaching you in the
17:01second chapter of this
17:04video. Okay. So now you got the basics.
17:06Let's get clear on how a tool is
17:07actually made and what the key parts are
17:09as you're going to be using them a lot.
17:11So, let's break this down using a simple
17:13example of a text capitalization tool.
17:15It takes in some text and the outputs
17:17the capitalized version of it. So, first
17:19to create a tool, we need a function. We
17:21need something that does work. In this
17:23case, it's super simple, right? It needs
17:24to take in text and it needs to make it
17:26uppercase. So, this can either be done
17:28through a basic Python function or you
17:30can use an LLM to do this as well.
17:31Basically, we need to build some way to
17:33capitalize the text that we give to this
17:35function and actually do the do the
17:36work. Next, in order for the AI agent to
17:39use this function, we need to wrap it in
17:41an API. So, we have the function and
17:42then the API wraps around it. And this
17:44is essentially making that functionality
17:46we created accessible over the internet
17:48via APIs. Without it, the function
17:51cannot be used by our agent. And in
17:53order to use the API that we've just
17:54created and use this function inside it,
17:56the API is going to expect the same sort
17:59of inputs that the function needs. So
18:01the input of the text that we want to
18:02capitalize and it's going to output the
18:04capitalized version. So this is very
18:06important to remember function. It takes
18:08in the input of the uncized text does
18:10work and outputs the capitalized
18:12version. We're basically then just
18:14building an API around it so that we can
18:16put it on the internet and then we can
18:18have an agent that knows how to call
18:19that API can send information into the
18:21input go through the function and then
18:23get spit out and then our agent catches
18:25it at the end. But the magic step and
18:27what has really caused the AI revolution
18:28to kick off is that we can explain to
18:31our agent how to use this API just by
18:34explaining how the API works in natural
18:36language. And this is where schemas come
18:40in. A schema is like a one-page
18:42instruction manual on how to use an API
18:45and therefore how to access the
18:47functionality inside that API. And when
18:48an AI agent is given one of these
18:50schemas, it too can read that
18:52instruction manual and determine things
18:54like what the tool does, what
18:56information it needs as an input, like
18:58we talked about before, and what
19:00information to expect as an output. Now,
19:02they may look scary, but they're
19:03actually really, really easy to
19:04understand, and we're going to cover
19:05them in the next chapter with this
19:06video. And the good part about it is
19:08that these days, schemas are
19:09automatically created by many of these
19:11no code platforms that you build agents
19:12on. But I'm teaching you this because it
19:14still helps to know what they are doing
19:16and what that what's really happening
19:17under the hood on these platforms. And
19:19there are still going to be times where
19:20you may need to roll up your sleeves and
19:21do it yourself. The incredible part
19:22about these schemas is that modern AI
19:24like chatpt can read these instructions
19:26and perfectly understand not just how to
19:28use it and like okay I need an input and
19:30then I expect an output but also when to
19:32use it. For example, let's say we had an
19:34agent and we gave it that capitalization
19:36tool that we just talked about and then
19:37we said can you please capitalize this
19:39text? Mary had a little lamb. The agent
19:41would then read over the schemas that we
19:43provided it and then it would see that
19:44there's a tool with a description saying
19:46this tool capitalizes text right in the
19:48instructions for the capitalization
19:50tool. We would have said this thing is
19:52for capitalizing text and it takes in
19:54some text and it gives you the
19:55capitalized version. And so the agent
19:56will read that and see okay this looks
19:59like based off the instruction they just
20:00gave this is the tool that they want to
20:02use. And then it will check the
20:03requirements and see that the tool takes
20:05in one input in string format which is
20:07just text which we have described as the
20:10text to be capitalized. So it reads all
20:11this. He says okay it it needs one
20:13input. It's in string format. So I know
20:15I need to give it some text and okay
20:17what does this text do? It's the text
20:18that they want to capitalize. Great. So
20:20now it knows it needs the input and it
20:22knows that this is where it's going to
20:23send the text to be capitalized. Then
20:24now that it knows what it wants, it goes
20:26back to our message and it intelligently
20:28extracts Mary had a little lamp. not,
20:30hey, can you please capitalize Mary had
20:32a little lamb? It's smart enough to know
20:33that we want that taken out. So, it will
20:35take that part, Mary had a little lamb,
20:37out of our input, and then it sends that
20:39to the API where our capitalization
20:42function does its thing. Then the API
20:43sends back the capitalized version plus
20:46a bunch of other response data as well.
20:48Then the agent looks at your original
20:50question, looks at this messy response
20:52it got back from the API, and then using
20:54its brain, the LLM, it writes a natural
20:56language response answering your
20:58question. It would say, "Here's your
20:59capitalized text colon Mary had a little
21:02lamb in all caps." That may sound
21:03complicated. It may have gone over your
21:04head. Please go back and just listen to
21:06it again. You really, really need to
21:08understand this process of uh the
21:10message comes in, looks at the schema,
21:11realizes, okay, it wants to use this
21:13tool. Okay. What do I need to do in
21:14order to use this tool? Okay. Well, then
21:16I'm going to grab it out of the input.
21:18I'm going to put it in here. And it can
21:19actually go back and forth. Say our
21:21capitalization tool needed some other
21:22input. Say you needed to provide uh the
21:25number of letters you wanted to be
21:26capitalized. It may see that this tool
21:28needs two inputs and I've only been
21:30given one. So then it will go back and
21:32ask me, hey, could you can you please
21:34tell me how many letters you want to be
21:35capitalized and you will see this magic
21:37in the agents that we're going to build.
21:39When the agent can ask you questions in
21:40order to help fulfill the needs of the
21:42tool, you have this very intelligent
21:43system that really will blow you away
21:45when you see it in action. And one thing
21:46many people miss about this process is
21:48the agent actually gets back raw
21:50computer data from the API or what we
21:52call JSON. But using the LLM, it can
21:55transform that into natural conversation
21:56and answer your question in a very very
21:59uh clear and concise way. So it's
22:00basically like having an employee who
22:02can read all this technical information
22:04and then explain it to you in plain
22:05English, which is another part of why AI
22:07agents are so powerful. And so when you
22:09understand this pattern that we've just
22:10gone through, I promise you, you will
22:11never see the internet the same way
22:13again. Every action online is just
22:15requests and responses. And therefore,
22:18we can build our own tools and AI agents
22:20to automate all of it. So instead of you
22:22manually searching the web, copying
22:24information, pasting it into
22:26spreadsheets, sending emails, an AI
22:28agent can do it all automatically using
22:29tools if you build it correctly. It's
22:31like having a digital employee who can
22:33press all of these API buttons for you
22:36thousands of times faster than any human
22:38could. And don't worry if this feels a
22:39little bit technical. In the next
22:41chapter, uh I'm going to show you how to
22:42create your own tools like this from
22:44scratch using platforms like Relevance
22:46AI, uh where you can build out powerful
22:47tools without writing any code. and will
22:49really start to click into place once
22:50you see the stuff in action in the
22:52building section. But before we get into
22:53that, let me reveal the power of AI
22:55agents which is unleashed when they are
22:57given multiple tools to work
23:00with. Now, obviously having an AI agent
23:02that just capitalizes text isn't very
23:04useful. I get that. The real magic
23:06happens when you give agents multiple
23:08tools and the ability to use them
23:09together in order to achieve complex
23:11goals. So, do you remember our
23:13definition? AI agents are workers that
23:15can understand instructions and take
23:16actions to complete tasks. When you give
23:19an AI agent a task, it's going to try
23:21its best to execute on it, but if it
23:23doesn't have the right tools on hand to
23:24do the job, it's going to be useless.
23:26And so, the more tools that you can give
23:28an agent, the more flexibility it has to
23:30solve problems just like a human would.
23:32So, let me give you a real example from
23:33my own business, right? Say I build an
23:35agent and give it the task. Find AI
23:37startups that have recently raised money
23:39and put them in a spreadsheet and add a
23:41summary of each of the businesses in the
23:42spreadsheet and then email me the link
23:44to the spreadsheet. When you give an AI
23:45agent a task like this and provide it
23:47with multiple tools to use, it can break
23:49down this problem just like a human
23:51would. For example, it might think first
23:54I need to search for AI startups using
23:56my web searching tool. Okay, let's do
23:57that first. Then I'll need to create a
23:59new spreadsheet with my Google Sheets
24:01tool. And then for each company that I
24:03find, I'll need to add a row to the
24:04spreadsheet. And then I'll need to write
24:06a summary of each business and put it in
24:07a new column. And then finally, I'll use
24:09my email tools in order to send the link
24:10to Liam. And that's all great, but then
24:12when you add on top of that powerful
24:14reasoning models like OpenAI's 01 and 03
24:16and even things like deepseat as the
24:18brain of the agent that can plan, take
24:20actions, then reflect and then plan
24:22again and so on. You have essentially
24:24created a truly intelligent AI that
24:26solves problems and approaches them just
24:28like a human would. So, say for example,
24:30the original plan was to use the web
24:32search tool to search for AI startups
24:34raising money. Probably a terrible
24:35search term, but what if that doesn't
24:38return any good results to the agent?
24:39Well, a human would go, damn, I need to
24:41change my search term or maybe I need to
24:43try find a different method of finding
24:44these companies on like LinkedIn or
24:46something. The latest in AI technology
24:47like these reasoning models, it allows
24:49these agents to do this exact same kind
24:51of reflection and replanning in order to
24:54achieve their objective. And this is
24:55when you can really see why we call them
24:57digital workers because they can do
24:59things like planning multiple steps.
25:01They will use different tools in a
25:03sequence and even adjust their approach
25:05based on the results from those tools.
25:07Now, I should mention that this
25:08technology isn't perfect yet, right? So,
25:10these multi-step tasks are often
25:12unreliable and agents typically need
25:14human supervision for more complex
25:16workflows. But things are moving
25:19incredibly, incredibly fast. In fact,
25:20we're already seeing the next evolution,
25:22which is multiple agents working
25:24together. Instead of just one agent
25:25trying to do everything, you can have
25:27one main agent that you give orders to,
25:29and then it can use all of the other
25:30agents underneath it as tools where it
25:32can send specific instructions. Like
25:34underneath the main agent might be a
25:36research agent, which is best at finding
25:38companies and has its own tools. Then
25:39you have a writing agent that's really
25:41good at writing summaries. Then you have
25:42an emailing agent, which has got all the
25:44emailing tools. And so each of these
25:45agents can be specialized in their
25:47specific task with multiple tools. and
25:49then they all work together to achieve a
25:50common goal. This is exactly what major
25:53companies like HubSpot and Microsoft and
25:54Google are building towards. It's these
25:57entire workforces of AI agents that can
25:59handle complex business processes
26:00automatically. In the next chapter, I'll
26:02show you how to build AI agents like
26:03this for yourself using no code tools.
26:05But first, we need to understand the
26:06different ways that these agents can
26:08actually be used in the real
26:11world. So, we understand how AI agents
26:14and tools work under the hood. Now,
26:15great. If you don't, please go back and
26:18take some notes, right? You should by
26:19now have a whole bunch of notes um from
26:21the stuff that we've covered already.
26:22And this stuff that you're learning took
26:23me two years in order to learn and and
26:25be able to apply effectively. So, you
26:27best believe it that it's going to take
26:28you two to three watches before it all
26:30sinks in. So, if you're feeling a bit
26:31lost and and overwhelmed, don't worry.
26:33That's how it feels with learning
26:34anything new or how it should feel if
26:36you're learning something that's
26:37actually pushing your boundaries and
26:38adding something to your to your
26:39capabilities. Next, we need to look at
26:41the different ways that AI agents can be
26:43used in the real world. There are two
26:45main categories of AI agents.
26:47Conversational agents and automated
26:49agents. Conversational agents are ones
26:50that humans interact with directly
26:52through chat on things like websites.
26:55You've got maybe you're chatting to it
26:56on WhatsApp. You've got interacting with
26:58it over the phone via phone call. You've
27:00got chatting to it via Instagram DMs or
27:02custom apps and websites. For example,
27:04OpenAI's GPT platforms allows you to
27:06create agents that you can chat with
27:08directly on your computer or on your
27:09phone. or using platforms like my own
27:11Agent, you can connect these agents that
27:13you build onto a WhatsApp number or onto
27:15Instagram. And I'll show you how to do
27:17this in the tutorial chapter of this
27:19video. So, in all these cases, you or
27:21someone else is there sending messages
27:23or instructions to the agent and
27:25explaining what you want to do and kind
27:26of chatting back and forth with it,
27:28whether it's on a website, WhatsApp,
27:29Instagram, or whatever. And within these
27:31conversational agents, it's not just
27:32text based. It's like I said, there's AI
27:34voice agents as well, which are an
27:35extremely exciting sector of the AI
27:37space right now. And these systems use
27:39multimodal models that can take in audio
27:42as input and then produce audio as an
27:44output. And so these agents can be
27:46chatted to over the phone or via audio
27:48rather than via text. This AI voice
27:50stuff is super cool. And in the tutorial
27:52section, I'm going to show you how to
27:53take the exact same AI agent that we can
27:55chat to on a website and then connect it
27:57to a phone number and talk to it on the
27:58phone. But then we get to what I call
28:00automated agents. And so these are
28:02slightly different from the
28:03conversational ones. The truth is that
28:04AI agents don't always need humans to
28:06talk to them and use them directly. All
28:09they need is some kind of input or
28:11instructions to trigger them and that
28:13tells them what to do. This means that
28:14we can build these automated agents that
28:16instead of waiting for some kind of
28:18human input, they are actually part of
28:20larger systems and processes and they're
28:22triggered automatically by events like a
28:24new email received or a form submission
28:26or they work on schedules like once a
28:28day and they essentially work in the
28:30background without necessarily having
28:32human oversight or input. For example,
28:34later in the video, we are going to be
28:35building an automated agent that is
28:37triggered by a new form submission. When
28:39the form is submitted, some of that form
28:41data is taken and sent to the agent,
28:42which then causes it to use the tools
28:44that we've equipped it with and follows
28:46the instructions in the prompt that we
28:47gave it in order to make decisions and
28:49take appropriate actions on our behalf
28:51in a fully automated way. We are still
28:53sending the message to the agent, but
28:55it's not a human needing to type it
28:57manually or speak it over the phone.
28:59There's no human step. The input is
29:01being automated in some way. And this of
29:02course opens up a huge number of use
29:04cases for AI agents in businesses
29:06especially. And of course I'll be
29:07showing you how to build both types of
29:08these conversational and automated
29:10agents in the tutorial section of this
29:12video. But the last step of building
29:14your foundation of knowledge before we
29:15move into that is to look at some real
29:17world examples of how businesses are
29:19using these AI agents right
29:23now. So firstly we have the personal
29:25assistant category. And this is what
29:27most people think of when they hear the
29:28word agent. something that you can chat
29:30to that's going to update your calendar
29:31and sort of send emails and even make
29:33phone calls for you. Um, now these are
29:35all nice to have features, but honestly
29:38uh this space is likely going to be
29:39dominated by the big tech giants. You've
29:41got OpenAI through Chatbt trying to do
29:43this with Tasks, Google through their
29:45suite of apps and connecting them to
29:47Gemini and Apple through Siri. These
29:49guys are going to eat up this entire
29:50market of personal assistance and your
29:52own personal AI agent that helps you do
29:54personal stuff. the real opportunity
29:56lies in business applications and how
29:58people like you and I can build and sell
30:00AI agents to businesses which we're
30:01going to be covering in depth in the
30:03final chapter of this video. So, we've
30:04got the next chapter which is going to
30:05be on building the four tutorials and
30:08the final chapter is all about how to
30:09sell and how to monetize your AI agent
30:11skills that you've just learned. One of
30:12the core use cases for businesses right
30:14now are what's called co-pilots. And
30:16these are AI agents made for specific
30:18roles in a business. We're going to be
30:20building one of these later in the
30:21video. And these specialized AI agents
30:23are essentially helping someone in a
30:25specific role in a business to do their
30:27job more effectively. Take a customer
30:29support co-pilot for example. It would
30:31have a knowledge base that allows reps
30:33to get answers to customer queries
30:34instantly and deliver them over the
30:35phone. So they've got the little
30:36co-pilot up on the side there. They're
30:38on the phone. They get a question, they
30:40can search and for an answer in the
30:41knowledge base, it gives them back and
30:42they can give it to them over the phone.
30:44This same agent could also have a tool
30:45that allows them to look up the customer
30:47information very quickly. Um, I could
30:49have another tool that it makes it very
30:50easy to send a summary of the call into
30:52the database so that the next rep who
30:54picks up the phone and talks with them
30:55knows exactly what was discussed
30:56previously. It's like giving every
30:58support rep some kind of AI assistant
31:00that makes them dramatically more
31:01effective. It also makes their customer
31:03support a lot more consistent as to what
31:05the company wants people to be saying,
31:06which is a a big problem with managing
31:08large customer support systems. And then
31:10we have lead generation and appointment
31:11setting agents. These are probably the
31:13most valuable type right now. And
31:15businesses are using these on their
31:16websites, through WhatsApp, on
31:18Instagram, and even over the phone to
31:20engage and have conversations with the
31:21interested people who are approaching
31:23the business 24/7. They can offer
31:25instant answers about products and
31:27services. And they're even smart enough
31:28to be able to capture emails and phone
31:30numbers mid-con conversation for later
31:32follow-up by sales team. Some can even
31:34book appointments on the spot and mid-
31:36conversation by using a tool to check
31:37the calendar availability and then using
31:39another tool to create a new booking
31:41once they've agreed on a time with the
31:43prospect. Another real world agent use
31:44case and one of my favorites is a
31:46research agent. And so these can help
31:48businesses by automatically researching
31:50leads that come in through their website
31:51or elsewhere. And when someone fills out
31:53a form, the agent can spring into action
31:55and start searching the web for
31:56information on the company, finding
31:58their LinkedIn profile of the person
31:59they're going to get on a call with and
32:00gathering any other valuable data that
32:02it can find. Then it can take all of
32:03this information and generate a summary
32:05of who this person is and what this
32:07company is also and decide whether
32:09they're a good fit for working with the
32:10company and if so then they can send the
32:13sales team some kind of detailed brief
32:14or suggested strategy on how to close
32:16this particular person on a call based
32:18on the research. So it's basically like
32:20having a an automated team of
32:22researchers who as soon as leads show
32:23interest in your business, they're out
32:25there figuring out everything about them
32:26and determining one whether they're a
32:28good fit for you and your products and
32:29services which is called qualification.
32:30Then secondly, if they are qualified,
32:32giving the sales rep something that will
32:34bring them up to speed on who this
32:35person or who this company is and how
32:37they can try to close
32:39them. So, we have covered a lot so far.
32:42So, before we dive into each of these
32:43agent builds that I'm going to walk you
32:45through over my shoulder, please make
32:47sure that you've got your notes taken
32:48out and the core concepts of this video
32:50so far understood properly. You should
32:52be clear on things like what is the
32:54definition of an AI agent? What are the
32:56five parts of an agent? How is building
32:59an AI agent like being a chef? And how
33:01many ingredients do you have to play
33:02with? What are the two main parts of a
33:05tool? And what do schemers do? So, pause
33:08the video now and try to answer these
33:09questions. And if you aren't 100%
33:11confident, you need to go back and watch
33:13it again. So, don't rush this or you're
33:15going to feel way out of depth when we
33:17get into the tutorials that we're going
33:18to be covering next. But if you are,
33:19congratulations. You are one step closer
33:21to AI literacy and becoming a much more
33:23valuable uh participant in this global
33:26economy. So before we get into the
33:27second chapter, there's just three very
33:28quick things from me. Firstly, if you
33:30are a business owner who wants to fast
33:32track to becoming an AI leader within
33:34your industry, at my agency, Morningside
33:36AI, we offer everything from AI
33:38education and upskilling programs for
33:39executives and staff to AI strategy and
33:42roadmap consulting and of course AI
33:44development services as well. So we
33:45would love to help you get ahead. So
33:47feel free to get in touch via our
33:48website in the description below. And
33:50secondly, at Morningside, we are hiring
33:52for all sorts of roles right now. So
33:53whether you want to build AI systems for
33:55some of the world's biggest companies
33:56that we have as clients or to help
33:58produce videos like these that are seen
34:00by millions of people or create
34:02educational material for thousands of
34:04businesses. Uh we have roles for all
34:05sorts of things right now. So you can
34:06apply using the link in the description.
34:08And please, even if you're just vaguely
34:09interested, I really recommend you just
34:11check out the link and see what roles
34:12we're hiring for. Uh you never know
34:14what's going to be on there. Um and it
34:16may be a very good way for you to use
34:17your skills to fast track into the AI
34:18space by working under myself and my
34:20team. And finally, if you have gotten
34:22any value so far in this video, please
34:23head down below and leave a like on the
34:25video. It helps me reach more people.
34:27Um, I put a lot of work into these
34:28videos and it also lets me know that you
34:30enjoy this kind of content and that I
34:31should make more of it. And of course,
34:32if you like this kind of content and
34:33want to see more of it, you can
34:34subscribe so that YouTube will put my
34:36videos up for you whenever a new one is
34:38released. So, there's also a little
34:39share button if you want to click that.
34:40That'll let YouTube know this is good
34:41content and that you're sharing it to
34:42other people. Not only will that help
34:44me, but you can share it to your friends
34:45and family who may also or you may want
34:48to help them to brush up on these skills
34:49or help them give a way to get on the
34:51front foot with AI. And that's what I
34:52really make these videos for. So, thank
34:54you for sitting through that little bit
34:55of housekeeping and self-promotion. Now,
34:57let's get stuck into the
35:01building. I have carefully assembled
35:03this chapter on building to give you the
35:05most bang for your buck possible in
35:07order to kick off your AI agent learning
35:09journey. We are going to be covering
35:10four different use cases across four
35:12different AI agent building platforms.
35:14These are all no code, so don't worry
35:16about that. And the chances of you
35:18falling in love with at least one of
35:19these platforms is pretty much 100% as
35:22you're going to rapidly start to connect
35:23the dots uh about how you can start to
35:25use these kinds of agents and these
35:27platforms in your own life or in your
35:28work or for your friends and family and
35:30those around you. So, here's a quick
35:31rundown of the builds we're going to be
35:33getting into. The first build is going
35:34to be a sales co-pilot built with
35:36relevance AI. And here we're going to be
35:38building three custom research tools
35:40from scratch, including an advanced web
35:42scraping tool, which is a a great skill
35:44that I want to teach you. And with
35:45these, we are going to be creating a
35:46conversational agent to help the sales
35:48reps at Big Boy Recruits, a hypothetical
35:51fantasy uh recruitment firm, in order
35:53for them to be better prepared for sales
35:55course. So that's the purpose of the
35:56sales co-pilot. The second build is
35:58going to be an automated lead
35:59qualification agent. And this will be
36:01built on a platform called N8N. And this
36:03time we will be helping Big Boy
36:05Recruits, our fantasy recruitment firm,
36:07to automatically research and qualify
36:09new leads and then send an email
36:10notification to the correct sales rep.
36:12And this is going to show you that
36:13automated style of agent where it's
36:15built into a process rather than having
36:16a human input necessarily. In build
36:18number three, we will be building a
36:19website and phone-based lead generation
36:22customer support agent. This will be
36:23built on voice flow and it's going to be
36:25able to do three things. Firstly, answer
36:27questions from a knowledge base,
36:28generate instant quotes using a custom
36:30tool we build and also do lead capture
36:33on interested prospects. We're then
36:34going to slap this agent onto a website
36:36widget so that you can chat to it via a
36:38website and via chat and text. And then
36:40we're going to take that exact same
36:41agent and connect them to a phone line
36:43so that we can call our agent over the
36:44phone and access all of the same
36:45functionality we just talked about. And
36:47finally, for build number four, we'll be
36:49using my own AI agent platform, Agent,
36:51to rapidly build a lead generation agent
36:53and connect it to a WhatsApp number that
36:55we can chat to. The leads that we
36:56collect are going to be automatically
36:57sent into an Air Table database for
36:59later review. And please don't skip
37:01around these builds as they're all kind
37:02of connected in some way where we're
37:04reusing parts from build one and build
37:05two, etc. But without further ado, let's
37:08get into building some agents. All
37:09right, people. Enough of the theory. Uh,
37:12now we get into the fun bit of actually
37:14building these agents out. So, I've done
37:16a lot of work and my team has done a lot
37:18of work. So, thank you to the my team
37:19members who have helped me put this
37:20together. Um, putting together four
37:22different AI agent builds for you. And
37:24this is really going to walk you through
37:26an A to Z all the different platforms
37:28that you really need to care about, all
37:30the different kind of core use cases and
37:32functionality. There's a lot more of
37:33course, but this is going to really give
37:35you the foundation that you need to
37:37succeed in the space. And hopefully
37:39it'll be the thing that kind of sparks
37:40your interest in it because I I want you
37:43guys to have fun with it. these big
37:44tutorials for me. Honestly, when I put a
37:46lot of work into it, I build up the sort
37:47of mental resistance to it because I
37:49know how much work there is going into
37:50it and I have to make this big whole
37:52session where I'm all uptight about it.
37:53But I'm just going to try and relax and
37:55enjoy this. And I really want you all to
37:56do the same. So, set a bit of time
37:58aside. You can either pause this video,
38:00put on your watch later, but I really
38:02want you to take your time with this.
38:05I'm going to be doing this more. So,
38:06when you do tutorials like this, there's
38:07a few different ways you can do it. I
38:09can either do all the building and then
38:11give you the templates and kind of just
38:12spoon feed it to you. And that's more so
38:14what you do for someone if you're trying
38:15to like really fast track them and they
38:17don't want to learn all the skills, but
38:18um I I know what I'm trying to build
38:20here for you guys. And I'm going to give
38:21you a sort of stream of consciousness.
38:23You just get to see me kind of jamming
38:25out and building these things. And I'll
38:26be explaining my thought process and the
38:28concepts etc along the way to reinforce
38:30what we've learned before. So I'm just
38:32going to dive into it with our first
38:34agent.
38:39And so what I've done is put together a
38:40big Figma board here which is going to
38:41be breaking down all these different
38:42builds. So under here there's some
38:43goodies you see. Oh, there's some
38:45goodies under each of these that I've
38:47put together. Um and we're going to go
38:48through them one by one. Starting off
38:50with agent one over here. I mean there's
38:51a lot of stuff here um that you guys are
38:53going to get. So you'll get the whole
38:54Figma and it includes all the templates.
38:56So if you do want to just kind of watch
38:57through this, pick it up. You can either
38:59do it and follow it step by step with me
39:00and see how I build it and really build
39:02those flexible skills that you're going
39:03to need to succeed in the space or you
39:05can just watch it and be like, "Okay, I
39:07kind of get what he's doing and then
39:09take all the templates from me at the
39:10end." That's I mean, completely up to
39:12you. Depends if you want to be a really
39:13really nerdy builder about it and get
39:14into the weed like like I like to do. Um
39:17or you just want to be like, "Hey, I
39:18want to do this my business. I want to
39:19roughly understand how these things work
39:21and what platforms." So, use this
39:23resource as you will. But we're going to
39:24jump into agent build number one here,
39:26which is our sales co-pilot built with
39:28relevance AI. So, running through this
39:30quickly, we have the purpose of this.
39:32This is basically going to look a bit
39:33like this. It's going to be a co-pilot
39:35and co-pilots work in that you have a uh
39:37it's basically a specific AI agent that
39:40you build for a specific staff or staff
39:42member or role. So, say this case, it's
39:44going to be a sales co-pilot. It'll be
39:46the thing that the sales rep uses to uh
39:49in their day-to-day as they're working
39:50on their jobs. You can add tools like in
39:52this case, you see we're going to have
39:54three different tools here for our
39:55agent. One's going to be a company
39:56researcher tool. So this is when the
39:58sales rep would be like, hey, I have a
40:00call coming up soon. Um, let's put in
40:01this I need to research this company cuz
40:03this is who I'm going to be on a call
40:04with. So they'll put in the company URL.
40:06This tool that we're going to create is
40:07going to go and research that company.
40:08It's going to bring back and give a
40:10summary. And then it's like, okay, well
40:11this is the LinkedIn URL of the person
40:12that we're going to be got on a call
40:14with shortly. It's going to pass in the
40:16LinkedIn URL. It's going to take that
40:17URL. It's going to pull all the
40:18information and write a summary about
40:20the person. So now we have the company
40:21summary and we have the person summary.
40:23And the final step here is going to be
40:24what I'm calling a pre-all report
40:26generator. And that's going to take both
40:28that company and prospect research that
40:29we've done. It's going to combine them
40:31together and be for this specific
40:34company. As you're going to see that
40:35this hypothetical company we're building
40:36this sales co-pilot for, it's going to
40:39generate a basically a pre-core report
40:41or a strategy uh a strategy prep for the
40:43sales rep so that they go onto those
40:44calls much more prepared and also sort
40:47of a personalized guide on how to try to
40:49close this person. So, um, all of these
40:51templates are going to be here. Each of
40:53these are templates for the tools. And
40:54this is for the agent as well. Um, but
40:57here's some more information. You guys
40:58can pick through this as you wish. But,
40:59um, this is the kind of end result and
41:01we're going to be able to chat to it.
41:02And this would be something you could
41:03build for a client. You could build it
41:04for your own business or you could just
41:06tinker around. You could build co-pilots
41:07like this on relevance for yourself. So,
41:09that's why I want to start with
41:10relevance because it also is a platform
41:11that we can build these tools on. So,
41:13it's a really, really good one to start
41:14with and let's get into it. So, the
41:16first step of course is to go to
41:18relevance AI. So, I'll put a little link
41:19up here. You guys will be able to get
41:20this Figma. It'll be on the school. Um,
41:22all of the information and all the
41:23resources for this are going to be like
41:24this Figma is going to be linked to the
41:26school. My free school community. If you
41:27haven't already joined, biggest AI
41:29community on school, biggest AI business
41:31community probably in the whole world.
41:33Um, so we can jump across that first
41:35link in the description. You'll be able
41:36to find this in the YouTube resources
41:37section. Um, pretty straightforward. Of
41:39course, when you click on this, it's
41:41going to ask you to log into relevance.
41:43So, if you haven't already, you can make
41:44an account. Um, it's fairly low cost.
41:46They have a free plan, then a team's
41:48plan, I believe. Um, so it's not too
41:49much, but it is a really, really
41:50valuable tool as you're going to see.
41:52So, you can sign in here. I'm going to
41:54jump in with my Google account. There
41:57may be a bit of setup for your account,
41:58but I'm sure you guys are smart enough
41:59to figure out how to set up an account.
42:00I'm sure relevance also makes it easy
42:02enough. So, then we get taken to this
42:04dashboard, but we see on this left hand
42:05side, we have tools. So these are the
42:07tools that I've talked about um where
42:08it's some kind of functionality that we
42:10can create and we can build it all on
42:11relevance no code and there's even sort
42:13of extensibility or you can add more
42:15functionality in to relevance by adding
42:18some low code components or even custom
42:19code. So relevance is a really great
42:21base for building not only tools but
42:23then the agents that you can connect
42:24that into and we're going to use the
42:25same relevance tools that we make now in
42:27multiple of the different agents that
42:28I'm going to make for the rest of this
42:29video. So first things first if we go
42:30back to the Figma here we see we need to
42:32make three different tools. So tool one
42:33is company researcher. It's going to
42:35take in a company URL. It's going to
42:36search the web and it's going to return
42:37a summary. So, that's the functionality
42:39we need. Let's go and create a new tool.
42:42Going to call this um
42:45research company. We can give it a cool
42:48little I'm going to zoom this up for you
42:49guys. Hopefully, that's the right size.
42:51Um uh search. Have fun with this stuff,
42:55guys. Like, if if it's putting stupid
42:57emojis on things to to enjoy it, then
42:59then that's what you need to do. Like,
43:01uh if you make it a chore, it's going to
43:02feel like a chore, right? Um, now we get
43:04to descriptions. Um, this is something
43:06we're going to see recurring, but
43:07basically as as you know, as we learned
43:09in the concept section, we need to have
43:11natural language descriptions of our
43:13tools and of our APIs so that the agents
43:15can read those descriptions and
43:16understand uh what the agent or what the
43:18tool does and what those different parts
43:20does. So, you'll see this recurring
43:22throughout this. But first things first,
43:23what does this research company tool do?
43:26Um, takes and oh, I got caps lock on.
43:30takes um and there's also some tutorials
43:32here if you want to go deeper. Relevance
43:34has some great documentation as well. Um
43:36but takes in a
43:40company
43:43penny URL
43:45and scrapes the website then returns a
43:51sum and AI generated summary. So then to
43:55build a tool we need to have some kind
43:57of input. You don't always need an
43:58input. It can actually just be
43:59triggering, but generally you're going
44:00to have some kind of input that the
44:02agent needs to pass into the tool. In
44:03this case, to research company, it's
44:05going to need to take in some text,
44:07which is going to be a URL. Um, we're
44:09going to say comp company
44:13URL. And then again, here we have
44:15another description. You see, describe
44:16how to fill this input. This is again
44:18going to help our agent within relevance
44:20AI and elsewhere as you'll see in
44:22tutorial number four. Why this is so
44:24important to add in the descriptions,
44:25right? So this is a
44:29URL for a company to be
44:34researched must be in the format https
44:39colon slash
44:43um dot dot dot dot dot. So we need to
44:45have the https for this to work. So
44:48that's going to be our input. Now if
44:49this seems a bit confusing just stick
44:50with me. It will make sense in a second.
44:51this stuff. If there's anything that
44:53I've learned from picking up so many
44:54different tools, like when I first got
44:56into Facebook ads, when I've got into
44:57building these kinds of agents, it's you
44:59feel completely overwhelmed, but that's
45:00all just part of the process. And what
45:02feels difficult now is not going to feel
45:03difficult forever. So, just please stick
45:05with me. Um, and it's a really, really
45:07great feeling once you go and be like,
45:08this was hard a few weeks ago and now
45:10it's really easy. So, we've got our
45:12first input. That's what the research
45:13company tool is going to do. And then we
45:15need to define our steps. So, the next
45:17step, we can go add. I'm going to hide
45:19this so we get a bit more space. Add
45:20step. Now, the cool thing about
45:22relevance is that it comes with a lot of
45:23great functionality out of the box.
45:25Here's one, extract website content, we
45:27have LLMs, we have Google searching, we
45:30have all sorts of AI generations,
45:32replicate um knowledge bases as well.
45:35There's so much cool stuff on here and
45:36this is why I really rate relevance as
45:38one of the best platforms. If you were
45:39to go all in and want to upskill, you
45:42can build so much on this. Um, so I'm a
45:44big fan. I love the the relevance team
45:45and what they're doing. So the initial
45:47plan for this build I was going to use
45:48the extract website content which is
45:50fairly straightforward. We can say oh
45:52one other thing um the company URL we
45:55are going to be using this company URL
45:56throughout our tool here. So we can
45:58change this text to say something more
46:00descriptive. So company
46:03u URL. You guys are going to got going
46:06to think like think and write like
46:07coders now cuz you need to uh use some
46:09kind of syntax and use some kind of
46:10variable naming convention. This is a a
46:13standard one or you can do things like
46:15company URL camel case but I prefer this
46:18format as I'm I'm mainly a Python kind
46:20of guy myself. Now that we have that
46:21named we can use it in these kinds of
46:23fields. So here you can see pops up use
46:25inputs. So basically when this tool is
46:27run it's going to take the inputs or the
46:29information we put in the inputs and
46:30it's going to pass it to different steps
46:32and use them as we describe within the
46:33within the builder here. So let's run
46:35this quickly. Let's put in my company
46:39https/
46:43morningings.ai. Um, and then we can
46:46click uh run here. So, that's going to
46:49go to my website and scrape the
46:51information off of it. There we go. So,
46:54it's got all of this, but you see it's
46:57just pulling back the first page. Um,
46:58and this is why I actually I shuffled
47:00this around. And I want to show you guys
47:02how to do something a bit more advanced.
47:03I know this is supposed to be a beginner
47:04tutorial, but this is not really that
47:06useful and I it's a very easy thing for
47:08me to just bump this up to a little bit
47:10more valuable. Um, while relevance tool
47:12here is great, we can do better. So,
47:14we're actually going to delete this. Um,
47:16you can use that step for all sorts of
47:18other things, but I really like what's
47:20called fire crawl
47:23firecroll web scraper. This is a cool
47:26app. Um, firecrawl.dev. Shout out to the
47:29guys at firecraw. Basically, if we then
47:31go HTT Oh, I should just see if they can
47:33do it. Morningside
47:37doti and now do a free scrape for us
47:40here. So, this is just going to do the
47:42single URL just like we got in
47:43relevance. But the difference here is if
47:46we then go to crawl, if you hover over
47:49this, it's going to crawl a URL and all
47:51of its accessible subpages outputting
47:53the content from each page. So, instead
47:55of just taking that front page, it's
47:56actually going to crawl through multiple
47:57things. So this is really the first cool
48:00thing or or first skill that I want to
48:01put in your tool belt is that you have
48:02things like fire call that you can use
48:04their relevance. They have things like
48:05map which is just going to output all
48:06the URLs that it finds. Then there's
48:08other things here where you can use AI
48:10to extract data. I'm not going to go
48:11into that. But what we want is this
48:12crawling functionality from firecraw. So
48:15I'll put a link in on the school post
48:16that comes with this video. It again
48:18first link in the description to go to
48:19the school and if you go to the YouTube
48:21resources section um there will be a uh
48:23a whole post on this and all the
48:25resources will be in there. It'll also
48:26be in my free course on school as well.
48:28So you can find it in the classroom
48:29section. So you want to sign up to FCL
48:32so you can get API key. It's very easy.
48:34We can just go through with Google.
48:35Again, this is not sponsored and there
48:37is zero sponsoring going on through any
48:39of these tools. I guess I'm kind of
48:40sponsoring my own tool because I'm
48:41putting it at the end. But I'm not
48:43getting paid a dime for any of this. I'm
48:44really just trying to put you guys on
48:45what I like to use, what's made me
48:47money, what's made me a more valuable AI
48:49automation expert or developer for my
48:52companies and for the companies you work
48:53with. continue and then you get to the
48:56dashboard here. It might look a bit
48:57scary, but what you can do is go to the
48:58API keys. So, you can click on create an
49:00API key here,
49:03YouTube. I'm having issues with mine. It
49:05should be fine for you. I've already got
49:06an API key, but once you get the API
49:07key, what you can do is take it and come
49:09back here to
49:11relevance. And you can go into the side
49:13panel here. Where is it? Settings. And
49:16then we have our API keys. So, we're
49:17going to need to add more into this
49:18later. So, keep an eye on this. This is
49:20something you need to be familiar with.
49:22Um, and you can scroll down to firecrawl
49:23and you just pop it into this firecall
49:25API key section here and you're good to
49:28go. And you can come back. Oh no, we
49:29don't want to duplicate
49:32that. Now we've got our firewall set up.
49:35We need to make a couple things. We need
49:37to do a couple tweaks here. We have, of
49:39course, the if you want to get those
49:41variables up, you can go bracket bracket
49:43um or curly bracket curly bracket, which
49:44is shift um to get those. I don't know
49:46why it's not popping up. There we go.
49:49Company URL. And if we hover over this,
49:50we can see it says scrape the provided
49:52URL only. Uncheck if we want to crawl
49:54instead. So if we want to get that crawl
49:56functionality that we just saw that we
49:58think we want to get all the data, we
50:00can uh uncheck
50:02it. And then we want to extract the main
50:05content. So you might have to just trust
50:06me on that one that we don't want to
50:07have all of the other rubbish. We just
50:09want the body of the website. Um a
50:11number of pages. We don't want this to
50:12take too long. You can expand this much
50:13more. Um but I'm just going to go for
50:15say five for now just to keep it uh keep
50:18it quick.
50:21And then now what we can do is run this
50:23again. We've still got my URL up here.
50:24We can go run
50:27step. Give it a second. You will at some
50:29point have to pay firecrol. Of course,
50:31it's not a free service, but they do
50:32have a free plan, so you should have no
50:33issues with getting that. So here you
50:34can see we're getting a lot more data
50:35back from this web scrape than we were
50:37with just the relevance version, which
50:39is great. So the next step is we have
50:41this data. We want to generate some kind
50:42of research summary um so that we can
50:44send that to our sales script once they
50:46have requested it. So now it gets into
50:48the fun part of writing LLM prompts. For
50:51this one, sometimes you need to really
50:53go and and make a big effort, which we
50:54are going to do later as you'll see. But
50:56in this case, we just want a quick
50:57summary. I'm going to throw one in here
50:59that I made earlier. All of this will be
51:00available on the Figma or it'll be given
51:02somewhere in the resources, right? But
51:03um if we just need something quick and
51:05dirty here, it's not really a massive
51:06part of the project, so it's okay to
51:08just whack one in there. So bang, I've
51:09got it in there. Can you please take
51:10this website content and summarize it
51:12into a 300word natural language summary,
51:14which clearly outlines rad where they're
51:15based, their values, etc. anything that
51:17would be helpful to know for a sales rep
51:19who will soon be on a call with them. Uh
51:20break it into key areas like overview,
51:22products and services team etc. And I've
51:24got a couple things here to make sure
51:25that it doesn't mess up which I it was
51:27doing for me a bit in testing. So we do
51:29want to put in if we go curly bracket
51:30curly bracket we want to put in the fire
51:32call data here which is going to be uh
51:34all the website data that came back from
51:36our scrape. We want to then insert it
51:38into this prompt here. So I hope you're
51:39starting to the things in your your cogs
51:40and your brain are starting to click
51:42into gear here. And then we get to
51:43select the model for this. It's pretty
51:45basic task. So, I want something quick
51:46and cheap. Um, for many of you, it's
51:48going to be easiest to work with the
51:49Open AI APIs because you've probably
51:51already played around with your API key
51:52before, which I'll show you how to do in
51:54a second. But, let's go with GPT40 Mini,
51:56but Relevance, of course, does have
51:58support for all of uh the other models.
52:00But my tendency for most tasks now is to
52:02actually go for some of the Google
52:03models that have come out. Again, you
52:05guys might be watching this in a year or
52:06whatever. It might be quite different,
52:08but at the moment, Google is really
52:09leading the way with making the cheapest
52:11models possible, and they're actually
52:12really good as well. the the price
52:14decrease on using things like uh Google
52:16Gemini Flash Light and Google Gemini
52:17Flash 2.0 and stuff like that. It's
52:20ridiculously cheap and it's also a
52:21really good model. So, it's a bit more
52:22difficult to get APIs on the Google
52:24side. So, I'm just going to stick with
52:25OpenAI for this tutorial. Um, so let's
52:26go GPT4 mini. And then we need to of
52:29course set up our API key. Um, if you
52:31scroll down, where is the OpenAI? So, to
52:33get your OpenAI API key, you need to go
52:38platform.openai.com.playground/playgroundplayground,
52:39sorry. Um, this will again be linked in
52:41the resources or you can just type up
52:43platform.openai.com. I'm sure you'll
52:44find it. If you haven't already, create
52:46an account um, and sign in. And then you
52:48need to go to your dashboard here. You
52:50go on the left side, you go to API keys,
52:52and you click create a new secret key.
52:54And then you're going to be able to copy
52:55that key and bring it back into uh,
52:57relevance. Paste it in here. And then
52:59you're good to go and start using the
53:00OpenAI suite of models. It's pretty
53:02easy, right? So now we have our
53:04information back from the scrape. We
53:06have our prompt in here. And then we can
53:09go run step and see what this company
53:11research tool is going to output for us.
53:12Boom. Morningside AI is a leading
53:14artificial intelligence development
53:15company dedicated to empowering
53:16businesses autonomous AI agent
53:18development, enterprise consulting,
53:19chatbot development team. Uh they
53:23got keep me out of my own team page.
53:25Damn. But uh yeah, so there you go.
53:27That's the that's the company research
53:28tool. That's step one. Um I hope you
53:30guys can kind of see how that works. Now
53:32a cool thing about relevance is there is
53:33this build section which we've just gone
53:35through. But you can also go to use and
53:36this can be really helpful when sharing
53:38these kinds of tools. So not this is not
53:39really only just an AI agent tutorial.
53:41I'm also teaching you how to build tools
53:42because you can build very valuable
53:43tools and something like relevance. And
53:45then you can go share. You can go
53:48publicly
53:49available. Oh, and I can click this and
53:52then I can give to my employees. I can
53:54give to the companies that I'm working
53:56with, the clients that I've sold these
53:58kind of services on. And then they get a
53:59nice and handy tool like this. I mean, I
54:01use these throughout my organization for
54:02like description generation, a lot of
54:04content repurposing. There's tons and
54:06tons of different use cases for building
54:07these kinds of tools. And then you can
54:08take this URL and you can share it
54:09around to whoever you want. So, this is
54:11a uh a great way to use relevance. Let's
54:13go back to our tool here. But for now,
54:15we need to keep moving along so we can
54:17get this this first agent done. So,
54:18that's the first tool. I'm going to run
54:19through it a bit quicker now that we
54:20know how these kind of tool buildings
54:22work. I'm going to create a new tool
54:24here. I'm going to call it this time if
54:26we go back to our Figma. And I recommend
54:28when you guys are building your agents,
54:29you're building systems and planning
54:30them out. This kind of laying out in a
54:32Figma, if you're not familiar, this is
54:34Figma. It's like a design software, but
54:35you can also use it for kind of
54:36whiteboarding. It's called a Fig Jam
54:38board. It's one of the types of uh
54:39boards you can do. And I use this all
54:41the time with my team as well. It's like
54:43if you can't take what I'm telling you,
54:45lay it out on a board so that I can
54:46review it and give you notes and and
54:47then we all agree this is the build. Um
54:50there's often a lot of uh communication
54:51issues with explaining functionality. So
54:53this AI agent layout, it's really
54:55helpful. You can if it's maybe a
54:57workflow automation, you can do box box
54:58arrow arrow arrow all of this laying out
55:00how it's going to be built. Here I've
55:02done it for you in a very basic format.
55:04So we've done this first one. Maybe I
55:05just make this green. Um second one is
55:07going to be prospect researcher. So this
55:09takes in a LinkedIn URL, searches the
55:11web, and returns a summary. And you're
55:13going to see how I tie this all into an
55:14agent
55:15shortly. So we're going to go research
55:18prospect. Sil takes in a linked in URL
55:22scrapes the profile and then generates
55:25an of the prospect input. We're going to
55:28need a link and URL the link linked and
55:32URL of the prospect.
55:35and we rename
55:40this. Now, we're going to add a step and
55:43relevance has got us here linked
55:46um get a LinkedIn profile or company
55:50post. So, this is cool cuz then we can
55:52pop in our LinkedIn URL
55:55here. Oh, we don't need two of
56:01them. So, we're going to get the user
56:03profile here. Then if I put in
56:08mine my LinkedIn profile, if you guys
56:10want to connect with me on LinkedIn,
56:12more than welcome to do so. I'll put in
56:13the description below. Um, we can do a
56:17little run step here. So, if we go back
56:19over to data here, that's great. We've
56:22got my about section. So, this goes very
56:24long way across because it's in uh it's
56:26in JSON here. It's got my company. It's
56:29got my company domain, where I'm from,
56:33years, company, founded, tons and tons
56:34of great information that you guys can
56:36use and we are going to use shortly. So,
56:37this is really cool. I would probably
56:39add one more step to this if I was
56:40taking this um and building it for for
56:42my own team. I would add in another
56:44LinkedIn scrape here um where we just do
56:46the same thing, but we also get the
56:48posts because posts can give you a bit
56:50more up to date um information on what
56:51they've been doing recently. that you
56:53can guys can add that and you just go
56:54add a step LinkedIn and you do the same
56:57thing as we've done here but you change
56:59this to LinkedIn post get user post so
57:02that may be a cool thing for you guys to
57:03add the uh functionality on at the end
57:05is a bit of a challenge you can pause
57:06this video and do that and then we need
57:07a llm step to take this again I'm going
57:10to grab a pre-written prompt that I did
57:12just to save some time going to drop
57:13this in here says fairly similar stuff
57:17um LinkedIn data I'm going to put all
57:21the data in
57:23there. And then we're going to use a
57:25GPT4 mini again. And we can give that a
57:28run because we've already got this data
57:30queued up here. And there we go. We've
57:31got a nice summary. Um, if I change this
57:34to nice and formatted. So there's a
57:36little button down there between raw or
57:38formatted. And you can copy the stuff
57:39out of here, of course. So Liam Mley, my
57:41followers. Damn, I got a lot more than I
57:43thought. So there we have the summary,
57:44my name, where I'm based, uh,
57:46information, my career experience. Super
57:48handy stuff. And this is going to be
57:50super helpful in the next step when we
57:51generate that pre-core report. So very
57:53quickly, we've created one more tool.
57:55I'm going to save this. So now we've
57:56got, if we go back to our Figma, we have
57:59two of these done. Now the final one is
58:02going to take in the company and
58:03prospect research and generate a
58:05pre-call report. So this one's going to
58:06be a little bit different. If I go
58:08create a new
58:11tool, preall report
58:16tool. Okay, free call takes in company
58:19and prospect summary and generates a
58:21free call report for sales
58:24direct. And now for the inputs for this,
58:26we need long text, not just normal text
58:28because we're going to be taking in that
58:29big company and uh and prospect
58:31research. So we go prospect summary
58:34summary of the prospect based on length
58:37and profile
58:41that and a prospect in this case is
58:43someone who's a potential customer. Just
58:45to clarify that if you're new to
58:47business and don't really get these
58:48terms. Company summary, uh,
58:56summary, right? We have our prospect
58:58summary and company summary in there. I
59:00hope you're following along. Next step
59:01is just an LLM step and we want to
59:03combine these two together. Again, I've
59:04got a little handy prompt for this to
59:06save us time. Now, in this case, you
59:07will see that the prompt is a bit
59:08bigger, right? So this is um for more
59:11important parts when you are creating
59:12tools whether it's it's for agents or
59:14just generally when you're using prompt
59:15engineering and LLMs to create value. In
59:17this case we are creating value because
59:19in this case we're taking in this
59:20prospect summary and this company
59:22summary. We're also giving it the the
59:24context of this fantasy or or
59:26hypothetical business that we are
59:28selling this agent to as like a co-pilot
59:30system. We've got Big Boy Recruits which
59:32is Dallas based recruitment firm
59:33specializing in software industry talent
59:35acquisition for SMBs. You're going to
59:37see this kind of recur across the
59:38different projects we do, but basically
59:39we're helping these big boy recruits to
59:41automate their business with AI. So,
59:43this takes in some context on that
59:45business and it's going to generate a a
59:47report that's going to help the sales
59:48rep say, "Okay, this is the company.
59:49This is what we sell. This is what we
59:51specialize in. This is the company that
59:52we're trying to sell to. This is who
59:54we're going to be talking to. What's
59:55some how can I personalize this call or
59:57what's the strategy I can go into this
59:58with? What are some angles that I can
1:00:00attack this call from?" And so to do
1:00:02this, I have a prompt writing tool that
1:00:04I use quite regularly and my team uses
1:00:06it as well. Um, perfect
1:00:09prompt. So, this tool does a lot of leg
1:00:11work for myself and my team all the
1:00:13time. I'm going to give it to you guys
1:00:14to use for free. You'll be able to clone
1:00:15it into your relevance account.
1:00:17Basically, what I'll usually do is I'll
1:00:18put on uh the dictation thing. You've
1:00:20got it on one of these keyboards. You've
1:00:22got this little thing. basically
1:00:24whatever on your computer allows you to
1:00:26speak into the computer and it takes in
1:00:28your voice and transcribes it into into
1:00:30text on the screen. I'll press that and
1:00:32then I'll explain as you can see here
1:00:33what is this prompt doing and why and
1:00:35I'll go this prompt needs to do this
1:00:36this this is going to take in this
1:00:37information it's going to do this the
1:00:39reason we're doing this is this this and
1:00:40I'll do like a big big body of text in
1:00:43there and then the next one if I have
1:00:44them I'll give some good examples of
1:00:46input and output pairs of how I want it
1:00:48to take in data and how I want it to
1:00:49spit it out. If you give it both of
1:00:51these and you hit run, it's going to
1:00:52print you out using the researchbacked
1:00:54prompting techniques that we use at
1:00:55Morningside. It's all crammed in here.
1:00:57There's a video that I recommend all of
1:00:59you watch. It's going to be in the free
1:01:00course anyway on school. So, when you
1:01:01get in there and watch my prompt
1:01:02engineering guide, um, this is basically
1:01:04the entire information of that prompt
1:01:05engineering guide smashed into this LLM
1:01:08step here. So, when you pass this
1:01:09information in, it applies all of that
1:01:10and it gives you out a prompt that is
1:01:12fully researched back and performs very,
1:01:14very well right out of the box. So,
1:01:15that's a little bit of extra value I
1:01:16wanted to throw in there for you guys.
1:01:17this is going to be available on the
1:01:18school um with the rest of the resources
1:01:20as well. So basically I put in the
1:01:22information here about what this
1:01:23particular uh task was. I said it's
1:01:26going to take in the prospect
1:01:27information. It's going to take in the
1:01:28prospect summary and the company
1:01:29information and it spat out this
1:01:31basically first go and I just had to
1:01:33insert these variables. So this prompt
1:01:34and everything else will be on the on
1:01:35the score resources as well. So in this
1:01:37case because we are doing a bit of
1:01:38strategy and sort of high level thinking
1:01:40rather than just summarizing and we may
1:01:41want to change the model here to
1:01:43something a bit smarter. We could go to
1:01:4403 mini which is one of the later ones.
1:01:46Um, again, when you're watching this, it
1:01:48might be 06 or 010 or whatever the hell
1:01:50they come up with next, but there's
1:01:52probably going to be some much better
1:01:53models. So, just use a smart one because
1:01:54it's really strategizing on how big boy
1:01:57recruits can position themselves for
1:01:58this call. So, enough of me yapping
1:02:00about that. Let me grab some inputs for
1:02:02this and we can test
1:02:05[Music]
1:02:06it. Shout out muscle. All right. So,
1:02:08I've got this information here about
1:02:10myself, my LinkedIn profile, and my
1:02:12company. And so again, remember that
1:02:13this is for Big Boy Recruits, a Dallas
1:02:15recruitment firm specialized in software
1:02:16industry talent. So it's going to look
1:02:18at my company, Morningside AI. It's
1:02:20going to look at me and my background
1:02:21and my profile on LinkedIn. Then it's
1:02:23going to spin, as we see here, um,
1:02:25review this, analyze this, map big boys
1:02:27unique value proposition, ra, and it's
1:02:29going to try and create a report that's
1:02:32going to allow the sales rep to sell me
1:02:33or close me better on there or find some
1:02:35angles to sell to at least. So if we go
1:02:37run tool, and now you see that I'm using
1:02:38a lot of just basic web scraping and LLM
1:02:40steps. I just want to show you guys the
1:02:41basics. The thing is tools can get very
1:02:43very advanced when you have like CRM you
1:02:45want to integrate into, but relevance
1:02:47allows you to do all of that. It's just
1:02:48within the scope of tutorial, it can be
1:02:50pretty difficult to be pulling
1:02:51information from all over the here cuz I
1:02:52have to set up a database, show you guys
1:02:54how to do it, too. So, this this keeps
1:02:56it quite confined, but it still gives
1:02:57you a good taste of it. So, if we look
1:02:59at this view, all key business
1:03:01challenges and opportunity, Morningside
1:03:03AI, this Liam's profile gives me a bit
1:03:05of a rundown of this mapping big boy
1:03:07recruits unique value proposition. Maybe
1:03:08it's going to be better if I change this
1:03:10to the format. There we go. Talks about
1:03:13mapping big boys recruits, strategic
1:03:15talking points. I've been following your
1:03:17journey of digital marketing, AI, ra um
1:03:20dive into opportunity. I work at big
1:03:21boys. This assist, and it's even gone
1:03:23and done a section on anticipated
1:03:24objections. So, the idea is that the
1:03:26sales rep is going to have a skimmer of
1:03:27this before the call, which ties into
1:03:29the value that I've listed here, which
1:03:31I'm going to do for all of these builds,
1:03:32by the way, which comes down to
1:03:33ultimately a better prepared sales rep
1:03:35should close more deals, right? if they
1:03:37know more about the prospect and the
1:03:38company and you have an angle to try
1:03:40sell through or suggestions at least. It
1:03:42should increase the conversion rate of
1:03:43the sales team. So, we've built this
1:03:44tool. We can change this to green now.
1:03:47And the final step is going to be
1:03:48heading over to our agent builder within
1:03:50relevance. I'm going to save this. If
1:03:52you pop over in the left panel here, you
1:03:54can go into our agents. And what we want
1:03:56to do is create a new
1:03:59agent. We're going to call it our sales
1:04:03co-pilot.
1:04:05Um, big
1:04:07boy big boy sales co-pilot. Sorry, I
1:04:11got, like I said, I got to have fun with
1:04:12the stuff where I go kind of insane. Um,
1:04:14this this agent is
1:04:18our sales co-pilot that helps reps to be
1:04:25better prepared for sales
1:04:28[Music]
1:04:30calls. Triggers, we don't need to do any
1:04:32of that. We go to core instructions. I'm
1:04:34going to again paste in some of the
1:04:35stuff that I've prepped earlier. So, if
1:04:36I paste this in here, you'll see that
1:04:37it's structured fairly similarly to the
1:04:39prompt that we just did before for the
1:04:41uh pre-core report generator. And this
1:04:43is again using another tool that I've
1:04:44created um for AI agent prompting. Um
1:04:47so, it's fairly similar stuff that I I
1:04:49include in that other prompting tool,
1:04:50but the agents is slightly different um
1:04:53to just regular LLM steps within
1:04:55different tools and workflow
1:04:56automations. So, I will include this as
1:04:58well. It's my AI agent perfect prompt
1:05:00generator and it's fairly
1:05:01straightforward to use. I'll include
1:05:02that in there as well. But basically,
1:05:04you put in all the information about
1:05:05what the agent is, why it's doing it,
1:05:06the different tools that you're
1:05:07connecting to it, and then it prints out
1:05:09this for you. So, I'll just run through
1:05:10it. We've got a role here telling it who
1:05:11it is, um, and kind of hyping it up and
1:05:13saying how good it is. Explains the
1:05:15task, um, talking about how it's helping
1:05:16to conduct detailed research on
1:05:18companies and prospects. Um, some
1:05:20specifics. Uh, don't need to worry about
1:05:21those too much. Just reiterating the
1:05:23task. And now here we can enter in the
1:05:25references to tools. So if we go slash
1:05:27tool and see in order to get the agent
1:05:29to function as well as possible, we need
1:05:31to tell it what tools it has available.
1:05:33This is really key across all the agents
1:05:35you build, especially if they're more
1:05:36conversational. You need to explain to
1:05:38them what tools they have and how and
1:05:40when they should use them. So if I go,
1:05:41you have three powerful tools. Research
1:05:45company. Um, yep, that's right. Purpose,
1:05:48input, company URL, use when needing to
1:05:50gather company information. This one of
1:05:52course is the uh
1:05:56prospect. And then this last one is our
1:06:01[Music]
1:06:02pre. There we go. So that's all whacked
1:06:04in there. Don't need to worry about that
1:06:05too much. But again, this will be
1:06:07included um this prompt and everything
1:06:08if you want to follow along and also if
1:06:10you just want to clone the whole agent
1:06:11um and use it in your own business or
1:06:13sell it or whatever you want to do. We
1:06:14also get to select the model here. Um
1:06:16I'm just going to keep it as GPT4 mini.
1:06:18I like some pretty fast responses here
1:06:19because agents as someone's using it, it
1:06:21can feel really irritating if it's not
1:06:23responding quickly. So, we've got all
1:06:25that built out. That's the core
1:06:26instructions in the prompt of the agent.
1:06:28Um, we can go down to the uh tools
1:06:31section. It's got all the tools
1:06:33connected in here because we mentioned
1:06:34them in the prompt. And we can just go
1:06:36through and do some quick settings on
1:06:37here. Um, I don't want to have to do an
1:06:39approval for it. Some tools you can say,
1:06:41look, they've got to give it a thumbs up
1:06:42before it can actually uh trigger it.
1:06:44Um, prompt for how to use. Just some
1:06:46quick descriptions we can pop in here.
1:06:48I'm not sure why relevance doesn't carry
1:06:49that over from the tool. I guess they're
1:06:50asking us to do it again for some
1:06:52reason. Um when you need to research a
1:06:54uh
1:06:55prospect linked and URL, we're going to
1:07:00say this is auto run as well. Um and
1:07:03then preall change this to auto run as
1:07:05well.
1:07:07Use this when you need to generate a pre
1:07:11call report from the company and
1:07:15prospect
1:07:18research. All righty. Um, and there's
1:07:20all sorts of other cool stuff.
1:07:21Relevance, as you can see, is like
1:07:23abilities, sub aents, metadata, extra
1:07:24stuff that you can build onto. Um, but I
1:07:26just want to get you guys started with
1:07:27the core of this. Um, all right. So, we
1:07:30can confirm
1:07:33that. And boom, we have our agent here
1:07:36ready to go. So this is where you can
1:07:37test your agents and use them if you
1:07:38want to. But in this case, I'm just
1:07:40going to give it a quick rundown and say
1:07:42see if the functionality is working as
1:07:43we as we planned. Um, hi, I am getting
1:07:47on a call with Liam Otley from morning
1:07:53side
1:07:55AI. Here uh has
1:07:59details or we can just go
1:08:06[Music]
1:08:15lenol report please to prep for the
1:08:19call. Sending your task to big boy sales
1:08:22copilot. And then we get to see all the
1:08:24debug and how it's actually walking
1:08:26through these different steps. Oh, let's
1:08:28see what he
1:08:31does. Yep. Okay, that's great. It's
1:08:34using the research company as we wanted
1:08:36it
1:08:37to. Should add in one more step there to
1:08:40research the prospect as
1:08:42well. There we go. Using the second
1:08:44tool. It is pretty satisfying when this
1:08:46stuff works. And this is just a really
1:08:48basic one, guys. I don't want you to
1:08:49think this is like, oh, well, that's
1:08:51pretty underwhelming, Liam. I'm trying
1:08:52to teach you the basics so that you can
1:08:54actually build on top of this. So if you
1:08:56get the bug, if you get the like you get
1:08:58a travel bug, if you get the agent bug
1:09:00and you see the stuff and you're really
1:09:01interested. Oh, look, there it is. Now
1:09:03it's filling out the prospect summary as
1:09:04the inputs. Surely we don't have to
1:09:06watch it do that. It's going to take a
1:09:09while.
1:09:11[Music]
1:09:12Um, I really write that out word for
1:09:15word like that. But when you start to
1:09:16see this magic and you add in other cool
1:09:18tools and functionality, you test it on
1:09:20yourself. You can build like things for
1:09:21maybe you want to do content, you make
1:09:22yourself a little content co-pilot,
1:09:25etc. There we go. To use all the tools
1:09:28and it should be spitting back and bam,
1:09:29there we go. So, I hope that was worth
1:09:31the wait. Let's go through it now.
1:09:32Here's your comprehensive precore report
1:09:34for your upcoming conversation with Liam
1:09:35Mley from Morningside AI. Pre-core
1:09:37report. Talent acquisition under
1:09:39pressure. Morningside AI operates in a
1:09:40highly competitive AI and tech market as
1:09:42they scale. Finding specialized talent,
1:09:44engineers, data scientists, AI with a
1:09:46proven record can be challenging. That's
1:09:49scarily accurate because that is
1:09:50literally one of the biggest constraints
1:09:51that we have had to scaling Morningside
1:09:53long-term is that it's just really
1:09:54really hard even with my channel. It's
1:09:56so hard to find the right people and get
1:09:58them to commit as as developers as well.
1:10:00So if you want to build a very big
1:10:01general AI development firm, an AI
1:10:03automation agency, you need the best
1:10:05talent and you need to get a lot of it
1:10:06in um so that you can scale up. So
1:10:08that's bloody spot on. Obviously this
1:10:10thing knows that's a good angle to sell
1:10:12through. Um but yeah, prospect analysis.
1:10:14Liam's a dynamic entrepreneur and
1:10:16thought
1:10:17leader with a robust background in
1:10:19e-commerce, digital marketing and AI.
1:10:21His journey reflects a passion for
1:10:22innovation and commitment to continuous
1:10:24learning. Again, that's pretty bloody
1:10:26spot on um hands-on experience. So there
1:10:29you go. That is the big boy sales
1:10:31co-pilot for big boy recruits. The cool
1:10:33thing you can do now once you have built
1:10:34this um is you can go share um there's a
1:10:38chat UI which I'm going to turn on now.
1:10:40There are chat widgets so you can put
1:10:41them on websites and stuff like that.
1:10:43What I want to do is just pull this up
1:10:44because this is what you'd be giving to
1:10:46your client likely or if you guys are
1:10:48going to start selling these to
1:10:49businesses which again we're touching on
1:10:50selling in the last section of this
1:10:52video. So how do you turn these into
1:10:53into a business and start making money
1:10:55from it and selling these as a as a
1:10:56service and building these businesses
1:10:58which is really where the money's money
1:10:59is made. So there you go. This URL you
1:11:00can obviously send to your client. If
1:11:02you're building it for your own team,
1:11:03you can send this to your team and say,
1:11:04"Hey, pin this because you're going to
1:11:06be able to use it. Add more
1:11:07functionality into it, etc. That is how
1:11:09you build a co-pilot on relevance AI.
1:11:16All righty, that is build number one out
1:11:18of the way and we are jumping into AI
1:11:20agent build number two, which if you're
1:11:22listening closely at the start of the
1:11:23section, we are talking about an
1:11:26NATbased inbound lead qualification
1:11:28agent that's going to be doing some
1:11:29pretty cool stuff for us, which is a
1:11:31really important function within a
1:11:32business um around lead qualification.
1:11:34Um, so this is a really cool one. Again,
1:11:36in this case, this is what we call an
1:11:37automated agent, not a conversational
1:11:38one. What we just built is a
1:11:40conversational agent. humans are
1:11:41directly talking to it and chatting back
1:11:43and forth and using it and we are
1:11:44operating it ourselves. In this case, as
1:11:46you can see on this little flow uh flow
1:11:48diagram here, this is a screenshot from
1:11:49the final product. We are actually
1:11:51baking this into a workflow that's going
1:11:53to be triggered on a form submission.
1:11:55We're going to do some research and then
1:11:56we're going to use the handy AI agent
1:11:58and tools agent within NA10 to trigger
1:12:00another workflow and then send some
1:12:02emails off. So, this capability of using
1:12:04AI agents in workflow automation really
1:12:06expands the possibilities of what you
1:12:08can build. And the software NAN that I'm
1:12:10going to teach you how to use is really
1:12:11at the the cutting edge and leading the
1:12:13charge when it comes to these automated
1:12:14uh AI agent workflows. And just quickly,
1:12:16it's good that we walk through the
1:12:17purpose and the value behind this
1:12:18automation so that we know why we're
1:12:19doing it. Right? So this inbound lead
1:12:21qualification use case is based on the
1:12:22fact that companies who market
1:12:23themselves well soon have far too many
1:12:25people reaching out to them. Many of
1:12:27which are not a good fit or what you'd
1:12:28call qualified for what they sell. Eg
1:12:31they're too small or they're not the
1:12:32right industry. Like you you have a
1:12:33business and they say we only help XYZ
1:12:36kinds of businesses. And if leads come
1:12:37to that business who are not qualified,
1:12:38then they obviously don't want to be
1:12:40taking calls or or doing anything
1:12:41further with them. So this process of
1:12:43researching a new lead and deciding
1:12:44whether or not to take a call is known
1:12:46as qualification, which is what this
1:12:48agent aims to automate. So the value
1:12:50here is that instead of having to pay
1:12:51someone to manually qualify and go
1:12:52through all of these leads or using
1:12:54arbitrary rules, which is what some
1:12:56businesses have to go to, it's like
1:12:57look, oh look, we've got so many leads.
1:12:59Let's just say if they don't want if
1:13:00they say that they're not on this, then
1:13:01we'll just cut them out. And that's
1:13:02potentially leaving money on the table
1:13:04by cutting out leads who would have
1:13:05actually been a good deal, but the rules
1:13:07kind of didn't see enough detail to be
1:13:09able to determine if they're a good fit
1:13:10or not. So, this automation is
1:13:12essentially immediately qualifying and
1:13:14triggering the next steps to the sales
1:13:15team and allowing them to do that human
1:13:17style research on these leads at scale.
1:13:19So, enough talking. Here's a little bit
1:13:21more information. Again, with all of
1:13:22these, it's going to be on the figure on
1:13:23the school. Then, I've also broken down
1:13:24how this agent would actually operate in
1:13:26the real world and sort of real time. A
1:13:28lead's going to be submitting the form,
1:13:30which is going to be this here. um the
1:13:32relevance company AI researcher. So
1:13:33something that we've just built in
1:13:34relevance, we're going to reuse in here,
1:13:36which is handy that you can start to
1:13:37move these components around and see how
1:13:38they can fit into different automation
1:13:39platforms. Um then the AI agent is going
1:13:42to look at this information from the
1:13:43research. Then it's going to determine
1:13:44based on a qualification criteria we
1:13:46give them um inside the prompt of this
1:13:48agent whether they are qualified or not.
1:13:50If they are qualified, it's going to uh
1:13:53use this tool here and call this second
1:13:55workflow. when we're going to
1:13:56essentially analyze that further and do
1:13:58a notification to either our agency team
1:14:00or our our SAS team and if they are not
1:14:03qualified we will use this tool here
1:14:05which will just send an email straight
1:14:06back to the person who submitted the
1:14:07form say hey sorry we're not open to
1:14:09working with businesses like yourself at
1:14:10the moment let us know if we can help
1:14:12you any other way so that's a rough
1:14:14rundown of the build let's jump into it
1:14:15so to kick things off of course you need
1:14:16a platform on natn.io
1:14:20io. All links and resources will of
1:14:21course be on the school uh post for this
1:14:23video. And once you're on this page, you
1:14:24can go to get started and you can create
1:14:27an account for free and just go through
1:14:29the setup process that they do. I'm sure
1:14:30you can figure that out. They do have a
1:14:3214-day free trial as of this filming. So
1:14:34that's very good if you're just jumping
1:14:35in, not having to pay anything. And they
1:14:36give you quite a lot of usage up here.
1:14:37As you can see, 1,000 executions. So
1:14:39what we're going to do, of course, is
1:14:40click on create workflow up here. I will
1:14:42be giving you the template. So if you
1:14:43want to just import it, you can. That
1:14:45will be on the Figma there. But just
1:14:46like the last tutorial, I'm actually
1:14:47going to be showing you the process of
1:14:49building these up from scratch that you
1:14:50can see how I how you go through the
1:14:52process of building these automations
1:14:54and the the testing and back and forth
1:14:56you need to do in order to get to the
1:14:57end result. That's probably actually a
1:14:58lot more important if you are to go out
1:15:00if I'm trying to teach you to fish, not
1:15:01just give you a fish, is to see how I
1:15:03deal with problems when we're building
1:15:04these. So, let's get started by starting
1:15:06off our trigger. We're going to go
1:15:09form has a nice
1:15:12form on new inn form event. And here we
1:15:15get to create an N8N form. So we can
1:15:17call it uh work with
1:15:21us. Provide
1:15:26your. So now we get to pick the field
1:15:29names. So we want to have the first one
1:15:31which is what
1:15:35is make that a required field. We add
1:15:37another one. What is your company
1:15:40website?
1:15:42DG HTTP this morning. Require
1:15:51field. Right. So, we've just built out a
1:15:53basic form here with first name, company
1:15:55website, which we're going to need for
1:15:56the next step. I've put a placeholder in
1:15:58here so they know that it needs to have
1:15:59https um at the front of it. And I'm
1:16:02asking for them to provide some
1:16:03information about your inquiry like and
1:16:05maybe you can say what can we help
1:16:07you? Um and that's going to be text
1:16:10area. So, they've got a bit more room.
1:16:11So, we can go test step here. Make sure
1:16:12that it's all looking nice. This is what
1:16:14the form's going to look like. What's
1:16:15your name?
1:16:20Obby. There we go. We've submitted that
1:16:22form. If you go back, there we go. We
1:16:24have the data in. So, this shows you the
1:16:26output. NAM works by having kind of this
1:16:28middle island here, which is what you
1:16:29set up. And then the left side is the
1:16:31input and the right side is the output.
1:16:33So, we've tested it and these are the
1:16:34outputs that our form is giving, which
1:16:36is what we are looking for. And the next
1:16:38step in this lead qualification process
1:16:39is to do some research on the lead. So
1:16:42we have the company URL and this is when
1:16:44we're going to go and make an HTTP
1:16:46request. So this is basically calling
1:16:48any API over the internet. And in order
1:16:50to set this up, we actually need to go
1:16:51back to
1:16:55relevance and we're going to find the
1:16:57company researcher tool that we made
1:16:58with relevance. And now you're going to
1:16:59start to see how this all fits together.
1:17:01um that building tools and relevance can
1:17:03also be very useful and an extremely
1:17:05useful skill in all areas of AI
1:17:07automation because now I can come on to
1:17:09research company and not only can I use
1:17:11it here this I mean this is why I think
1:17:13relevance is such a great platform um
1:17:15you have the use here so I can send this
1:17:17across to a client I can share it with
1:17:18it as I showed before I can use it just
1:17:20here myself I can run it in bulk on a
1:17:22spreadsheet or but more importantly in
1:17:24this case I can go to the API and now I
1:17:27can call this might look scary just
1:17:29don't worry I can call this
1:17:31functionality basically send in a
1:17:33company URL and get back the research. I
1:17:35can access this over the internet
1:17:36through an API and they give me it here
1:17:38and they tell me exactly how to call it.
1:17:40So now we this is on actually a post
1:17:41request. So we can copy this. Remember
1:17:44how we talked about get and post
1:17:46request. This is a post request because
1:17:47we're posting some data to relevant AI.
1:17:49So we change this method to post. We put
1:17:52the URL in here. We do not need
1:17:54authentication in this case but you can
1:17:56turn it on. So you can make a private
1:17:57here and then you can have an
1:17:58authenticated. might sound a bit
1:17:59complicated but for now don't worry
1:18:00about it we don't need to have an
1:18:02authentication step on
1:18:04this and then if we look at the request
1:18:07body here it tells us how we can send
1:18:09data to relevant AI and if we go copy
1:18:13here come back we're going to send a
1:18:16body it's going to be in
1:18:20JSON and we can change this to using
1:18:23JSON and then we can paste in basically
1:18:26what we have been given from relevant AI
1:18:27now this looks a messy. Let's pop this
1:18:29open a bit
1:18:31more. And we actually need to change it
1:18:33to an expression here. So fixed means
1:18:35that we're not accepting any dynamic
1:18:37data in. So when the form is submitted,
1:18:39we actually need to take in some data
1:18:40from that form which we have over here.
1:18:42We need to inject it into this uh HTTP
1:18:44request to relevant AI to get this
1:18:46company research done. So we can't have
1:18:48it as a fixed um JSON body here. Need to
1:18:51change it to expression and then we can
1:18:53pop this out here and it gets a bit
1:18:54easier. So, we have params. We have the
1:18:56company URL. And then we need to pop in
1:18:59here. Oh, the company
1:19:03URL. Pop that in there. And here on the
1:19:05right side, we get to see what that
1:19:06would look like given the test data that
1:19:08we've just put through. So, you can see
1:19:10I've got the company URL, https
1:19:12morningside.ai in quotations here, which
1:19:13is what we want. So, we can go back now.
1:19:15And now we can give it a test to see if
1:19:17it's going to be able to communicate
1:19:18with relevant AI and get us the data we
1:19:21want. There we go. We have our result
1:19:24back from relevant AI which is the
1:19:26summary. As you can see when we go back
1:19:28to this um and back to build this is
1:19:31exactly what we'd expect. You know you
1:19:33put in the URL does this scrape a fire
1:19:35call writes the summary spits out the
1:19:37summary and this summary is what we're
1:19:38getting back out over here which is what
1:19:40we want. So bank that's great another
1:19:42step
1:19:43done. And those of you who are a bit
1:19:45confused about what this is this is just
1:19:46the body of the request. So because we
1:19:48are sending a post request remember how
1:19:50we have get and post request. Get
1:19:51requests are typically just with a URL
1:19:53and with a bunch of stuff tacked on. A
1:19:54post request, we need to send a a JSON
1:19:57body like this. And it does look quite
1:19:59confusing, but if I take this and go
1:20:01JSON
1:20:06form, paste this in and beautify it. And
1:20:09you can see how we have basically the
1:20:10project which is the relevant project
1:20:12that we are calling. And this tells the
1:20:14API this is the project that we want to
1:20:15interact with. and the project expects
1:20:18the params, the inputs, which is the
1:20:20company URL, and we're injecting that
1:20:22company URL from our information here in
1:20:25NA10 that we've dragged across. It might
1:20:26seem tricky a few times, but trust me,
1:20:28this stuff becomes like riding a bike
1:20:29once you get up and running. So, um, a
1:20:31few more of these and you'll be you'll
1:20:32be completely fine. All right. And so
1:20:34now that we have our company
1:20:34information, the next step is going to
1:20:36be setting up the agent, which is really
1:20:38the coolest part in my opinion about
1:20:39make right now is that we come in here
1:20:41and we can click on um agent and we can
1:20:44set it to as a tools agent, which means
1:20:46we connect our own tools. And if we just
1:20:48back out of
1:20:49this, we can do cool things like set up
1:20:52the model. And right here, this is so
1:20:53cool because we get to see exactly what
1:20:55we were just talking about earlier in
1:20:57this video, but we have the different
1:20:59parts of an agent, the different
1:21:00ingredients, right? So here's the chat
1:21:01model. This is the brain. This is the
1:21:03LLM that's going to be powering the
1:21:04whole agent using the soup example. Like
1:21:06this is the one meat that we get to
1:21:08choose, right? This is a specific model.
1:21:10In this case, we going to be using
1:21:11OpenAI again because you already have
1:21:13your API key and I can't really be
1:21:14bothered going and showing you a whole
1:21:15another provider, but it's the same
1:21:16process for all of them. You can if you
1:21:18want anthropic, you can then pick all
1:21:20the anthropic models. Uh but in this
1:21:22case, just to keep it simple, let's just
1:21:23go back
1:21:26to and open AI model.
1:21:30And if we go back again, you can see
1:21:32that we now have uh the memory and the
1:21:34tools that we can connect. So, of
1:21:35course, we have the tools, which we've
1:21:37talked about a lot in this video
1:21:38already. We can connect multiple
1:21:39different tools here, as we're going to
1:21:40do in a second. And then we also have
1:21:41memory set up here, which is a little
1:21:43bit outside of the scope of this video.
1:21:44As I said, in most platforms, it comes
1:21:46builtin, but NAT is a bit more of a
1:21:47developercentric platform. So, if you
1:21:49wanted to play around with memory, um
1:21:50different forms of of managing memory,
1:21:52you can do that here. In our case, we're
1:21:53not going to be touching it. And in
1:21:55order to connect a knowledge base, if we
1:21:57wanted to set up a knowledge base for
1:21:58our agent, we would connect it as a
1:22:00tool. So you can see here we've got
1:22:01in-memory vector store, pine cone vector
1:22:03stores, etc. These are vector databases
1:22:05that we can connect just like in the
1:22:07other tutorials we're going to do in
1:22:08this video. When you upload some
1:22:09documents to make a knowledge base,
1:22:10they're essentially being put in a
1:22:12vector store like these, but the
1:22:13platforms manage it for you and make it
1:22:15a lot easier to do. So in this case,
1:22:16we're just going to be doing two
1:22:18different tools. Firstly, we're going to
1:22:19be calling an NATM workflow. So I'm just
1:22:21going to finish off the the basic setup
1:22:23of this. Um, and then we're going to add
1:22:24another tool on. It's going to be the
1:22:26Gmail
1:22:28tool. And then before you know it, we
1:22:30have our AI agent structure built out.
1:22:32So, we're using the Open AI models.
1:22:34We're going to pick the model shortly.
1:22:36We're going to be using the tool to call
1:22:37the second NATM workflow, which is going
1:22:39to be uh triggering the the email
1:22:41notifications for our sales reps and the
1:22:43classification of the uh of the lead.
1:22:45That's going to make a bit more sense
1:22:46when we actually do it. So, just stick
1:22:48with me on this. And then this Gmail is
1:22:49going to be sending back a hey sorry you
1:22:52didn't qualify for what we do um sorry
1:22:55we can't help you let us know if we can
1:22:56do anything else for so to start setting
1:22:58things up we can start from left to
1:22:59right here with the chat model um the
1:23:00openai model that we want to use and you
1:23:02need to set up your openai account so
1:23:04you can click create new credential and
1:23:05you need to go and add in your API key
1:23:07here I'd suggest you go and make a new
1:23:08one on platform openai.com and you can
1:23:11add a new one in here and you can name
1:23:12the key nat so you start to know which
1:23:14keys are used for which different
1:23:16platforms um and once you put that in
1:23:18there you can Just click save and then
1:23:19it will run a little test and then
1:23:20you're ready to go. You should have this
1:23:21set up here. Then we get to select the
1:23:23model that we want to use for our agent
1:23:24here. And in this case, I want to go for
1:23:26something quite smart. So, I'm going to
1:23:27go for
1:23:2803. We have 03 mini here. This one
1:23:31appears to be a little bit more recent.
1:23:32So, they'll sometimes put the dates on
1:23:34the end of it. And if you just look at
1:23:35the current date, you can kind of see
1:23:36how how close to the current date it is.
1:23:38Um, but I'd say this one's a bit more
1:23:39recent, so it's going to be hopefully a
1:23:40bit better. And next, we're going to
1:23:42skip this cuz we need to set up a whole
1:23:43different workflow to connect it to. Um,
1:23:45we're going to jump straight into this
1:23:46Gmail one. You need to set up again
1:23:47another connection as you go through all
1:23:49of these different automation platforms.
1:23:50You do need to do these these
1:23:52connections between uh your own account,
1:23:54say your Gmail or your calendar or all
1:23:56these different apps. You need to go and
1:23:57create a new credential. Um and you can
1:23:59just do the sign in with Google here.
1:24:01Super easy to do. I'm sure you don't
1:24:02need my help with that. Once you've set
1:24:04up that connection, you can close this
1:24:05and you will see the connection that you
1:24:07set up here. Basically, that's what
1:24:08we're going to be sending emails
1:24:09through. And then we can get into
1:24:10setting this up. So, because this Gmail
1:24:12tool is going to be used to send a a
1:24:14reply back to the person who submitted
1:24:15the form and say, "Hey, sorry, you're
1:24:17not a good fit for us." You want to send
1:24:18it back to the person who submitted the
1:24:19form. Now, if you scroll down here, uh,
1:24:21yep, you see that I've I've forgotten to
1:24:24add the email form in. So, this is a
1:24:25good example of needing to go back a
1:24:27little bit. So, we can go back to the
1:24:28form submission here. Um, scroll down,
1:24:31say, what is your
1:24:35email? And we can set it as an email.
1:24:37So, it's going to automatically force
1:24:39them to provide a valid email for us.
1:24:41And I want to maybe shimmy this up a
1:24:42bit.
1:24:45Um, so it's name, email, then company,
1:24:49website. Um, we can do another test here
1:24:52just to give it some proper
1:24:55data. Oh, it's not going to let us do
1:24:57that because this isn't set up properly.
1:24:58So, we can just delete that for
1:25:01now. And we're actually going to delete
1:25:03that. Otherwise, it's going to be a bit
1:25:04of a pain.
1:25:06So, we want to test it
1:25:13again. So, we can submit another form
1:25:15here and go back to NA10. And there we
1:25:19go. We have the information. We have the
1:25:20email now. That's great. And so, now we
1:25:22can come and set up our tools again.
1:25:23We've got the workflow there and we've
1:25:25got the
1:25:28Gmail. Um, and now we have our
1:25:30connection set up. Yep.
1:25:36Oh, and we haven't got the data here
1:25:38because we need to run this
1:25:41again and get the research. So, now the
1:25:43research is done in there for us to set
1:25:44up the Gmail tool here. We can go to two
1:25:47and we'll be able to pull in the email.
1:25:49So, again, they submit this form. We
1:25:51realize that they're not a qualified uh
1:25:52person for our offer. And then we're
1:25:54going to send an email back and say,
1:25:55"Hey, sorry, you're not a good fit." So,
1:25:56we can say the subject here is um thanks
1:25:59for your interest. I'm going to change
1:26:02the email type to text here and I'm
1:26:04going to write a basic message in. I'm
1:26:05just going to snag it from the one that
1:26:07I've done previously. So, we need to
1:26:08change this from fix to an expression
1:26:10because we want to be pulling in their
1:26:11name here. So, I'm going to paste in
1:26:13what I have here. Again, this would be
1:26:14included um in the resources. It's just
1:26:16going to save us time if I don't have to
1:26:17type this all out manually. Um but you
1:26:19can see here, I'll just delete this so
1:26:20you guys know what we're doing. If I go
1:26:23hi, or hi, and then we can add in what
1:26:26is your first name? Hi, name. In this
1:26:29case, it's going to be filling in my
1:26:30name here as an example. So, highly mly,
1:26:32thank you for your interest in big boy
1:26:33recruitment services as you specialize
1:26:35in recruitment for software and
1:26:36development agencies. We're not a good
1:26:38fit based on your company's industry.
1:26:39Please let us know if you'd like to
1:26:40connect us with one of your partners who
1:26:41specializes in dealing with your needs.
1:26:43Cheers. Huge Jackman here to sales, big
1:26:46boy recruits, BBR, uh, Dallas, Texas.
1:26:48So, if you guys remember huge Jackman,
1:26:50comment down below um for the OG fans.
1:26:52And then that is our Gmail all set up.
1:26:54And just quickly so that you've got a
1:26:56bit more knowledge around how this Gmail
1:26:57uh tool works, we have all these
1:26:59different steps that we can use. We're
1:27:00using the send one. So it's sending an
1:27:02email. You can use reply, you can use
1:27:04get, delete, all these other functions,
1:27:05but the easiest one and the most common
1:27:07one you're going to use is going to be
1:27:08that send one, of course. Now, we need
1:27:09to set up this NATM workflow which the
1:27:11agent is going to call as a tool um when
1:27:13they are a qualified prospect. So I'm
1:27:15going to delete this one and just save
1:27:16this for now. Then we're going to go
1:27:18back
1:27:20to home. We're going to create a new
1:27:23workflow.
1:27:25And this is a really cool skill that I
1:27:26want to teach you. The fact that you can
1:27:27build all these workflows and then
1:27:28connect them to agents and it can just
1:27:30be taking data in and kind of shooting
1:27:32it off in all directions and triggering
1:27:33all these complex multi-step processes
1:27:36because it's a super valuable way of
1:27:37using agents. Um, so I really want to
1:27:39teach you that. And obviously this one's
1:27:40going to be starting off a bit different
1:27:42to the other one. We actually want this
1:27:43to be set up as when executed by another
1:27:45workflow. So that's going to be what the
1:27:46trigger is here. And we are going to be
1:27:49able to define using fields below. Let's
1:27:51just add in one here that is a lead lead
1:27:54name. Um, for now we can just leave that
1:27:57there. But that's all that we need to
1:27:58set up. We need to go back over to the
1:28:00other one. Just needed to set this part
1:28:01up. Let's rename this qualified lead
1:28:04lead classifier and
1:28:07notifier. So we can save that there. And
1:28:09we have a bit more work to do on this
1:28:10other one. If we go back to this, we can
1:28:13rename this here. So let's call this our
1:28:16lead
1:28:18qualification agent.
1:28:21And so now we have our other workflow
1:28:23set up. We can come here. We can call
1:28:24another workflow with the tool and we
1:28:27can call it lead is qualified. And so
1:28:30now is when we get to tie back into what
1:28:31we learned in the foundation section
1:28:32because we are now writing descriptions
1:28:34for our tools. Remember how we had
1:28:35schemas and scheas are basically
1:28:38written instructions or instruction
1:28:40manuals on how to use tools and how to
1:28:42use the APIs that wrap around them. Um
1:28:44this description is going to be
1:28:45basically those descriptions that you
1:28:46put in the schema. But NAT is going to
1:28:48be basically constructing it for us in
1:28:50the back end. And we just get to put in
1:28:52here, okay, what's the name of this
1:28:53tool? It's going to be called lead is
1:28:55qualified. It's giving us a nice example
1:28:56of how we can write a description for
1:28:58the tool. So call this tool to get a
1:28:59random color. The input should be a
1:29:00string with a comma separated names of
1:29:02colors to execute. So in our case, we
1:29:03can say call this tool when the lead is
1:29:05qualified according to our criteria. The
1:29:07inputs should be lead name, lead email,
1:29:13company, company summary, and request in
1:29:18info. We're basically telling that AI
1:29:20model or the brain what this tool can
1:29:22do. So when we send it some kind of
1:29:23input, it then it looks over our tools.
1:29:25It looks at the Gmail description and it
1:29:27looks at this uh workflow tool
1:29:29description and goes, "Oh, well, I have
1:29:30a tool that does this and a tool that
1:29:31does this. What have they just sent me?
1:29:33Okay, now I think I know what I need to
1:29:34do from here." So this is the rough gist
1:29:36of what we want to do as a description
1:29:37for this tool. I'm actually going to
1:29:38beef it up with a bit of a a bigger one
1:29:40here. Um if the lead is qualified to
1:29:42work with big boy recruits, eg they are
1:29:44software based business like SAS or
1:29:45development agencies and trigger this
1:29:47tool and send the lead data in the
1:29:48following format. It's just dummy data.
1:29:50So name a name email um an email message
1:29:54I want new div qualified true company
1:29:57information and company information
1:29:59which is a summary of the relevance tool
1:30:01to do the company research that we have.
1:30:03So, might seem a bit crazy at the
1:30:04moment, but stick with me because it
1:30:06will make sense in a second. We're
1:30:07basically just told it when it's going
1:30:08to trigger this tool and the format to
1:30:09send the data in. And then for the
1:30:11workflow, we get to choose here the one
1:30:13that we just set up, which is our
1:30:15qualified lead classifier and notifier,
1:30:17at least in my case. And then we see the
1:30:18workflow inputs that we've just set up.
1:30:20So, if we go back over to our other
1:30:22automation, so when we open this up and
1:30:24define our inputs here, you can see over
1:30:26here we are getting just one of them
1:30:28that we've put in as an example so far.
1:30:30So now we need to set up all these
1:30:32inputs correctly. And we have the name
1:30:34that we want. So we've got the name and
1:30:35we want the email. So we can add another
1:30:40email. What else do we need? We need the
1:30:47message.
1:30:49Um, honestly don't think we need this
1:30:51qualified one here. And then we have the
1:30:54company information.
1:31:02So, if we just test this, head back over
1:31:05here, and
1:31:09we refresh this
1:31:13list. Oh, back out. Save
1:31:16it. And this pops up. It says that these
1:31:19inputs are outdated. So, there we go. We
1:31:21have lead name, email, message, etc. And
1:31:23then we can actually automatically fill
1:31:24out a lot of these inputs. Maybe I will
1:31:26put this back in here just to show you
1:31:27guys how these work. So if we go um
1:31:35qualified. Um and then we just call it
1:31:37uh true
1:31:39here. I just put it as a string.
1:31:46And we go back here and we add in one
1:31:49more which is a
1:31:52qual
1:31:55qualified which is also a string. And so
1:31:58we have this qualified field here as
1:31:59well. If we go back um I'll just test
1:32:02this. Save it
1:32:05again. Come back over and update
1:32:07this. I think we can actually make it
1:32:09even cooler. So let's go. Um it changes
1:32:12from a string to a boolean. So that's
1:32:13either true or false. Um, and if we test
1:32:16this, save it again, and we change this
1:32:20to take away these little things. Sorry,
1:32:22pisses me off if I don't have this set
1:32:24up right. Um, and then we update this,
1:32:27you'll see this turns into a switch. So,
1:32:28that's a true or false, right? In order
1:32:29to trigger this tool, it should be
1:32:30qualified by default. So, it's a little
1:32:32bit redundant, but it's still cool to
1:32:34show you how we can get AI to fill out
1:32:35these fields. Um, in a lot of cases, we
1:32:37don't, but for this qualified run, we
1:32:39can. So, we can set this as an AI
1:32:41generated field. We can say if the elite
1:32:43is qualified based on our
1:32:47criteria this set to
1:32:50true. And then for the rest of
1:32:53them we can just fill this out
1:33:04quickly. Give this a
1:33:07second. Open this back
1:33:10up. And we can fill out some of these
1:33:12fields. So we can go lead name that we
1:33:14want to pass through to the other
1:33:15automation is going to be that. So we
1:33:17can fill a lot of these in email
1:33:21um message and the company information
1:33:26um we can get from um this technically
1:33:28but um maybe we could do a cool AI
1:33:30generated one here which is um a short
1:33:34summary of the
1:33:37company and the industry that they are
1:33:40in.
1:33:42company's details. So, this is basically
1:33:43telling the AI model of our agent how to
1:33:46fill this field out, which is one of the
1:33:47reasons that AI agents are so powerful.
1:33:49So, that's all set up. Now, we've got
1:33:51all of our tools set up. The last step
1:33:52is just to set up a prompt for our
1:33:54agent. And I am going to cheat a little
1:33:56bit here and just throw one in that I've
1:33:57done before to save us a bit of time
1:33:59here. So, we want to set the prompt
1:34:03here. We can paste in this information
1:34:05here. And this is basically just telling
1:34:07the agent who it is and what it's
1:34:08supposed to do. You're a lead
1:34:09qualification agent. Your job is to
1:34:11analyze the form submission and company
1:34:12research provided and then decide
1:34:14whether they are qualified to work with
1:34:15big boy recruits. Ra we specialize in
1:34:18XYZ. Um we are specialist in capturing
1:34:20talent for ra. We only work with
1:34:22softwarebased businesses, EG SAS
1:34:24companies or development agencies. These
1:34:25companies are willing to pay much more
1:34:27developers than your average marketing
1:34:28company or local business. Therefore, we
1:34:30only work with them. Your job is to
1:34:31determine if the lead you are provided
1:34:33with is a good fit for big boy recruits.
1:34:34And if so, call the lead is qualified
1:34:36tool and send the elite information to
1:34:38it. If lead is not qualified, then you
1:34:39must trigger the Gmail send email tool
1:34:41for us to respond to them letting know
1:34:43letting them know we are unable to work
1:34:44with them. And then we have a response
1:34:45format here which we can probably just
1:34:47delete. And then we can add in here is
1:34:49the lead
1:34:50to information for you to
1:34:54analyze. Let's pop this out to make it a
1:34:56bit easier.
1:35:00We can add in um just go name um company
1:35:07URL
1:35:09in go message
1:35:13request and of course we go a
1:35:17company and we take it. So that's from
1:35:20the relevant step for the research that
1:35:22we did to scrape using our firecrawl
1:35:24tool. And then we provided all the
1:35:25information to this agent and it's going
1:35:27to be injected with all of these values
1:35:29on each form submission and then it's
1:35:31going to make a call on what tool it
1:35:32needs to use. So we're pretty much
1:35:35there. We can even give it a run here
1:35:37and try to test the step and see which
1:35:39tool it's going to try to
1:35:44choose. If we go back we can see okay
1:35:46look it's used the chat model as the
1:35:48brain and it's triggered the NA10
1:35:51workflow as expected. You can see here
1:35:52that it's sent off information to our
1:35:54other workflow. It sent the lead name,
1:35:56the email, the message, and the company
1:35:59information, and it set it as qualified
1:36:00as well. So, all of these fields have
1:36:02been filled out. We've got a nice AI
1:36:03generated summary here from the model
1:36:05and brain. And we have the qualified set
1:36:07to true. And so, the final step now for
1:36:09us is to head back over to our other
1:36:12automation and just finish it
1:36:14off. Oh, we need to save that. I will
1:36:17just run that again for you so you can
1:36:19see it in slow motion. It's using the
1:36:20LLM as the brain and the tools agent and
1:36:22it's deciding whether it's qualified or
1:36:24not. And if it is qualified, then it
1:36:26will send it to this workflow. Bam,
1:36:28we've sent it. And there you go. If we
1:36:31head back over to um let's save
1:36:35this. Head back over to our qualified
1:36:37lead classifier notifier. Now, we can
1:36:40add on a quick few steps here. I'm just
1:36:41sort of going to rip through this. Um
1:36:43it's not super important. Um but it just
1:36:45shows you a little bit more
1:36:46functionality of what you can build in
1:36:47on N10. So, we're going to add in a
1:36:49messenger model step here, and we're
1:36:51going to choose uh 40
1:36:55mini. And what I want this to do is to
1:36:57take in that information that we sent to
1:36:58the workflow about the company research,
1:37:00etc. So, we know this is a qualified
1:37:02lead now, but we just want to split it
1:37:03between either our SAS team or our
1:37:05development agency sales team. So,
1:37:07they're specialized in dealing with
1:37:08different cases. So, I'm going to cheat
1:37:09and just throw in a prompt here, which
1:37:10you guys will be able to get access to.
1:37:13Um, which is basically saying we have a
1:37:15new inbound lead. Um, change this to an
1:37:16expression. Sorry. Um, we have a new
1:37:19inbound lead that we need you to
1:37:20classify into either SAS or development
1:37:22agency. Here's the lead information. Um,
1:37:24we need to go back step and test
1:37:27this. There we go. We should have some
1:37:30information. Um, and now we can put
1:37:33these in. So, you see how there was
1:37:34nothing here before I went back and
1:37:35tested the trigger so that it gives us
1:37:36some null values here that we can fill
1:37:38out. Here's lead information. Um lead
1:37:41name uh message
1:37:45um
1:37:46name
1:37:47request company
1:37:51information if the company is a SAS
1:37:53output SAS if the lead has development
1:37:56agency upput agency. So we're looking
1:37:57for just agency or SAS as the outputs
1:38:00here. Um simplify the output. Yep,
1:38:04that's all good. So we can no point in
1:38:06us testing that step there because all
1:38:07the values are null. Um but the next
1:38:10step is a basic
1:38:12router flow
1:38:15if so this is a basic conditional
1:38:17routing. So we have the conditions we
1:38:19can go expression here. So we can go
1:38:23um the content here. So this is the
1:38:24output from the open AI step. If the
1:38:27content which is the response from the
1:38:29LLM step the classifier it's either
1:38:30going to be agency or it's going to be
1:38:32SAS. So if it let's just to to make it a
1:38:34bit more flexible. If we go string, if
1:38:36it contains
1:38:40agency, great. So, if it contains agency
1:38:42on the true side, we want to go
1:38:45Gmail and we want to send a message. Um,
1:38:50and then if it is false, we want to do
1:38:54basically the same thing. Now, I've got
1:38:55a preset email just to help us out here.
1:39:09Um, okay. So, here we're not getting
1:39:11much data on the input side here and we
1:39:13can't seem to simulate it because it's
1:39:14of course triggered by another workflow.
1:39:16What we can do is just save it here. go
1:39:18into
1:39:19executions. And if we go back to our
1:39:21other workflow, the
1:39:27agent, and we can go to the form
1:39:29submissions
1:39:31one. If we go to executions, and we just
1:39:34run one of these that we just did
1:39:36before, copy to
1:39:38editor. And we just run this
1:39:44again. This is basically just going to
1:39:46trigger this again. and so that we get a
1:39:48fresh execution and we can sort of pull
1:39:49that data back into the workflow. Oh,
1:39:52we're having issues
1:40:02now. It's cuz I need to delete this.
1:40:10So if
1:40:15we stop that and we just run it
1:40:23again. Boom. Triggered it. And that's
1:40:26all done. Now if we go back to this and
1:40:28we go to executions, go to the most
1:40:30recent one that succeeded. It's going to
1:40:32load in. Oh, hang on. This one's
1:40:35probably it.
1:40:41Oh no, that's not
1:40:47it. Okay, so this one here, if we click
1:40:51this, yep, we've got all the data in
1:40:53here. So, what we can do is copy this
1:40:55into the editor and then we've got the
1:40:57data that we need that's already loaded
1:40:58in so that we have some values to put
1:41:00into our Gmail steps Gmail. So, that's a
1:41:02handy little trick to to know how to do.
1:41:05And now we have all of this information.
1:41:07So, that's what that's what I was trying
1:41:08to get. Um, the same setup and we're
1:41:10going to send this to I'm just going to
1:41:12use an example here and call this um
1:41:15it's the same email. You wouldn't pull
1:41:17this in necessarily. I'm just using this
1:41:18as something that I can show you
1:41:20at show. Say new agency lead. Let's do a
1:41:24text. We say um new agency lead man. Go
1:41:29get
1:41:29him. Um turn into an
1:41:32expression. And then we say we can just
1:41:35throw this company data in there. It's
1:41:37going to be messy. You can play around
1:41:39with this more when it comes to
1:41:40formatting, but just to show you the
1:41:41functionality. If we go uh test a step
1:41:44here, that's going to sent an email to
1:41:46this. This is like my agency sales reps
1:41:49uh email. Of
1:41:51course, um let's rename
1:42:00this. Can duplicate this. Right click,
1:42:04duplicate, bring it
1:42:06here.
1:42:08Oh, connect this up. And we change this.
1:42:12You change this to your like SAS guy.
1:42:14You change it to a different different
1:42:15email. Um, of course, and then you can
1:42:18say new SAS
1:42:22lead. Right. So now we have done all of
1:42:24that basically all built out. The data
1:42:27is going to come in from the agent. It's
1:42:28going to send in the company summary.
1:42:30This is going to classify it into being
1:42:32a uh agency lead or a SAS lead because
1:42:34those are the only two types of
1:42:35businesses that we work with. So all of
1:42:37them will be qualified when they come
1:42:38through here. And then it just sends an
1:42:40email to our uh agency sales rep or our
1:42:43uh SAS oh rename this to our SAS sales
1:42:46rep, new SAS lead for them to continue
1:42:50and do the next steps and follow up
1:42:54with. Right. So to test this we can turn
1:42:56this on to active and you can see that
1:42:58you can now make calls from your
1:42:59production form URL. Um we can go okay.
1:43:03If we double click on this we can open
1:43:05this up. We can click on production URL.
1:43:07Copy this and open this up in a new tab.
1:43:11And now we can give it a
1:43:16spin. So of course my agency Morning
1:43:18Side AI does development services. So
1:43:20this should be qualified and it should
1:43:21also route it to the agency email. So if
1:43:24I now go
1:43:26submit, we go back into NADN, we go into
1:43:29executions, we can see this one is
1:43:32running. If we go to
1:43:35Gmail and I pull up my spare
1:43:41inbox and there we go. We see it has
1:43:43succeeded. And then if we go
1:43:47to and then if we go to our lead
1:43:49qualifier and notifier and we go to
1:43:51executions, we will also see that we
1:43:53have a new one that has succeeded here
1:43:55which was just uh a few seconds ago and
1:43:58that's gone through. It is um outputed
1:44:01it as agency which is the the
1:44:03classification that we wanted. It has
1:44:05gone through and has sent a new email.
1:44:06And if we go to here we have new agency
1:44:08lead there. There we go. All the
1:44:10information. So that's working number
1:44:11one. Now we can go back to our form and
1:44:13we can try it again but this time with
1:44:14let's say an unqualified business. Let's
1:44:16go. What is your name? Ray Croc Ray
1:44:21McDonald.com. Um
1:44:26McTum I need more guys more people
1:44:30flipping damn
1:44:33burgers. So essentially Ray here has
1:44:35come to our recruitment agency and
1:44:37they're asking, "Hey, I need people to
1:44:39do flip patties for me in my fast food
1:44:41restaurant." Um, and because Big Boy
1:44:43Recruits in Dallas, which is a
1:44:44hypothetical company, of course, um,
1:44:47doesn't do that. It's going to qualify
1:44:49them or it's going to disqualify them
1:44:50and then send an email to our good
1:44:52buddy. Oh
1:44:54no, some poor dude at McDonald's is
1:44:56going to get an email now. Um, because
1:44:58send an email. Um, but it's going to be
1:45:00running and of course it's going to be
1:45:02sending. if they're unqualified, it will
1:45:04send an email to them and say, "Hey,
1:45:05sorry you're not a good fit for us. Let
1:45:07us know if we can help or we can connect
1:45:08you to our
1:45:11partners." And while that is running, I
1:45:13would just put together the final one
1:45:14here to test the functionality, which is
1:45:16if we go Liam
1:45:18admin.com, and we set up my SAS
1:45:21https, my SAS agent, if you haven't
1:45:24already used it, we're going to show you
1:45:25how to use it in the last tutorial of
1:45:26this video. So, you guys will get to see
1:45:27that. Um, which is my own no code AI
1:45:30agent building platform. And what can we
1:45:32help you with agents? Um, so this should
1:45:34be a SAS one and it's going to qualify
1:45:36as SAS. Um, we
1:45:41have the McDonald's one has succeeded
1:45:45here. And you see, yep, as expected, we
1:45:47were not qualified. The McDonald's
1:45:49person was not qualified for our offer.
1:45:51So, it looked at the qualification
1:45:52criteria we provided in here, said,
1:45:54"Hey, no, that's not a good fit." So,
1:45:55I'm going to use this tool. And you can
1:45:57see that it sent the email and it said,
1:46:00"Hey, thanks for your interest. Um, but
1:46:03we're not a good fit for you." So,
1:46:04someone at McDonald's just got an email.
1:46:06Apologize for that, but we didn't
1:46:07trigger the other workflow, which is a
1:46:09key part. And we're not going to send
1:46:10emails to our sales team saying, "Hey,
1:46:12look, new leads." Now, I have sent
1:46:14another one through here which just
1:46:15finished
1:46:17executing. And we can see this. It's
1:46:19gone through. It's researched um
1:46:21agentive. you'll see um Agent is a
1:46:24leading service delivery platform for AI
1:46:25agent AI automation agency owners um etc
1:46:29and it's called the tool because we were
1:46:30qualified because we're a SAS business
1:46:32right and again if we go back to
1:46:35here and we look at the most recent one
1:46:38which is this oh we have another one
1:46:43here then you'll see new SAS lead has
1:46:46been triggered because we are a SAS of
1:46:48course um the LLM step here has outputed
1:46:51just SAS So that means that it should
1:46:53send an email to the SAS team, which if
1:46:55we go to my
1:46:58inbox, tada, new SAS lead, right?
1:47:02So I know that may have taken a while,
1:47:04but uh we got there eventually. And you
1:47:06can see that we've built out all of this
1:47:07functionality. We have our AI agent
1:47:09calling our tools if they are qualified
1:47:10and triggering this other workflow.
1:47:12Again, you can build so much cool stuff
1:47:13by connecting an agent to multiple
1:47:15different workflows. We have a little
1:47:17relevance AI researcher tool that we're
1:47:18reusing here and we have people getting
1:47:20denied um with an instant email sending
1:47:22them back. So, hope this been a cool one
1:47:24to show you how NATM works. I really,
1:47:26really like this agent functionality
1:47:27that they have. I think you guys are
1:47:28going to be able to build some awesome
1:47:30stuff if you keep going down this rabbit
1:47:31hole. So, that has been agent build
1:47:33number two. Stick with me as we jump
1:47:35into agent build number three, which is
1:47:37a pretty damn cool one, focusing on both
1:47:39chat and voice-based agents all in the
1:47:41same build. So, let's get the ball
1:47:43rolling.
1:47:49All right, so that is two builds out of
1:47:50the way. Well done if you made it this
1:47:51far. We have another big one here. Um,
1:47:53this is going to be breaking down how to
1:47:55use voice flow. Let's take this off
1:47:57here. Um, to build an agent that is
1:47:59going to be both accessible through a
1:48:00website chat, so you can chat to it on a
1:48:02website as a as a chat widget, which
1:48:04you're probably familiar with, but we're
1:48:05going to connect the exact same agent
1:48:07and exact same functionality that you
1:48:08get through that chat widget also to a
1:48:10phone number on the website that will be
1:48:12able to call and have the same
1:48:13experience. So, this is going to show
1:48:14you how on voice you can build both chat
1:48:16and voice experiences. Um, and this is a
1:48:19new feature for them as of the time of
1:48:20filming. And this agent is what we can
1:48:21classify as an AI customer support and
1:48:23lead generation agent for both website
1:48:25and phone. And we're going to build it
1:48:26on voice flow, of course. And the
1:48:28purpose of this agent is that it's
1:48:29designed to be able to answer common
1:48:30questions from potential customers via a
1:48:32website chatbot and also via a phone
1:48:33number that can be called. Not only can
1:48:35it answer questions to help them sort of
1:48:37move them towards a uh a purchase, but
1:48:39it can also generate instant quotes for
1:48:41interested parties. This is intended to
1:48:43increase the number of leads that they
1:48:44get because people who see a contact
1:48:46form may be like, "Hey, I want to get
1:48:47instant response. I want to know
1:48:49instantly how much this is cuz I'm
1:48:50shopping around." Um, and rather than
1:48:51just filling in a contact form and
1:48:53waiting. Um, having this instant
1:48:54quotation can give people confirmation
1:48:56on the price. Um, so they're ready to
1:48:58take a step forward and end up getting
1:48:59the sale ultimately. So that instant
1:49:01quotation feature is a cool one. Um,
1:49:02very easy to do with the custom tool on
1:49:04relevance that we're going to build. And
1:49:05finally, this agent is going to be able
1:49:07to actually capture lead information
1:49:08from those who have been given a quote.
1:49:10So after they've been given a quote,
1:49:11then we'll move to say, "Hey, give us
1:49:13your details and we will follow up. Our
1:49:15team will follow you up and set an
1:49:16appointment for the service." And the
1:49:17value here of the system is that
1:49:18customers are going to often want
1:49:20instant answers so that they can make a
1:49:21purchase. So by offering easy ways for
1:49:23them to get this information, we can
1:49:24increase the chance that they're going
1:49:25to purchase from the business. Um,
1:49:27companies typically have to spend money
1:49:28on some kind of customer support or
1:49:30sales team in order to get these kind of
1:49:32answers given to customers when they
1:49:33need them. But this agent can
1:49:35essentially be a oneanddone solution um
1:49:37to both help increase the sales of the
1:49:39business by increasing that likelihood
1:49:40of purchasing because they now have more
1:49:42information um while also saving the
1:49:44business money that they would typically
1:49:45spend on some kind of support staff um
1:49:47say if this chatbot can handle a dozens
1:49:50and dozens of responses a a week that
1:49:52would typically have to have gone
1:49:53through a support person then we're
1:49:54saving the company money and also
1:49:56helping them increase their chance of
1:49:57generating more revenue. So um here's a
1:49:59rough layout of the design here. We are
1:50:01going to have a website. I'll give you a
1:50:02template for this. It's very easy to set
1:50:04up and we're just going to throw in a
1:50:05number um that's going to be connected
1:50:06to the voice agent that we build and
1:50:08we're going to be setting up voice flows
1:50:09web chat widget as well. And this agent
1:50:11is going to have access to a knowledge
1:50:13base to answer questions um that
1:50:15prospects may have about the business
1:50:16and their services etc. Um it's going to
1:50:18have a tool that is allowing them to
1:50:20generate an instant quotation. So it's
1:50:21going to take in some information. This
1:50:23is going to be for a cleaning business
1:50:24or a hypothetical cleaning business. And
1:50:25then we're going to be able to generate
1:50:26an instant quote for them based on the
1:50:28property type and the size of the
1:50:29property they need cleaned. and we're
1:50:31going to be able to capture the lead
1:50:32information afterwards and log it into a
1:50:34CRM. In this case, we'll just use Google
1:50:35Sheets, but it's fairly easy to swap
1:50:37that out to whatever CRM you want. So,
1:50:38it's going to look a bit like this. I've
1:50:40actually added a little bit more. And we
1:50:41are using another relevance tool in
1:50:42here. This is from a different project
1:50:44from my accelerator, but I'm going to be
1:50:45pinching that and putting it in here for
1:50:46you all. And this is the uh tool number
1:50:49two here, the generate instant quote.
1:50:51So, we're going to be slotting that in
1:50:52there, taking in some information,
1:50:53answering questions, etc. The process of
1:50:55building on Voice Flow is one of my kind
1:50:56of favorite experiences um in the
1:50:58automation space. I really like the the
1:50:59way they've built out their uh their
1:51:00flow builder. Um so I'm sure you guys
1:51:02will enjoy building this uh step by step
1:51:04with me. And then the general usage
1:51:05pattern of the system is that the
1:51:06person's going to arrive on the website.
1:51:08They'll either click to chat with the
1:51:09chatbot and engage with this
1:51:11functionality or they'll enter the phone
1:51:12number into their phone. And then the
1:51:13agent will jump in and respond either
1:51:15through text or through uh voice and
1:51:17determine what they're needing help
1:51:18with, which is this section here. And
1:51:19then it's going to be routing using this
1:51:21router section here to the correct tool
1:51:23whether they want a question answered or
1:51:25they want to get a quote. Um, and then
1:51:27each of these branches will execute on
1:51:28that uh, functionality depending on
1:51:30their intent. So, it's going to look a
1:51:32bit like this. We'll have a phone number
1:51:33and we'll have a chatbot like this. This
1:51:34is actually an agenda chatbot from my
1:51:36own software, but we'll be swapping this
1:51:37out to a voice flow one in this build.
1:51:39So, without further ado, let's jump into
1:51:40voice
1:51:42flow. So, when you click that link on
1:51:44the Figma, it's going to take you to the
1:51:45signup page. You can sign up there and
1:51:47then once you're in, you're going to get
1:51:48a page that looks a bit like this. The
1:51:49first thing that we want to do, of
1:51:50course, is to create a new agent up here
1:51:52on the top right. Let's call this
1:51:54Bonor's cleaning um, website. and phone
1:51:56agent. Um, let's just start with a basic
1:51:58template here. Import knowledge for this
1:52:00import knowledge. We can actually just
1:52:02skip that for now. And then we get into
1:52:03the flow builder on voice flow. So just
1:52:06a quick orientation if you are new to
1:52:07the voice flow platform. This is where
1:52:08we can add in our knowledge which we
1:52:10will do shortly. The workflows are where
1:52:12we access the flow builder. In most
1:52:13simple builds like this, you just going
1:52:14to have one workflow. So you don't need
1:52:16to worry too much about that. Now we
1:52:17have integrations like uh the widget
1:52:20which we're going to be using to deploy
1:52:21this on a website. We have the phone
1:52:23number integration which we're going to
1:52:24be doing later as well. Then we have API
1:52:26keys etc which you don't need to worry
1:52:28too much about right now. We have some
1:52:29publishing features here which we'll
1:52:30double later. We have access to
1:52:32transcripts. So once we deploy this you
1:52:34can access all of the transcripts either
1:52:35by voice or through chat here and and
1:52:37sort of dig through the answers and and
1:52:38see how the uh people are interacting
1:52:40with the agent that you've built. Um
1:52:42something that a lot of people neglect
1:52:43after they've put one of these into
1:52:44production. And then we have things like
1:52:46analytics um etc. But obviously we need
1:52:48some data before we see anything useful
1:52:49there. And then the settings page is not
1:52:50too much you need to worry about right
1:52:51now. Just sort of on a need to know
1:52:52basis. The more important stuff, of
1:52:54course, is up in this first tab here,
1:52:55which is content. So, we have messages,
1:52:57we have prompts, we have components.
1:52:59We're going to be working a lot with
1:53:01prompts shortly. So, that's the main one
1:53:02we need to be worried about. But for
1:53:04now, we can just go into workflows, and
1:53:05we open up this first workflow and edit
1:53:07it. And here we have the template that
1:53:09VoiceFlow gives for us. Um, which I'm
1:53:11actually just going to nuke this, and
1:53:12we'll start fresh. And if you see on the
1:53:13Figma, we have a design here that we're
1:53:15roughly working towards. I'm going to be
1:53:16showing you the sort of step by step.
1:53:18So, we need a welcome step.
1:53:21So, we're going to start off by going
1:53:22here and dropping. I'll try to zoom in a
1:53:24bit for you guys here. Talk message. So,
1:53:27message is how you send a message um via
1:53:29the chat. So, start is when maybe you
1:53:31click on the widget, it pops up. And
1:53:32this message that we're about to put in
1:53:34is the first message the bot is going to
1:53:35send. So, we say um up here, hey, hey,
1:53:39welcome to
1:53:41corners. I'm going to zoom this up for a
1:53:43bit for you guys. And right away, we
1:53:44have a little tip and trick that I want
1:53:45to give to you because we are building
1:53:47this as a chat and voice assistant. We
1:53:49want to over voice. You don't want to
1:53:51overuse punctuation because it leaves
1:53:52these big long pauses when the the voice
1:53:54agent is going, "Hey, welcome to
1:53:56Connor's cleaning." So, we wanted to
1:53:58just say, "Hey, welcome to Connor's
1:54:00cleaning." A bit more natural. There's
1:54:01times where you'll see me on the side of
1:54:02sloppier punctuation, but that's just to
1:54:04ensure that when we get to testing it on
1:54:06voice, it sounds as natural as possible.
1:54:08So, hey, welcome to Connor's cleaning.
1:54:09And then, of course, we're going to wait
1:54:10for them to reply and say something back
1:54:12to us. We go to listen and then capture.
1:54:14And this is going to capture the
1:54:15information. We want to change this from
1:54:17capturing entities which is like say I'm
1:54:19looking for a price or an address. Um
1:54:21we're just going to go the entire user
1:54:22reply and the reply is going to be saved
1:54:24into this variable here. So we actually
1:54:25want to change this to um first user
1:54:30reply because we're going to need it a
1:54:32little bit later. Um the users first
1:54:37reply. And I like to name these as we
1:54:39go. So we can call this
1:54:41welcome. Drag this out here. Get a new
1:54:43one. And we're going to be doing a uh a
1:54:45set step here. So, we're going to be
1:54:47setting some variables. And I'm going to
1:54:48add a new variable to set. And we're
1:54:51going to do it based off a prompt here,
1:54:52which is a cool feature in voice they've
1:54:54added recently. And we're going to be
1:54:55able to select a prompt that's going to
1:54:56take this information from this first
1:54:58reply and then generate some kind of
1:54:59output from it and set a variable. And
1:55:01the variable we're going to set here is
1:55:02called last response. So, this is
1:55:04typically what you're going to put as
1:55:05the last response from the AI or from
1:55:06the agent. Um, last response here. And
1:55:08last utterance is typically the most
1:55:10recent uh message from the user. So
1:55:12utterance is coming from the user the
1:55:13most recent last utterance and the last
1:55:15response is what the the AI or the agent
1:55:18or the system has last responded to. So
1:55:19we want to set the last response to
1:55:22something that is generated through this
1:55:24prompt here. So we can create a new
1:55:25prompt here and this is basically going
1:55:27to take in the data from the chat and
1:55:28the conversation so far and we'll be
1:55:30able to generate things off of it. So
1:55:31say we add in here the conversation
1:55:33history. That's a good thing to have in
1:55:35in most cases. And I'm going to be
1:55:36dumping in some of the prompts here just
1:55:37to save us a bit of time. But basically,
1:55:39we're saying summarize the customer's
1:55:40question below and ask them to confirm
1:55:42that that's what they meant. And so,
1:55:43we're not actually going to be
1:55:44generating the last utterance here.
1:55:45We're going to be adding in the last the
1:55:48first user reply that we got. I mean,
1:55:50it's going to be included here in the
1:55:51conversation history, but there's no
1:55:53harm really in hard coding it or at
1:55:54least putting the variable in here to
1:55:56make sure that it's in there. Um, we're
1:55:57just saying summarize the customer's
1:55:59question and basically say a
1:56:00confirmation statement. So, just to
1:56:01confirm this is what you're looking to
1:56:02do. So, you imagine this over the phone.
1:56:04Hey, um um yeah, I'm not really sure
1:56:07what I'm supposed to be doing here, but
1:56:08I was thinking if if you guys were
1:56:10possibly so all of that information is
1:56:12taken in, we can flick back to them and
1:56:14say, just to confirm this, this sounds
1:56:15like you're looking for this, yes or no.
1:56:17Um so, ensure your tone is empathetic.
1:56:18Speak directly to the end customer. Keep
1:56:19your answer brief and two sentences max.
1:56:22So, if we go back here and actually we
1:56:25can name that prompt. So it's um
1:56:28summarize
1:56:30problem and then we need to send the
1:56:33response. So this prompt is going to
1:56:35take in the information we provided
1:56:37here. It's going to use this prompt to
1:56:38take in um the conversation history so
1:56:41far and this information from the user
1:56:43in the first question. It's going to
1:56:44generate a a question to ask back and
1:56:47it's going to save it to this variable
1:56:49that we have here. So apply output to
1:56:51variable last response. There is
1:56:52actually an easier way to do AI
1:56:54responses like this, but in our case, we
1:56:55need to be saving this variable. So,
1:56:57it'll make sense in a second. But, we
1:56:58can go into here and we can go last
1:57:01response. And then it's going to send
1:57:04that information back to them. So, let's
1:57:05just do a quick test here. Click start.
1:57:08Hey, welcome to Connor's Cleaning. Oh,
1:57:10actually, we need to ask a question.
1:57:12Hey, welcome to Connor's Cleaning. Um,
1:57:15how can I help?
1:57:23We'll
1:57:24say I need cleaning services for my
1:57:28house. Sounds like you're looking for
1:57:30cleaning services for your greenhouse.
1:57:31Is that correct? Want to make sure I
1:57:33understand your specific needs before we
1:57:34proceed. So, I obviously spelled house
1:57:36with G house. So, we thought it was a
1:57:37greenhouse, but that's what we want.
1:57:39Some kind of confirmation message just
1:57:40saying like, hey, look, is this what
1:57:41you're actually looking to do before we
1:57:42then go and trigger the different tools
1:57:44that we are equipping our agent with? By
1:57:45the way, there is a way of changing
1:57:47between trackpad and mouse. So I am
1:57:48panning around with my mouse here. You
1:57:50can also do a trackpad method which is a
1:57:52lot easier to use if you're if you're
1:57:53having trouble using it. Okay. So after
1:57:54that they're obviously going to say yes
1:57:56or no whether like have I got the
1:57:57question or have I summarized what
1:57:59you're looking for correctly and we can
1:58:00go to a choice step here and we can set
1:58:03up some triggers. We can set the intent
1:58:05to yes and then we can add another
1:58:07trigger and we can set it as no. So this
1:58:09is basically using AI to analyze what
1:58:12they've said and grouping it around
1:58:14these certain things which are called
1:58:15intents. So, what is the intent of them
1:58:16of this uh of the response? And in this
1:58:18case, they have some pre-built ones, but
1:58:20we are going to be building our own
1:58:21custom intents later. But for now, just
1:58:23know that if you're looking to sort of
1:58:25split traffic or split people coming
1:58:26through the system, these choice blocks
1:58:28with the default uh intents from voice
1:58:30flow, yes and no, are ready to use out
1:58:32of the box. And if they don't say yes or
1:58:33no clearly or we can't pick it up, we
1:58:35can add a no match here. We can say
1:58:37sorry, I didn't get that. Can you say
1:58:39yes again? A yes or a no is enough. And
1:58:43then we can say to follow a path after
1:58:45these reprompts which we'll call no
1:58:46match. And then we have this uh no match
1:58:49path which we'll set up in a second
1:58:51here. I'll just put it as a a filler for
1:58:53now. Basically if if they don't say yes
1:58:55or no um this is setting up error
1:58:56handling. Um, and basically if people
1:58:59particularly over the phone, um, there's
1:59:01so many different ways that the
1:59:02conversation can go and end up and
1:59:03you'll want to, while I don't focus on
1:59:05it too much in this build, um, as you're
1:59:07building production grade assistants,
1:59:08you'll need to build a lot more of these
1:59:10fallbacks and these reprompt and these
1:59:12no match things to handle edge cases
1:59:14where people use it in a weird way that
1:59:15you don't expect. So, I want to give you
1:59:17a little taste of that in this tutorial,
1:59:18but it is nowhere near representative of
1:59:20what it takes to actually get something
1:59:21into production that you can trust on a
1:59:23on a customer's website. Okay, so we've
1:59:25got this choice block set up to
1:59:26determine if we have got their
1:59:27summarization of the problem correct.
1:59:29And we can take this up to here and we
1:59:31can set another uh variable. And so this
1:59:33is really the core part of the
1:59:35application which is determining what
1:59:36their intent is, what are they looking
1:59:37to do and which tool are we going to
1:59:39route them to it. So this is a very
1:59:40handy skill to have which I'm going to
1:59:41teach you which is how to set up some
1:59:43kind of intent classification system. Um
1:59:46which is really really essential to
1:59:47building agents on on voice flow and any
1:59:49kind of agent where the platform itself
1:59:50isn't automatically handling that for
1:59:52you. So if we go set a new variable and
1:59:54we're going to do it through a prompt.
1:59:56We're going to set a new variable here
1:59:58called
2:00:00desired action. So basically people
2:00:02coming through and asking questions can
2:00:04be saying hey look I I just have a quick
2:00:07question about where you guys are based
2:00:08and then we're going to route them to
2:00:08the knowledge base. And then someone may
2:00:10be saying hey how much does it cost for
2:00:11this? And then we're going to route them
2:00:12to the pricing uh the instant quotation
2:00:15um system that we're setting up. So
2:00:16needs to be able to determine what
2:00:17they're looking for and we're going to
2:00:18route them depending on that. And that's
2:00:20what this router is going to do. So the
2:00:22action that the prospect wants to take,
2:00:25the most likely action that the prospect
2:00:26wants to take. Then we need to make a
2:00:29new prompt here. We're going to call
2:00:31this intent
2:00:33classifier. Classifies the intent of the
2:00:37uh prospect into asking the knowledge
2:00:40base or generating an instant
2:00:44quote. Add in the conversation history.
2:00:46It's always good to have that in there.
2:00:47And I'm going to put in the prompt that
2:00:48I've written previously. And this is a
2:00:49pretty basic one as well, which is just
2:00:51saying what does the customer want to
2:00:52do, ask a question, get a real-time
2:00:53quote, or something else entirely. You
2:00:55must output a label for this only. Your
2:00:57options are ask a question, get a quote,
2:01:00or other. And you guys can just pause it
2:01:01and see what I've got in there. But
2:01:02basically, anything asking a question is
2:01:04going to be about the business and the
2:01:05services. And if there's anything about
2:01:07pricing or directly related to getting a
2:01:09quote and like they're ready to move on
2:01:10this, then we're going to route them to
2:01:12get a quote. And anything else is going
2:01:13to go to other. And because in the next
2:01:14step, we are going to be looking out for
2:01:16either this is the output or this or
2:01:18this. a really clear statement saying
2:01:20this is all you need to output just this
2:01:21and not hey I took a look at the the
2:01:23conversation history and it seems like
2:01:25the user wants to do get a quote we just
2:01:27want just get a quote and we can
2:01:28explicitly state that with this big caps
2:01:30lock block here and as with the other
2:01:31builds all of these prompts are going to
2:01:32be available in the resources for you to
2:01:34follow along with okay so now we get to
2:01:37the cool bit which is routing this. So
2:01:39if we go condition add this in here and
2:01:43we go add path condition builder and we
2:01:46say if desired action
2:01:49is
2:01:51ask
2:01:52a
2:01:57question. Oh that's that's all we need
2:01:59there. So as you can see that's added
2:02:01one in here. And if we want to add
2:02:02another one in, we go if desired action
2:02:05is um what do we have the label as
2:02:11output? What did we what's the exact
2:02:14label that we had in here? So this
2:02:16prompt is going to be outputting these
2:02:18labels. So we need to make sure they
2:02:19match up. So ask a question. Get a
2:02:21quote.
2:02:25Yep. Get a quote.
2:02:29and other one is
2:02:33other and it's already got an else path
2:02:36in there for any error handling as well.
2:02:37So what this now allows us to do is to
2:02:39build out our different tools. So we can
2:02:41go up to here to ask a question. Just
2:02:43throw this in for now so we can get an
2:02:45idea of what it's going to look
2:02:47like.
2:02:50Um other this is going to be sort of
2:02:52error
2:02:54handling. And if it's else, that means
2:02:56that the LLM step here hasn't outputed
2:02:59any of the labels that we told it to,
2:03:00and it's likely thrown in a bunch of
2:03:02rubbish. Um, so this is sort of an error
2:03:04handling step. Um, should say, "Sorry,
2:03:07something went wrong, at least during
2:03:09this prototyping phase. So now what I'd
2:03:10like to do is make this look a little
2:03:11bit prettier. Um, we can go through and
2:03:13add things in here like this is the uh
2:03:16confirm problem um intent classifier
2:03:20router." And then we can go here add a
2:03:23note can say tool number number one
2:03:26answer from knowledge
2:03:35base two you guys don't have this so be
2:03:37easy for you guys to see understand I
2:03:38know this might look pretty
2:03:42confusing and then this other one we
2:03:44don't need to worry about too much so to
2:03:45keep things quick I'm not going to test
2:03:46this just yet I'm going to test it once
2:03:47we've got that functionality set up on
2:03:49either side at least for this top one
2:03:51first so now We need to go and set up
2:03:52our knowledge base. And to do this, we
2:03:53can click on the back button here. Go to
2:03:55our
2:03:57knowledge. And here we have a data
2:03:59source which we can upload. I'm going to
2:04:00upload a file, but you can put in URLs
2:04:02to different websites, etc. I'm going to
2:04:04be uploading a file
2:04:06here. And I'm going to upload this
2:04:08Connor's cleaning FAQ kind of document,
2:04:10which you guys are going to have access
2:04:11to in the resources. Basically, it's
2:04:13just about us, location, our services,
2:04:15ra um some frequently asked questions,
2:04:18etc. So, I've just AI generated this.
2:04:20Um, and if you're doing any kind of
2:04:21prototype builds, I recommend you do the
2:04:22same just to throw it in there and see
2:04:24if the knowledge base is working as
2:04:25expected. Obviously, you'd swap this out
2:04:27with actual customer data or or your
2:04:28client's FAQ. I'm just going to throw
2:04:30that in there for now. And you guys can
2:04:32do the
2:04:34same. And when you're setting up your
2:04:36knowledge base, you can also set up the
2:04:38settings for it. So, in this case, it's
2:04:40using by default Claude 3.5 Haiku. And
2:04:42you can see how many tokens this is
2:04:43going to cost you for Voice Flow's
2:04:44usage. Um, what is Haiku? Haiku seems to
2:04:47be the cheapest. Oh, you've got GBT40
2:04:49mini. Let me just chuck on GBT40 mini
2:04:51here. We want this thing to be pretty
2:04:53deterministic. So, I'd say 0.1 is fine.
2:04:55Max tokens. Um, we can increase that
2:04:57just in case it needs to give a longer
2:04:59answer. And chunk limit of of three
2:05:02should be enough. So, that's just so
2:05:03this stuff is a little bit more
2:05:04advanced. That's that vector database as
2:05:06I was talking about. Basically,
2:05:07knowledge base is going to be sending
2:05:08the message that we ask it and querying
2:05:11it and getting back chunks of
2:05:12information. Because our knowledge base
2:05:14is quite small, we don't really need to
2:05:15have too many chunks. If you put this
2:05:16up, you just be getting the whole
2:05:17document back. Anyway, so max tokens,
2:05:19the number of tokens that it's going to
2:05:21include in the response. So, we want to
2:05:22increase that to 480 um so that it can
2:05:24give a longer response if they need.
2:05:25Maybe just tone that down a bit. And
2:05:27these are of course kind of controls
2:05:28that you have on how much you want to
2:05:30allow the app to spend. And those
2:05:31settings obviously the main ways that
2:05:32you can control how much um your
2:05:34knowledge base is using and how much
2:05:35your you or your client are ultimately
2:05:37spending on the AI features for the
2:05:39knowledge base. So once we've got that
2:05:40set up, we can go back to workflows
2:05:41here, open this up
2:05:43again, and then to plug in our knowledge
2:05:45base, we can go to the dev section here.
2:05:47We can go to KB search, pop this in
2:05:49here. I will
2:05:51uh we need to delete that and reconnect
2:05:55this up to the top. And we're going to
2:05:57delete this as well. And so we can go
2:06:00into this knowledgebased step here, and
2:06:02we can enter the query. So what we're
2:06:03going to say is we basically want to
2:06:05throw the information that we've got
2:06:06from the user already about what they
2:06:08want which is we have here as the first
2:06:10message they gave us which might be a
2:06:12bit longwinded. Then we have the summary
2:06:13that they have confirmed and then we can
2:06:15put these into the query that's going to
2:06:17be asking the knowledge base hey this is
2:06:19what we want information on can you give
2:06:20us some information back. So we can go
2:06:24user first message put a curable in
2:06:27there. We can go first use reply and
2:06:29then we can also go summarized problem.
2:06:31Now you can see why if we put last
2:06:33response here, why we have this variable
2:06:35saved instead of just sending it
2:06:36automatically. So I actually don't like
2:06:38using last response because that's
2:06:39something that you like to update quite
2:06:40a lot. So I'm actually just going to
2:06:41switch this to um changing it to
2:06:46summarized problem just so we don't get
2:06:49any kind of overlaps that cause problems
2:06:50down the line. A summary of the user's
2:06:55problem.
2:07:02Then we put this in here. Get to spit it
2:07:04out. So when you put a variable in a
2:07:07message, it's just going to print out
2:07:08and and spit out the the value that's
2:07:10inside that variable. So we've set the
2:07:12summarized problem variable and then
2:07:13we're just going to spit it out and send
2:07:15it into the chat or over the over the
2:07:16phone. So now we can come back out to
2:07:17our knowledge base and we can take out
2:07:19the SL response and replace it with
2:07:21summarized summarized problem. Then we
2:07:23can save the chunks that come back from
2:07:24the knowledge base. I don't want to get
2:07:26too in depth on what chunks are
2:07:28specifically. It's a little bit more
2:07:29advanced, but for now, we can just know
2:07:30that it's going to return some
2:07:31information from that knowledge base.
2:07:32It's going to chop up that document we
2:07:34put in. And when we put in this
2:07:36question, it's going to basically ask
2:07:38that knowledge base, can we get three
2:07:39chunks that most closely match the
2:07:41information that is in this query that
2:07:42we sent to it. So, we can save these
2:07:44chunks, which is going to be three
2:07:45because we set that up in the knowledge
2:07:46base settings into this chunks variable.
2:07:48And the chunk limit is three still. If
2:07:50we click this, we can add in a chunks
2:07:52not found path. But for this tutorial,
2:07:53we don't need to worry about that
2:07:54necessarily. And then we're going to use
2:07:56those chunks that came back from the
2:07:57knowledge base to generate an answer
2:07:58based on the original question. So if we
2:08:00put this here, we go talk, we go prompt.
2:08:02And for this prompt, if we go here, we
2:08:04can create a new prompt. We can add in
2:08:06the conversation history just for good
2:08:08measure here to give it the full context
2:08:09of what's going on. And then I have a
2:08:10prompt here. You are an AI customer
2:08:12support rep from Connor's Cleaning
2:08:13helping customer with the question. Use
2:08:15the provider details below to answer the
2:08:17customer's question. Ensure you keep
2:08:18your answer brief and speak directly to
2:08:20your end customer. You are speaking to
2:08:21them over the phone. It's the input data
2:08:24provider details which is the chunks
2:08:25variable. I'll just put that in
2:08:27again. Chunks to make sure it's set up
2:08:30properly to the user's original uh
2:08:34question. We can put all this
2:08:36information back in. So, first reply,
2:08:38this is what they asked us as soon as
2:08:39they picked up the phone or the first
2:08:40message they sent when we asked them
2:08:42what can we help you with. And then we
2:08:44also just put in for good measure our
2:08:46summary of the problem that they
2:08:48confirmed. And we can go to our
2:08:51summarize problem
2:08:53variable and throw that in. That should
2:08:55be good to
2:08:56go. And I like to make these look purple
2:08:59or some kind of cool color. Um, call
2:09:02this a KB
2:09:04query. And we can change this to
2:09:07generate answer from chunks. Let's say
2:09:11from Kh. And if you want to make this a
2:09:13bit easier for you to kind of understand
2:09:15at a glance, you can add in your
2:09:17descriptions on these. So if we go to
2:09:18edit again and we go here, this takes in
2:09:22chunks from the from the KB and their
2:09:25original question and writes a short and
2:09:28sweet answer. And we have this in here
2:09:30that you are speaking to them over the
2:09:31phone because we want to make sure that
2:09:32we're building this with the phone in
2:09:33mind, which is more tricky than just
2:09:35chat. So long text outputs don't really
2:09:37work that well over the phone. So that's
2:09:38kind of why we're putting that in there
2:09:39as well. And we can call this um
2:09:42generate
2:09:44answer. All right. So now we can
2:09:46actually give this a spin. We can start
2:09:48it right from the beginning
2:09:52again. Oh, we may need to if we just
2:09:54click run. Okay, there's no training
2:09:56needed yet. We can run
2:09:59test. Where are you
2:10:01located? Sounds like you're asking about
2:10:03a business location. Could you confirm
2:10:04if you'd like to know the specific
2:10:06address or where Connor's Cleaning
2:10:08operates? Okay, it's good this popped up
2:10:09because as you can see, it's asking for
2:10:10a non- yes or no answer. We're just
2:10:12looking for a confirmation in yes or no.
2:10:15And so this would technically break the
2:10:16system and that'd be saying, "Oh, I'd
2:10:18like a specific address." And this is
2:10:19looking for yes or no. And so it would
2:10:20send it to this no match. So what we can
2:10:23do to fix this is to go into the
2:10:24summarize problem prompt, modify the
2:10:26prompt, and then say they should be able
2:10:29to answer only with yes or no. This is a
2:10:32confirmation step, not asking for more
2:10:35information.
2:10:38So now if we run that
2:10:48again, whereabouts are you
2:10:53located? Sounds like you want to know
2:10:54the specific location of our cleaning
2:10:56business. Is that correct?
2:10:58Yep. Now it goes to the router here.
2:11:02It's going to determine that I said yes.
2:11:04Bam, bam, bam. And there we go. So that
2:11:06that all happened pretty quickly, but
2:11:08you can see it's sort of broken down
2:11:09through here. Um, if I click over here,
2:11:11it's going to remove all of this. Okay,
2:11:13so let's break down how this happened
2:11:14step by step. Um, so you can see this
2:11:16through here. It's still using 3.5 haiku
2:11:18for some reason. I'll need to double
2:11:20check why that's still using the model
2:11:22we didn't select. But basically, it
2:11:24comes through this step here is the
2:11:25intent classifier. So you can see that
2:11:26it set it as yes. That is the correct
2:11:29intent and predicted intent yes. And
2:11:31that routed it to
2:11:33this. And then using the model again, it
2:11:37analyzed the information that it was
2:11:38given. And then it set the desired
2:11:40action variable to ask a question which
2:11:42is one of the labels that we wanted and
2:11:43that is correct. And then it said
2:11:45condition matched taken path one. So it
2:11:47set the desired action variable to ask a
2:11:49question. We were checking for it here.
2:11:51Then it said okay great. Now I'm sending
2:11:52it up here to the knowledgebased query.
2:11:55It says it's query received. We passed
2:11:56in all the information whereabouts you
2:11:58located and then the summary that we
2:12:00gave it. And then we got two chunks back
2:12:02from the KB and the AI response here
2:12:04finally took in all of this information
2:12:06and it gave us the final output and
2:12:08generated this response. We are located
2:12:10at 247 ra and at the end here it's
2:12:12saying is there a specific area you're
2:12:14interested in. For a basic build like
2:12:16this I'd probably change the prompt to
2:12:17say don't ask another question because
2:12:19in this case you then need to set up a
2:12:20looping mechanism where it can keep
2:12:22answering questions for them and then
2:12:24break out into any of these other
2:12:25intents um as needed. But for now that
2:12:28is a knowledge base and that's how you
2:12:29can ask questions. And so that is tool
2:12:30number one knocked out which was easy
2:12:32enough. So great we can go on to tool
2:12:34number two now. So for this second quote
2:12:35I'll be able to give you a relevance
2:12:36tool that we're using for it. But let's
2:12:38just jump into answering this question.
2:12:39So they've said here that they wanted
2:12:41something related to pricing or quoting.
2:12:43Remember in here in the router we have
2:12:45set the intent classifier to say it's
2:12:47going to go to the get a quote if
2:12:48they're asking about pricing or have
2:12:50directly requests a quote. And this will
2:12:52take them to a real-time quotation tool
2:12:53that takes the property type and size
2:12:55and then returns an estimate. So that's
2:12:56the people that are going to be getting
2:12:58to this next branch of the agent. So
2:13:00desired action is get a quote. We can
2:13:04say, "Okay, sure." To give you an
2:13:06instant quote, I
2:13:11just Okay, sure. To give you an instant
2:13:14quote, I just need the properties type
2:13:15and size and square feet. Then I'm going
2:13:18to add another chat step or message,
2:13:20sorry. And we can ask a
2:13:24question. Is the property a house or an
2:13:29apartment? The next we can do one of
2:13:32these choice steps again and this time
2:13:34we get to create some custom intents. So
2:13:36we can go to triggers here and we want
2:13:38to select an intent. See it doesn't have
2:13:39a house. So we can go create an intent
2:13:41and we say house. The
2:13:46users property type is a house. And then
2:13:50we can add in some examples here. So
2:13:52obviously house this is just giving
2:13:54examples to voice flows AI engine to
2:13:56help us better to classify the different
2:13:58intents as they come into this step. So
2:14:00we can say home and then you can add in
2:14:02some AI generated ones here which
2:14:04usually pretty easy and uh which usually
2:14:06pretty good. Okay, residence, dwelling,
2:14:07property, abode, living place. Uh this
2:14:10this gets tricky um because some of
2:14:12these could overlap with apartment. So
2:14:14property is probably abodess too broad
2:14:17living place.
2:14:18uh dwelling uh
2:14:21residence potentially we could get away
2:14:23with. We could say like
2:14:26manor mansion family
2:14:31home single family
2:14:35home and by now we've given enough
2:14:37examples where we can just go to create
2:14:39and now we need to do the same for
2:14:41apartment. So, we'll go to create an
2:14:43intent apartment. The user's property
2:14:47type. The user's property type is an
2:14:49apartment. And then we can put an
2:14:55apartment. Um, see what else it's got
2:14:58for us. Um, townhouse,
2:15:02penthouse, duplex, flat, probably not
2:15:06right. Loft probably not right.
2:15:08Townhouse penthouse. Think that's a good
2:15:10bunch.
2:15:12And then we can add in a no match here
2:15:13as well.
2:15:16Um, and we can add in a reprompt. Sorry,
2:15:19I didn't get that. Is it a house or
2:15:26an If it is a house, we can come up here
2:15:28and set it as logic here and go
2:15:34set. We can
2:15:37go value or expression. Select the
2:15:40variable to set. We're going to go
2:15:42property
2:15:45type the
2:15:48the save that variable or create that
2:15:51variable, sorry. And then we can enter
2:15:53this
2:15:54in and set this as
2:15:58house. So we're setting the property
2:16:00type variable to house when it's been
2:16:02triggered um by this particular
2:16:05route. And then we need to add another
2:16:07one. So, I'll probably just duplicate
2:16:11this. That's by right clicking on one of
2:16:13these blocks. Um, and then we can
2:16:15connect up apartment to
2:16:17it. Property type. Instead of being
2:16:19house, we can make it
2:16:23apartment. And from here, now that we
2:16:25know the property type, we can set the
2:16:27size of the apartment or the house. So,
2:16:29we go and how many
2:16:32square and how many square feet is it?
2:16:37We can connect both of these up here.
2:16:39And then we're going to save the entire
2:16:40user reply or on a capture step. So
2:16:43anything that they respond to after
2:16:44this, we want to capture the entire
2:16:48thing. And we want to go here, set up a
2:16:52set
2:16:54step. Sorry, that's not right.
2:17:05using a prompt. And what we're going to
2:17:07do is use AI to analyze the response and
2:17:10then extract the number of square feet
2:17:12from it. We could have used a step here
2:17:14where we change this to instead of the
2:17:15entire user reply, we change it to an
2:17:17entity. Um, but it's not as reliable as
2:17:20doing it this way. So, I'd prefer to
2:17:21just get it right the first time. Um,
2:17:23cuz a entity of the number of uh number
2:17:26of feet may be a bit harder to pick up
2:17:28than a clear word like an entity of
2:17:30house or apartment or a name, etc. So
2:17:33just to make sure that it works every
2:17:35time for us, we want to create a new
2:17:37prompt. And I'm going to put this in
2:17:39here. Extract the number of square feet
2:17:40from what the user said in numerical
2:17:42format only. Each 500 include nothing
2:17:44else in your response. This will be
2:17:45saved as a variable and passed to an API
2:17:47to a quote generator function. So giving
2:17:49it a bit of context about what's been
2:17:50going on. And also put in the memory
2:17:52there just for good
2:17:54measure. Oh, we didn't name the prompt.
2:17:57Call this extract square.
2:18:04And then we need to set the variable
2:18:05that we want to save this to, which is
2:18:06going to be proper
2:18:08property size. So now we're saving that
2:18:12response, the size of the
2:18:16user's, the size of the user's property
2:18:18in square
2:18:19feet. And then we need to do one kind of
2:18:22tricky step, but it's just something
2:18:23that you guys will pick up as you go and
2:18:25uh as you build more of these. But the
2:18:27next step after this is going to be
2:18:28sending that information to a relevance
2:18:30tool. And the relevance tool is
2:18:31expecting not what's called a string,
2:18:33which is just a number of letters or
2:18:36just text. Um, you can have a number as
2:18:38a text, which is confusing, but
2:18:39basically it's just the format uh in
2:18:41which it's being received in. So,
2:18:42because we're sending it over an API, we
2:18:44need to be specific about the format. So
2:18:46I can't take this is essentially going
2:18:48to be saving that 500 or say if we say
2:18:50it's a 500 foot property. This is going
2:18:52to be plucking that out and giving us
2:18:55500 as a string. In order to get the
2:18:57response we want from relevance we need
2:18:58to convert it to a number and then send
2:19:00it. So a little block of uh custom code
2:19:03here. Um know this was a no code
2:19:06tutorial but I hope you can forgive me
2:19:07for
2:19:10this. And all we need to do is put this
2:19:12in here. So property size this is the
2:19:13variable. We're going to go property
2:19:15size and we're just casting this
2:19:19variable as a number and then
2:19:21reassigning it to the variable that it
2:19:23was before. So we're taking whatever
2:19:24came out of this, we're saying, okay,
2:19:25can you just make it a number um and
2:19:27we'll save it and sort of overwrite the
2:19:29existing
2:19:30variable. So now we have the number 500
2:19:32in that case that we're ready to send to
2:19:34the next step. And then if that's gone
2:19:35correctly, we can talk to them and
2:19:40say one sec while I get
2:19:47And then we have this JavaScript fail
2:19:49route here which you can just sort of
2:19:50throw down there for now. Um there is a
2:19:53chance that this prompt outputs not just
2:19:56a number. As you can see we're asking it
2:19:57for just this and therefore the property
2:20:00size variable if it just is 500. Um but
2:20:04sometimes it can say hey your size is
2:20:09500. and then we end up with a variable
2:20:10that's not actually convertible into a
2:20:12number. Um, so we do need to add a
2:20:15little bit there as a as a potential
2:20:17fallback. Um, probably won't be doing it
2:20:19in this video if I'm honest. It's fairly
2:20:21basic. Um, but you would add this in
2:20:22here some sort of looping back in to
2:20:24make sure that it is actually in the
2:20:25right format or just make sure that your
2:20:26prompt is actually only you put like a
2:20:28strict instruction only output just the
2:20:30number um so that you get less errors
2:20:32there. So the next major step after this
2:20:34is to get our relevant AI research. If
2:20:36we go back to the Figma here, you'll see
2:20:37that we have this relevant AI tool which
2:20:40is going to allow us to uh generate the
2:20:42instant quote. So, I've pre-made this
2:20:44and I've actually sniped it from some of
2:20:45my accelerator resources. So, if you
2:20:47open this link, it's going to allow you
2:20:50to clone this into your relevance
2:20:51account. So, up here you can click clone
2:20:54and we'll just clone it into Morningside
2:20:58AI. And now this thing is going to be
2:21:01taking in a property type and a size and
2:21:02square feet. And then we have a basic
2:21:04LLM step here. This is a really really
2:21:06simple one. Again, like I said earlier
2:21:07in this video. The building really
2:21:10powerful and advanced functionality does
2:21:12take a lot longer. And if I was to do
2:21:14things that weren't just a very basic
2:21:15LLM step like this, then it would take a
2:21:17lot longer with different platforms you
2:21:18have to sign up for. The idea is to
2:21:20teach you how these things can connect.
2:21:22And then you can come into relevance
2:21:23here. Once you know how to connect voice
2:21:24flow to relevance through an API call,
2:21:26you can come in here and throw whatever
2:21:27the hell you want in and make it as
2:21:28advanced as you want. But in this case,
2:21:30we literally could have done this in
2:21:32voice flow if I'm being fully honest.
2:21:34But the idea of being able to access an
2:21:35external tool via an API is really the
2:21:38skill that I'm trying to teach you here.
2:21:39So, what this is doing is taking in the
2:21:41property type. We've got two options
2:21:42here and the square footage in a number.
2:21:46See here, it's it's a number. I can't
2:21:47just type in here. It has to be a
2:21:49number. So, that's what the API is going
2:21:50to be expecting. You can see here that
2:21:51it says it's expecting a number. And
2:21:54then we pass this into a basic prompt
2:21:55here that's saying the customer is
2:21:56requesting this um and these are the
2:21:58rough prices for the different square
2:22:00footage and different types. and the LLM
2:22:02in this case GPT4 mini just going to
2:22:04look over that take in their inputs and
2:22:05then give an output. So if I just go up
2:22:07here and I set it's an apartment and I
2:22:10mean I don't know what five how big is
2:22:12an apartment in square feet I don't know
2:22:15um but we can run
2:22:19that it's going to give us some kind of
2:22:22output and then it gives us the
2:22:24estimate. So it's saying regular
2:22:25maintenance cleaning for 60 deep
2:22:26cleaning for 120 move and move out etc.
2:22:29So that's the output that we're going to
2:22:30be sending back to voice flow and we're
2:22:32going to be turning that into a nice
2:22:33message to send through chat or through
2:22:35the phone. So in order to set this tool
2:22:36up in voice flow and be able to interact
2:22:38with it, we need to we can hide this
2:22:40make sure the tool's been saved and you
2:22:41have it in your account and then you can
2:22:43go to use here and just like we learned
2:22:45in theuh NA10 tutorial we can go to API
2:22:48here and then we get an API for us to
2:22:50use and again is a post request and it
2:22:53tells us how to use it. So we have
2:22:54params basically the inputs it's
2:22:56expecting of the property type and a
2:22:58string. You see how it's got two uh
2:22:59quotations? That means anything inside
2:23:01it is a string. And the square footage
2:23:02here, you can see it isn't in quotation.
2:23:04So, it's not a string. It's in this
2:23:05case, it's a number. And of course, we
2:23:07have the project ID here, letting
2:23:08relevance know which tool that we've
2:23:10created that we're actually trying to
2:23:11interact with. And so, all we need to do
2:23:12now to interact with this and set up our
2:23:14quote generator is to copy this link.
2:23:16So, this is the endpoint URL that we're
2:23:17going to be calling over the internet.
2:23:19And this is what it's expecting in the
2:23:20in the body of that post request. If you
2:23:22go back to voice flow here and we need
2:23:25to go to dev
2:23:27API, we're going to change this to post.
2:23:30Put in that
2:23:31URL. And in this case, seems that we've
2:23:33got a little bit of a different page on
2:23:34relevant for some reason. So we can
2:23:36actually do the authentication step,
2:23:37which is helpful. So you can click
2:23:38generate API key here. May have changed
2:23:40by the time you're watching this video,
2:23:41but should roughly be the same. And we
2:23:44can click deploy here. Make sure that
2:23:46the API is up and running for us to
2:23:47interact
2:23:48with. Actually, now that we've deployed
2:23:50it, we don't need the authentication
2:23:52anymore. That was a bit strange. Uh,
2:23:55actually, I will do the API key just so
2:23:57you guys see how this works. We can make
2:23:58it private here. So, now that we have
2:24:00our API key, we need to see how it's
2:24:02expecting to receive that API key via
2:24:04the HTTP request, which we're going to
2:24:05be sending from voice flow. Um, so if we
2:24:07scroll down here, curl is usually the
2:24:09one I like to go to. And there is a
2:24:10little bit I haven't really explained in
2:24:12terms of headers and bodies when it
2:24:13comes to API calls. But for now, just
2:24:15know that when you have a curl request
2:24:17like this, maybe it's even easier on the
2:24:18uh JavaScript. Here we have what's
2:24:20called headers. And this is basically
2:24:22like the the the envelope that you put
2:24:25the information in. So this up here is
2:24:27the information that we're putting in
2:24:28the request. It's going inside the
2:24:29envelope. And the headers and the method
2:24:32and the endpoint URL are like the stamp
2:24:34and the information that you put on the
2:24:35outside of the envelope to make sure it
2:24:36gets where it wants to go. So you can
2:24:38say that this endpoint here, that's what
2:24:40we call the endpoint. Same as up here.
2:24:41That's the the same as we have just
2:24:43here. Endpoint is like the address where
2:24:45the envelope is being sent to. We have
2:24:47the post, the method. Maybe this is like
2:24:50super fast mail or like overnight
2:24:52delivery or maybe like a parcel versus
2:24:54an envelope. Basically, the type of
2:24:56delivery that you're doing or type of
2:24:57request. And the headers include
2:24:58important stuff like the type of content
2:25:00that's inside it. So you might say this
2:25:02like this is there's written text or
2:25:03there's a letter in here. So it might
2:25:05seem a little bit complicated, but we'll
2:25:06fill this all in. Now, anytime you are
2:25:08making an API call, you'll see these
2:25:09headers around and you'll see the
2:25:11endpoint and the method and then the
2:25:12body which we're going to be setting up.
2:25:14So, this will all make sense in 2
2:25:15seconds, but let's just say for now we
2:25:17have content type and it's going to be
2:25:18application JSON. That's one of our
2:25:20headers. So, if we go back to Connor's
2:25:22cleaning, we open up the headers. We can
2:25:25go content type here as we saw here,
2:25:28right? So content type, we need to set
2:25:29it to
2:25:30application/json. Very common one that
2:25:32you're going to be
2:25:33using. And we have another one which is
2:25:36authorization. So this is basically the
2:25:38majority of the endpoints you're going
2:25:39to be seeing which is authorization and
2:25:41content type. And a lot of the time it's
2:25:42going to be an API key and application
2:25:45JSON. So then we can go back to
2:25:47relevance. We can get our API
2:25:50key. And now to set up the body of the
2:25:54request which is this. We can copy this
2:25:58and change this to raw. And we'll paste
2:26:01this in. we have the params or the
2:26:03inputs in this case, the property type
2:26:05and square footage for the tool. And
2:26:06what we want to do is insert the
2:26:08variables that we've got over here into
2:26:11these. So we can go inside these
2:26:12quotations here and we can go property
2:26:16uh type in this case and then for square
2:26:20footage we can go
2:26:22property
2:26:25size. And so because the property type
2:26:27is expecting a string as we can see if
2:26:29we go back here um to build
2:26:32It's expecting a string and this is
2:26:34expecting a number. The string must be
2:26:37wrapped around in uh these quotations
2:26:40and the square footage is a number and
2:26:42so we can just put the number because
2:26:44we've already converted it into a number
2:26:45here. Right? So that's probably the
2:26:47trickiest part of all the stuff I'm
2:26:48going to teach you today. And now that
2:26:49we have this set up, we can actually
2:26:50send a test request and we can say uh
2:26:53house. We can say 500 and test this API
2:26:58call. And there we go. It's complete. We
2:27:00have got back our answer the estimate
2:27:02etc. and all of the information is
2:27:04coming back from relevant AI as expected
2:27:06and now what we need to do is extract
2:27:07the information which we want which is
2:27:09this answer from this API call and from
2:27:11the response. So we sent a request and
2:27:13we got a response. Remember how we were
2:27:15using those terms before. We can click
2:27:16on answer and we can save it to and we
2:27:20can call it uh
2:27:22raw quote
2:27:24data because it is quite raw
2:27:28um raw quote data
2:27:33from. And now we have the information
2:27:35back from relevance AI. All we need to
2:27:37do is make it pretty and use AI to
2:27:39generate a message that summarizes the
2:27:41quote to the customer. So we can go a uh
2:27:45prompt step here. We can create a new
2:27:48prompt. Let's add in the conversation
2:27:50history for context. And I've got this
2:27:53prompt
2:27:55here. Write a short and clear
2:27:57explanation of this quote for the
2:27:58customers. uh we can go
2:28:01property property
2:28:03type for the customer's property type
2:28:06and we can put in the raw quote data
2:28:08here. So that's going to insert this
2:28:10with the customer's apartment or the
2:28:12customer's house and then we're just
2:28:13going to dump in the raw data that came
2:28:15back from relevance AI. Your response
2:28:16will be read over the phone. So it must
2:28:18be all in one paragraph and no longer
2:28:19than three to four short sentences. It
2:28:21should read like and I've given an
2:28:22example here of how I wanted to give the
2:28:24outputs.
2:28:28And so we can change this to
2:28:31quote quote response. And now we are
2:28:35ready to give this a spin. So if we run
2:28:38this whole thing from the top
2:28:40um
2:28:46run, how much for a weekly clean? So I'm
2:28:50going straight to the asking a question
2:28:51about price. You're asking about a
2:28:53pricing for our weekly cleaning service.
2:28:55Is that correct? Yes.
2:28:57Bam. Bam. Okay, sure. To give you an
2:29:00instant quote, I just need the property
2:29:01type and size and square feet. Is the
2:29:03property a house or an apartment? It's a
2:29:10house. How many square feet is it? 500.
2:29:13I don't know. Is that a big That sounds
2:29:14very small to
2:29:16me. I don't know if 500 I have I have no
2:29:19I have no clue about the sizes of houses
2:29:20if 500 is normal or
2:29:23not. And there we go. Based on your 500
2:29:26foot house, we have four cleaning
2:29:27packages. Regular maintenance cleaning
2:29:29at $90 for standard weekly cleaning,
2:29:31deep cleaning, ra. So, we're giving them
2:29:33a quick summary based on their 500 ft
2:29:35house or x number of square ft x
2:29:38property type and giving them a quick
2:29:40summary. So, that's pretty cool. We're
2:29:42using relevance tool and we're getting
2:29:43this information back. Now, there is one
2:29:44more step that I have added onto this
2:29:46this I get this thing where if I've gone
2:29:48this far with you guys, I may as well
2:29:49add in like the rest of it to make it
2:29:51actually a bit more useful. um which is
2:29:53a quick lead capture using Google Sheets
2:29:55and Make.com. So, I couldn't leave you
2:29:57guys hanging on this. I thought I may as
2:29:59well throw it in there. So, stick with
2:30:00me because this is really where you're
2:30:01going to be like, "Oh, this is this is
2:30:03uh opening my eyes to to what you can do
2:30:04with these kind of platforms." So, the
2:30:06reason we're adding this on is because
2:30:07the person has asked about pricing or
2:30:09they're directly interested in some kind
2:30:10of services and we've given them an
2:30:12instant quote and now we're trying to
2:30:13immediately follow that up with, hey,
2:30:14look, give us your details and we'll be
2:30:16in touch and we'll get that service
2:30:17booked in right away. So, we can jump
2:30:20into a uh message block here. This is
2:30:22I'll just paste this in to keep things
2:30:24nice and quick here. Please provide your
2:30:25name and phone number and I'll get one
2:30:27of the team to call you to find a time
2:30:28that works. Now, one thing I will change
2:30:29is as you saw on that last run, it had a
2:30:32question at the bottom. It's like, which
2:30:33one are you interested in? I would
2:30:35probably try to remove that.
2:30:38Um, do not ask a question at the end.
2:30:43Just give them this description.
2:30:52So that's just going to end the message
2:30:53and saying, "Look, this is a quote."
2:30:54Bam. And then the next message I get is
2:30:56this. Please provide your name and phone
2:30:58number and I'll get one of the team to
2:30:59call you to find a time that works. And
2:31:01so we want to save this entire user
2:31:02reply with a capture step. Capture step.
2:31:05There we go. Change this to entire user
2:31:08reply. So we want to capture all of the
2:31:09information that they send or over the
2:31:11phone or either through chat. So in this
2:31:12response, they're going to say their
2:31:13name and their phone number, right? and
2:31:14we need to extract those
2:31:17out. I'm going to put this here and do a
2:31:20set step. We're going to use a prompt.
2:31:23So, we're setting this with
2:31:26AI. And we're going to create a new
2:31:28prompt. Um, let's just set the variable
2:31:30first. Let's say this is going to be put
2:31:32into last
2:31:35response. Make a new prompt
2:31:39here. And this is the prompt that we're
2:31:41going to be using. So, we have just
2:31:42asked the user to provide their name and
2:31:43phone number. We need to attempt to
2:31:44extract the information and then confirm
2:31:46it with them. Here is their reply last
2:31:48utterance which they've just provided
2:31:49and we've captured. If there is a valid
2:31:51name and phone number present, then you
2:31:52must do a confirmation eg okay quickly
2:31:54to confirm your name is this and number
2:31:55is this. Is that correct? However, if
2:31:57one or both are missing or appear to be
2:31:59invalid, you must output only retry as
2:32:01your response and nothing else. This
2:32:03retry variable will be checked and if it
2:32:05matches exactly, then it will trigger
2:32:07another attempt to capture. So either a
2:32:09write a short and sweet confirmation
2:32:10message or b output retry for another
2:32:13attempt at capturing. So what we're
2:32:15doing here is using AI to analyze the
2:32:17response and say look we're looking to
2:32:18pluck out a name and phone number and we
2:32:20want to also confirm that cuz this is
2:32:21likely going to be over the phone. And
2:32:22if the AI doesn't see a clear valid
2:32:25phone number and a valid name then it's
2:32:26going to output only the word retry. So
2:32:29give us a response. If it's good to go
2:32:31and we can move on to the next step. If
2:32:33it says retry then we're going to try to
2:32:35retry. Um, now of course this retry
2:32:37doesn't actually do anything unless we
2:32:38build the functionality in to look for
2:32:40that retry keyword, which we'll do in a
2:32:43second untitled prompt. We can change
2:32:45this to um extract what attempt extract
2:32:51name. And then we can go to um
2:32:54conditions here and
2:32:58go if last response which is the
2:33:01variable that we're saving the output of
2:33:03that prompt into here. Um whether it's
2:33:05going to be retry or the say just to
2:33:06confirm is this your phone name and
2:33:08phone number. Um if last response is
2:33:12retry or even just to make it a bit more
2:33:14flexible contains retry unless unless
2:33:17the person's name is like Bill Retry
2:33:20Smith then uh this should be fine. And
2:33:21actually, just to make this look a bit
2:33:23cleaner, I might change this to um if
2:33:24last response um does not contain retry.
2:33:30Then we're going to go up
2:33:35here and we are going to send the last
2:33:39response because as we said, it's going
2:33:41to either generate the confirmation
2:33:42message or it'll output retry. So if
2:33:44it's valid, it'll and we put up last
2:33:46response here. They'll say their name
2:33:48and phone number. We'll analyze it and
2:33:50then we'll go great. It doesn't contain
2:33:51the word retry in it. And bang. Hey,
2:33:53just to confirm this is your name and
2:33:55this is your phone number. And then for
2:33:56the choices here, we can go and use our
2:33:58handy dandy. They're going to be giving
2:34:00us a yes or no answer to this. So, we
2:34:01can put in these two triggers, yes and
2:34:05no. And so, if they say no to the
2:34:07confirmation, say no, that's not my
2:34:09correct phone number or name, then we
2:34:11need to have some kind of retry um logic
2:34:13here. I usually like to make my retries
2:34:16um a orange color. And then my failure
2:34:19is a red. Um, so we
2:34:25say, "Okay, let's try that again. Can
2:34:27you please give me a full name and phone
2:34:28number, please?" And then we're going to
2:34:30send them all the way back to this step
2:34:31here. So they're basically going to
2:34:33recapture their information and then put
2:34:35them through this process. And this is a
2:34:36loop that can be done over and over and
2:34:37over again. So, and then we also need to
2:34:39deal with this else step. So if the word
2:34:42does contain retry and it has said,
2:34:43"Hey, look, this isn't a valid phone
2:34:45number or a name," then we need to deal
2:34:47with that as well. So, we can come down
2:34:48here and go to message. Pop it under
2:34:52here. And I've got a message for this.
2:34:54Sorry, I didn't quite get that. Can you
2:34:55please give me a full name and phone
2:34:56number so a member of our team can get
2:34:58in
2:35:00touch. Um, right click on this and you
2:35:03can go block color. Change it to an
2:35:04orange. And then we're going to be going
2:35:05back to here as well. So, it's helpful.
2:35:08You can click on these arrows. So, the
2:35:10lines here, and you can change them to
2:35:12the same color. So, we want to make them
2:35:14a bit more obvious that they're coming
2:35:16in from places we
2:35:19expected. We can make this an orange one
2:35:22as
2:35:24well. And this one,
2:35:27too. So, if they come in and they say,
2:35:29"Hey, my name's Bill and my phone number
2:35:30is
2:35:3102111." And it comes in here and it
2:35:33goes, "Hey, that doesn't look like it's
2:35:34proper." It's going to send a retry as
2:35:37the output. We're going to pick it up
2:35:38here and it's going to say, "Sorry, I
2:35:39didn't quite get that." and it's going
2:35:40to come back up and they're going to be
2:35:42expected to give it again and it will go
2:35:43through and then once we got a valid
2:35:45name and phone number and it's not
2:35:46outputting retry, then it's going to go
2:35:48through here. It's going to spit that
2:35:50out and say, "Hey, just to double check,
2:35:51this is your phone number and email
2:35:53before we proceed." Yes and no. No is
2:35:55going to be handled there and it's going
2:35:56to take them back to the first step
2:35:57again. So, that's some nice um error
2:36:00handling and sort of looping that you're
2:36:01going to need to be building into a lot
2:36:02of your conversational AI agents,
2:36:04especially on voice flow, right? And so,
2:36:05the last steps here are some quick
2:36:06variable extractions. So we can go to
2:36:08div logic here and go to set and we're
2:36:11going to extract the name and the phone
2:36:13number. So we'll go prompt. Holy moly,
2:36:16it is it's bloody hot here. The variable
2:36:19that we want to save this, we were going
2:36:21to be extracting the name. So we'll add
2:36:23a new variable called name. The user's
2:36:28name. Um, and this is going to be called
2:36:31uh extract
2:36:35name. Here's a quick and easy prompt.
2:36:37You can pause it to take a look at that.
2:36:38It's just going to extract their
2:36:39name. We need to add another variable.
2:36:42This is going to be their phone
2:36:46number. And then we need another
2:36:51prompt. Just paste this one in here.
2:36:53Pretty basic. Again, pause it if you
2:36:55want to take a
2:36:56look. And I haven't named that prompt.
2:36:58It's going to annoy
2:37:00me. Extract phone
2:37:05number. And now we're going to have
2:37:06their name and their phone number
2:37:07extracted out of this response.
2:37:11And oh, and actually we need to add in
2:37:15the conversation history there so that
2:37:18it actually has the this information in
2:37:22it. And then we
2:37:30go great. Let me get that added into our
2:37:32system. This buys us a bit of time as we
2:37:35use our um make web hook which we're
2:37:37going to set up now. So the next step is
2:37:39to get a Google sheet set up and to use
2:37:41make.com to uh take this data and shoot
2:37:44it into a Google sheet. So to do that um
2:37:46I will leave a link on I mean you can
2:37:48just search it up. It's make it's
2:37:50make.com. All right. I'll save you the
2:37:52hassle. So you sign into make.com create
2:37:54an account whatever you want to do or
2:37:57need to do.
2:38:04Go to scenarios and then we're going to
2:38:06create a new
2:38:08scenario. I'm going to build from
2:38:10scratch here. I'm going to get all that
2:38:12rubbish out of the way. I don't know why
2:38:13it's acting like I'm some rookie here.
2:38:16Um, and then we need to go to web hooks,
2:38:20custom web hook. We're going to
2:38:23um we're going to add in a new web hook.
2:38:26This is going to be conors cleaning lead
2:38:29capture. going to save that. It's going
2:38:32to create this web hook here. I'm going
2:38:34to copy this edges to clipboard. We're
2:38:37going to come back to our build and then
2:38:39we're going to go to
2:38:41the API step. So, what this is doing if
2:38:44you're a bit new to to web hooks and and
2:38:45API calls and stuff. What we can do here
2:38:48on make is set this up to basically
2:38:50listen. It's a URL. You know how we had
2:38:52the endpoint? The end point, this thing
2:38:54here that it's given us that I've just
2:38:56copied to our clipboard. That's like the
2:38:58address, remember? So if you put it
2:38:59write it on the on the letter, that's
2:39:01where it's going to go. This allows us
2:39:02to basically send data um via API call
2:39:05um from voice flow to make and it's
2:39:08going to catch it here. And this little
2:39:09lightning bolt means that anything we
2:39:11build after is going to be triggered
2:39:13whenever it receives one of those uh
2:39:15whenever a new bit of mail arrives. It's
2:39:17going to then trigger this multi-step
2:39:18process. So we're going to put this into
2:39:21voice flow. We're going to trigger it
2:39:22and make sure it knows what data to
2:39:24expect. And then we're going to be able
2:39:25to use that data and put it into Google
2:39:27Sheets. So stick with me here, but this
2:39:29is another very very essential skill in
2:39:31AI automation is how to set up a webbook
2:39:33um and use it within different apps. So
2:39:35we have our webbook
2:39:37here. We've copied the address to
2:39:39clipboard. We're going back setting up a
2:39:41get request here. So it's not a post
2:39:43request this time. We're just uh getting
2:39:45and we're technically not getting data.
2:39:47A get request is a much more kind of
2:39:48quick and dirty request. Um and as
2:39:51you'll see, we kind of just tack on a
2:39:53bunch of information after this. Um we
2:39:54can do it through what's called
2:39:56parameters. Here we want to be sending a
2:39:59property. Oh, let's just say
2:40:01property
2:40:04type. Actually, let's do this properly.
2:40:07Let's go time
2:40:09um
2:40:11timestamp. So, in this Google sheet,
2:40:14you're going to want to know when the
2:40:15different leads came in. So, we can go
2:40:16uh time stamp and get the time stamp
2:40:18from voice flow. That's a default
2:40:20variable that they are automatically
2:40:22filling out for you. So, one of the
2:40:23things, one of the rows in the
2:40:25spreadsheet is going to be uh the time
2:40:27stamp. We're going to add another and
2:40:29we're going to go um name. We're going
2:40:32to put in the name here. So, now all the
2:40:35cogs in your head should start turning
2:40:36as we put this together. And we're going
2:40:38to go uh phone number and we go bracket
2:40:44phone number
2:40:47number. And if we add another, we can go
2:40:51property
2:40:52type. Let me go prop property
2:40:58type. We can go
2:41:00property
2:41:05size and go property size. We can add
2:41:10another one. We can go
2:41:12quoted prices so that the sales team
2:41:15knows what we actually told them. in
2:41:17case you're playing around with pricing.
2:41:18Um, raw quote data. And I'll probably
2:41:22throw in one more here, which is their
2:41:24first question. Um, first question,
2:41:31uh, user first reply. So, that might
2:41:33give context to the sales team like what
2:41:34did they actually contact us for in the
2:41:36first place? Maybe helpful, maybe not.
2:41:37But we can just send this all over to
2:41:38the Google sheet.
2:41:44And so now we can see this is uh as this
2:41:47thing's spinning around, it's basically
2:41:48waiting for us to send some data to it.
2:41:50It's basically sitting there at the
2:41:51mailbox like waiting for it to come
2:41:54through. Um and we can go send here and
2:41:57I'm going to put in uh gosh dug myself a
2:42:00hole here. Um let's go name
2:42:04Liam. Um house probably size. Um
2:42:09um lots of money. Um how much for
2:42:12cleaning yarning?
2:42:13um send. And if we go back, bam,
2:42:16successfully determined. And what we've
2:42:18done and determined means is that make
2:42:20has received the the the request that we
2:42:23sent. And it now it knows that we're
2:42:24going to be sending it a time stamp and
2:42:26a property type and this this really
2:42:29really key skill to understand because
2:42:30now when I go oh save now when I go to
2:42:33here and I go Google Sheets and I go add
2:42:37a
2:42:38row, um you will need to set up your uh
2:42:41Google Sheets connection here. So, you
2:42:43just sign in with Google, add your
2:42:44connection in. Um, I am going to have to
2:42:49create a new spreadsheet for this
2:42:55quickly. I'm going go timestamp.
2:43:00So I'll zoom this up. Time stamp name
2:43:05phone
2:43:06property type property
2:43:15size. 1 2 3 4 5 6 7. Yeah, we've got
2:43:19them all. Right. So then I can
2:43:22go. So call this my Connor cleaning
2:43:26support agent leads.
2:43:30So now I want to go back to make and I
2:43:33want to click here to
2:43:37um Connor. There we go. Connor's
2:43:40clearing. Why do I keep spelling
2:43:42cleaning
2:43:47cleaning? Now we have the spreadsheet
2:43:49set up. The sheet name is just going to
2:43:51be sheet one. Sheet one. Does the table
2:43:54contain headers? Yes, it does. And now
2:43:57we get to put in all of our values. So,
2:43:58bam. Time stamp. Pop that in there.
2:44:01Name, pop that in there. Phone, pop that
2:44:04in there. Property type. So, you see
2:44:08what such a key skill that uh I really,
2:44:10really want you guys to learn. Um,
2:44:12because this is a lot of like if you
2:44:14just have a couple of these key things
2:44:15using these APIs and say voice flow or
2:44:17in NAD, building tools and relevance and
2:44:19then using them via API, knowing how to
2:44:22set up web hooks and then send data in
2:44:24between them and to set this data
2:44:25structure. You send that first initial
2:44:27batch over to make in a test test
2:44:29request. It's going to go okay this is
2:44:31what they're going to send me in future
2:44:32and then that's locked in and you can
2:44:33send thousands and thousands of requests
2:44:35through it and it will operate as you
2:44:36expect and first
2:44:38question. So uh it's got all those other
2:44:41rows we don't need to worry about. Save.
2:44:43All right. And so that should be
2:44:44working. We can switch this to
2:44:46immediately. Oh, I need to save it.
2:44:49Save. Um I'm going
2:44:51to immediately as data arrives. So, this
2:44:54is going to be waiting all the damn
2:44:55time. Actually, I'm I've got a whole
2:44:57bunch of other stuff running in make for
2:44:58my my various businesses. So, I'm just
2:45:00going to have this set up. If you wanted
2:45:01this to run around the clock, you turn
2:45:03that on. I'm just going to turn it on to
2:45:04run once. And this is just going to be
2:45:06sitting there waiting for data. So, what
2:45:07we can do is just finish this off
2:45:08quickly. So, I'll just paste this
2:45:13in. Got this all done. Thank you. That's
2:45:16the end of it. Um, what we want to do is
2:45:19have a failure path. So, in case this
2:45:21fails, we just pop that down there and
2:45:26say, "Darn it, something went
2:45:31wrong." I'm going to put that as a
2:45:36red. Um, and then this one is like that.
2:45:40All righty. So, if we look at make, it's
2:45:42still expecting us to send some data to
2:45:43it. Let's hope that we've done all of
2:45:46this right and it's going to work first
2:45:48time, but you never really know, dear.
2:45:50So, let's give this a run and we'll try
2:45:52to go through that second path and we'll
2:45:54try to get that stuff added into the
2:46:01spreadsheet. I need weekend cleaning,
2:46:05please.
2:46:11So asking a question is going straight
2:46:12to price. You're looking for weekend
2:46:15cleaning services and want to know the
2:46:16pricing. Is that correct? Yes. So our
2:46:17router is going to pick that up and say
2:46:19he's looking for pricing. Let's send him
2:46:20to the quote tool. Okay. Sure. To give
2:46:22you an instant quote, I just need the
2:46:23property type and size and square feet.
2:46:24It's a property. Let's go aart
2:46:27mint. It's detected that correctly. And
2:46:30how many square feet? It's 500.
2:46:34typical
2:46:37apartment
2:46:39size
2:46:40and square
2:46:44feet. Okay, so 500 isn't actually
2:46:47ridiculous. Um, great. And then we get
2:46:49our answer based on your 500 foot
2:46:50apartment. Bam, bam, bam. There's all
2:46:52the prices. Please provide your name and
2:46:53phone number and I'll get one of the
2:46:54teams to call you with a time that
2:46:55works. Let's uh give this a a test. See
2:46:58if it's actually picking up on the fail.
2:47:00Sure. My name
2:47:03is and
2:47:05phone
2:47:11is boom. Sorry, I didn't quite get that.
2:47:14So, it detected that it wasn't right.
2:47:16So, we got the retry output. We got the
2:47:18retry output from this, which is what we
2:47:20wanted. Sorry, I didn't quite get that.
2:47:21Can you give me again? Um Liam
2:47:26Otley number is um 021
2:47:34021. That's what numbers look like here
2:47:36in New
2:47:38Zealand. Okay. Just to confirm quickly,
2:47:40your name is Liam Mley and your phone
2:47:41number is that correct? Yes, sir. Name
2:47:45and phone
2:47:47number. Oh, what has it done
2:47:50there? Oh, dame.
2:47:53I don't know why it's I just want the
2:47:56phone number. So, we have got a little
2:47:57bit of an error there. I just go back
2:47:58and tweak the prompt. Make sure it's
2:48:00like only get the phone number. We don't
2:48:01want anything apart from numbers here.
2:48:03Great. Let me get the added to the
2:48:04system. And then if we go back to make,
2:48:06we see these are all green now. And we
2:48:07see these dots. So, this is the
2:48:09information that came through. Namely,
2:48:10Mly phone number. And so, here's the
2:48:12little mistake where we had a new line.
2:48:14Apartment. Bam, bam, bam. All of that
2:48:16information. And then it's added it into
2:48:18Google Sheets. Here we see updates,
2:48:21updated number of rows, all the values
2:48:22that it's updated and we can go
2:48:28to. Okay. Holy moly. Right. We're ready
2:48:32to put this thing on our website and to
2:48:34also put it on a phone number. So, let's
2:48:36just finish the job. Guys, I'm uh
2:48:38getting real hungry, but we can uh we
2:48:40can push through. So, I will just turn
2:48:42this on actually so that if we are
2:48:44testing it on the web and over the
2:48:45phone, um it's ready to receive. Um, if
2:48:49we do want to be pedantic, I would go
2:48:51back and I would change I have to do it.
2:48:54It's going to piss me
2:48:56off. I'll put their name. Oh, that's
2:48:59why. I'll put their phone number only.
2:49:02My bad. I'll put their phone number
2:49:11only. And we do have these fail points
2:49:14here. Um, I'm not going to bother
2:49:16filling them out. I think you guys can
2:49:17figure out based off how I've handled
2:49:19this, how you can handle these as well.
2:49:21So, what you'll find is when you're
2:49:22building these, these kind of fail like
2:49:24error handling um is kind of a an
2:49:27endless thread that you keep pulling.
2:49:28It's like, oh, well, now I've got to
2:49:29handle this, this, and this this. So, um
2:49:31I'm not going to this is a prototype.
2:49:33I'm not going to be doing all of the the
2:49:34error handling for you here. In the
2:49:36template, there is actually a little bit
2:49:37more of it. Um some better examples. So,
2:49:39maybe if you import that, you can just
2:49:40steal the work that I've done there. But
2:49:43what we need to do now is we can publish
2:49:45this thing. We'll call it
2:49:48V1 first
2:50:00drop. All right. So now it's published.
2:50:03We can add the agent to a website. Let's
2:50:04click on that. And that takes us to this
2:50:06integrations tab.
2:50:08Um let's put this down. I don't need to
2:50:10see that. Um, they've got a new version
2:50:13of it. That's good to know. I said
2:50:15installation is pretty straightforward.
2:50:16So, we can just click copy here. And
2:50:18then I'm going to open
2:50:21up I'm going to open up brackets here
2:50:24just to give you a demo of a website. I
2:50:26use this in all my tutorials. It's
2:50:27really easy to spin up. Um, I will leave
2:50:29a link to this template if you want it.
2:50:31Um, and also some instructions on how
2:50:32you can open up a website. I know this
2:50:34looks like code and it's all scary, but
2:50:36um, this is just allowing me to spin up
2:50:38a website very quickly. So, I'll leave
2:50:40the template file. All you need to do is
2:50:42once you've downloaded the template
2:50:43file, you need to download brackets,
2:50:46which is the software. You can go file,
2:50:48then open folder, and then you want to
2:50:50click on the folder when you've unzipped
2:50:52it, and it's going to open up the whole
2:50:53folder. And then you'll get all of this
2:50:55uh opened up like this. And you see all
2:50:57of these files ra. All you need to do is
2:50:59click on the
2:51:00index.html. And then you'll see
2:51:02something similar to this. Well, I'm
2:51:04going to scroll down to the bottom of
2:51:05the index.html. I'm going to delete this
2:51:07old voice agent I was testing on here.
2:51:10Drop this in here. Paste that. And then
2:51:12save it. Command S. Click this little
2:51:14button up here. And it will show us a
2:51:18local version of the website running on
2:51:20our computer here through brackets. So
2:51:22here's my man with a magnificent
2:51:24beard.
2:51:26And we have the tester agent bubble down
2:51:30here. And there we go.
2:51:36I want to know where you
2:51:40live, my
2:51:44guy.
2:51:49Yep. Woo. Okay. Uh I didn't even program
2:51:53that in there. Maybe you just thought it
2:51:54was inappropriate. Um but we've got it
2:51:57working on a website. Now, if we pop
2:51:58back over to um uh Voice Flow here and
2:52:02you go to the integrations, the widget,
2:52:04you see we've got this test your agent
2:52:05thing. So, down here, we can play around
2:52:07with the look and feel of it. I'm not
2:52:09really going to get into this here.
2:52:10There's quite a lot to play around with,
2:52:11but basically all of what you see on
2:52:14here can be changed around. Different
2:52:16logos, different text here, different
2:52:18icon, etc., different colors, and you
2:52:20can just make it look and feel however
2:52:22you want it to. So, I'm sure you guys
2:52:24are big enough and ugly enough to figure
2:52:25that out yourself. we'd probably want to
2:52:27switch over to uh this here. One thing
2:52:29you would want to do is turn off powered
2:52:30by voice flow so it's not uh sending
2:52:33traffic to them when it's on your own
2:52:34website. And that's about it. For the
2:52:36sake of time, I'm not going to go
2:52:37through the entire flow again here. Just
2:52:40know that the functionality that we
2:52:41built that I just showed you in the
2:52:42builder is going to work cuz we just
2:52:43deployed it. Like this is exactly what
2:52:45we're interacting with. So, it's all
2:52:47working here. The only step to do now is
2:52:49to put it on this phone number so we can
2:52:51have a chat to it over the phone, which
2:52:52we're going to do now. To do that, we
2:52:53need to go to the telefan bit here. It
2:52:55is in beta right now, but for most of
2:52:56you watching it is not going to be by
2:52:58the time you you are watching this. So,
2:52:59we need to set up a phone number from
2:53:01Twilio, import it, and then connect our
2:53:03agent and its functionality to that. So,
2:53:04we can go import number. You'll see that
2:53:07we have this information here. So, we
2:53:08can use Twilio or Vonnage. Twilio is
2:53:10usually the go-to here. So, if we click
2:53:12on learn more, then they're going to
2:53:14help us. Basically, the best way to make
2:53:15sure you're getting the most up-to-date
2:53:16information is go to the docs of the
2:53:18platform. Finding and reading and
2:53:19extracting information from
2:53:20documentation on these kinds of
2:53:22platforms is another key skill that you
2:53:24need to pick up to succeed in the space.
2:53:25So if we go to the docs here, we go to
2:53:29um voice phone number setting up Twilio
2:53:34integration and they have a video here
2:53:36adding a phone number to your agent. So
2:53:37if you ever get stuck, you know, you've
2:53:38got documentation here and for all of
2:53:40the other platforms, but they'll keep
2:53:41updating these videos if things change,
2:53:43which they likely will as this voice AI
2:53:45space really takes off. So, if we go to
2:53:48Twilio, you will need to sign up and
2:53:50create an account on
2:54:04Twilio. All right, so we are logged into
2:54:06Twilio. You'll need to create an account
2:54:08for most of you, but Twilio is a uh
2:54:10phone number provider that you can
2:54:11connect to and interact with over the
2:54:13internet. super helpful when you can buy
2:54:14lots of numbers from different locations
2:54:16and stuff. When it comes to phone
2:54:18numbers, it can there's a lot of rules
2:54:19and regulations around different like it
2:54:21varies a lot from country to country. So
2:54:23depending like if you're in Germany, I
2:54:24believe in order to get a German phone
2:54:26number, you need to have a company
2:54:27registered and get the number through
2:54:29your company registration and provide
2:54:31those details. So can be difficult. I'm
2:54:33just going to show you how to use a uh a
2:54:34US-based number here. So we can go over
2:54:36to the phone numbers on the left here.
2:54:38There may be some setup that Twilio walk
2:54:40you through. It can be kind of annoying
2:54:41sometimes. They say you need to do all
2:54:43of these declarations and forms and
2:54:45stuff, but for the most part, it should
2:54:46be fairly straightforward if you follow
2:54:48their setup instructions when you create
2:54:49your account to then come over and go to
2:54:51your phone numbers and manage and go to
2:54:54buy a
2:54:56number. Now, unless you have other
2:54:57purposes you want to use this for, you
2:54:59can just snag any random one if you're
2:55:00following this tutorial. Um, if you're
2:55:02obviously doing this for a client, you
2:55:03could get one that's matched to their
2:55:04location or their their state or even
2:55:06their city. And when you click buy, you
2:55:08can see there's all these kind of
2:55:09registrations and RAR you need to do.
2:55:11But thankfully, voice is uh is ones that
2:55:14don't need all of that. And you've got
2:55:15global routing, etc. So, you can come
2:55:17down here and buy this. It's going to be
2:55:18a dollar a month. I know cost of
2:55:20starting up a business is ridiculous
2:55:21these
2:55:22days. How dare they? But just to walk
2:55:25things through and do it with you, I'm
2:55:26going to buy this number even though I
2:55:27really don't need
2:55:29it. There we go. I've got my number. I'm
2:55:33just going to copy this so I don't
2:55:35forget it. 517. And then I go to active
2:55:39numbers. I got my guy. Where is he? 517
2:55:43down
2:55:44here. Copy this bad boy. Go back to
2:55:46Twilio. Put our number in here. Our
2:55:49Twilio account SID. You need to pop back
2:55:51to Twilio. You go to your account
2:55:54dashboard. This was a bit confusing when
2:55:56I first did it. So, I'm glad I'm glad I
2:55:58can help you guys through it because it
2:55:59was a bit of a pain in the bum for me.
2:56:01You think all this information is going
2:56:02to be like tucked away in the side
2:56:04panels, but no, it's just on the
2:56:05dashboard. Who the hell puts your or
2:56:07tokens and stuff on a dashboard? I mean,
2:56:09it's when you know it's there, it's
2:56:10helpful, but it was a bit confusing at
2:56:11the time for me. Um, I'm going to paste
2:56:15this in
2:56:18here. We're going to copy this or token
2:56:22and paste this in here. And we're going
2:56:24to say this is my you YouTube tutorial.
2:56:29And I'm going to import that bad boy.
2:56:31Select environment. I'm just going to
2:56:33put it on development for now. uh we can
2:56:35assign
2:56:36that. And with that, you can see if I
2:56:38click unassign here, that means I'm
2:56:39going to take it off. So, it's it's
2:56:40assigned and it should be ready to have
2:56:42a chat with. So, if I click to copy this
2:56:44number here, I go back to my little
2:56:47uh uh website code here. If I go find um
2:56:52plus one. Oh, there we go. So, we have
2:56:54the call. Um this is in the header. Uh
2:56:58this is just me p wanting to put it on
2:56:59the site for you guys. So, you guys
2:57:01don't necessarily need to copy this. I'm
2:57:03going to save that. If I go back to my
2:57:05website here, now I have the actual
2:57:06phone number. And if I pull up Skype on
2:57:10my phone, you guys, if you're in the US,
2:57:12you can just or wherever your number is
2:57:13based, you can just call it stage plus
2:57:17one. All right, let's give our mate a
2:57:20ring.
2:57:24Welcome to Conor's Cleaning. How can I
2:57:25help? Hey, uh, I was just wondering, uh,
2:57:28where you guys are located.
2:57:38business. Is that correct? Yeah, that's
2:57:41correct.
2:57:49We're located at 247 Beacon Street in
2:57:52Boston, Massachusetts 0216. We serve the
2:57:55greater Boston area, including
2:57:57neighborhoods like downtown Boston,
2:57:58Cambridge, Somerville, Brooklyn, and
2:58:00several others. Is there anything else I
2:58:03can help you with today? Sweet. Okay, so
2:58:05that's done the
2:58:07uh knowledge the knowledgebased route.
2:58:10Now, let's give it another spin. Call
2:58:12him again. Okay.
2:58:15Hey, welcome to Connor's Cleaning. How
2:58:17can I help? Hi there. I'm just getting
2:58:20interested in getting uh some cleaning
2:58:22for my place on the weekends.
2:58:29It sounds like you're interested in
2:58:30booking a cleaning service for a
2:58:32licensed weekend. Is that correct? Yep.
2:58:41Okay. Sure. To give you an instant
2:58:43quote. I just need the property type and
2:58:45size and square feet. Is the property a
2:58:47house or an apartment? It's a house.
2:58:56Sorry, I didn't get that. Is it a house
2:58:58or an apartment? It is a house, a home.
2:59:06And how many square feet is it? It's
2:59:08about a thousand.
2:59:15One sec while I generate your quote.
2:59:18Based on your requirements, we can offer
2:59:20four different packages. Regular
2:59:22maintenance cleaning at $90, which
2:59:24covers standard weekly or bi-weekly
2:59:26cleaning. Deep cleaning at $150 for a
2:59:30more thorough top to bottom refresh,
2:59:32move in, move out. Cleaning at $170 for
2:59:36preparing or cleaning after a home
2:59:38transition, and post renovation cleaning
2:59:40at $200 for handling construction dust
2:59:43and debris. Please provide your name and
2:59:45phone number and I'll get one of the
2:59:46team to call you to find a time that
2:59:48works. Sure. My name is Liam Otsley and
2:59:51my phone number is 021 uh 1111 9792.
3:00:03Okay, just to quickly confirm, your name
3:00:05is Liam Oxley and your phone number is
3:00:0802119792. Is that correct? Yep, that's
3:00:11spot on.
3:00:19Great. Let me get that added into our
3:00:20system. All done. Thank you.
3:00:24Boom. All
3:00:27righty. That is done. That is dunzos.
3:00:30All right. So, you guys learned a lot in
3:00:32that one. voice integration, website
3:00:34integration, um connecting web hooks to
3:00:37make, setting up make automations,
3:00:39sending data over, connecting relevance
3:00:41AI tools into into voice flow. Um basic
3:00:46integration with a CRM, in this case,
3:00:47it's a sheet, but there's so much in
3:00:49there, guys. I hope you really really uh
3:00:50learned a lot from that. This has been a
3:00:52big one. And we've still got uh one more
3:00:54to go. So, I hope you're sticking with
3:00:56us. Um but going back to our Figma here,
3:01:00um we have ticked off all of this. So,
3:01:01we have it as a web chat widget and we
3:01:03have it as a a phone number. Now, as far
3:01:05as I know, you can have both options for
3:01:06the same agent on voice. You can have it
3:01:08on the website and over voice. You don't
3:01:10need to duplicate it and sort of define
3:01:11what modality it's going to be. So,
3:01:14we've ticked off all the boxes for this.
3:01:15All of the resources will be in here.
3:01:17All of the prompts, a template for that
3:01:18whole final build as well. If you just
3:01:20want to snag all my hard work and go and
3:01:21sell it to someone, again, I don't I
3:01:23really don't care. Um, that's what these
3:01:24videos are for. And we're getting into
3:01:26agent build number four now.
3:01:33All righty. So, last but not least is an
3:01:36agent built on my own software. So, I
3:01:38didn't want to make this. It's not about
3:01:39me selling you or getting you to use my
3:01:41software. So, I thought I'd put it at
3:01:42the end just so you know that I wasn't
3:01:43really This is about you guys learning
3:01:45and my software happens to help you put
3:01:47an agent onto WhatsApp very very easily.
3:01:49So, that's why it's included in here.
3:01:50But again, this is nonsponsored,
3:01:52nonpromoted, non whatever. I'm just
3:01:54really trying to share with you what I
3:01:55think is a really valuable skill set to
3:01:57have. All right. Now getting into AI
3:01:58agent build number four. This is going
3:02:00to be tada a WhatsApp based ARI customer
3:02:04support and lead generation agent built
3:02:06on agentive my software. So this is a
3:02:09noode uh AI agent builder that is built
3:02:11on top of OpenAI's assistance API. So
3:02:13you're technically using your OpenAI
3:02:15account getting very very cheap rates on
3:02:17the uh token usage that you're running
3:02:19through this agent. But Agent just
3:02:21allows you to build on top of it very
3:02:22easily. but more importantly to deploy
3:02:24these agents not just onto web chat
3:02:26widgets like we've done with voice flow
3:02:27but easily onto things like WhatsApp and
3:02:30Instagram etc. So that's really the key
3:02:31thing that Agenda focuses on doing right
3:02:33now is making it easy for you to get
3:02:35your agents onto these platforms. So as
3:02:36you can see it's a fairly similar build
3:02:38to what we just did on voice flow in
3:02:39terms of functionality. We're going to
3:02:40be having a uh a knowledge base that we
3:02:43can ask questions over. It's going to be
3:02:45able to generate another instant quote.
3:02:46So we'll just quickly connect that same
3:02:48relevance tool here. And finally, we're
3:02:50going to do a lead capture, but this
3:02:51time it's going to be done through Air
3:02:52Table. So, I want to mix it up and show
3:02:54you how you can connect your agents to
3:02:55Air Table, which is a very, very common
3:02:57integration that you're going to need to
3:02:58know. And the difference between this
3:03:00agent, you're going to see that it's
3:03:01much much faster to build. This is not
3:03:02meant to be a side-by-side comparison of
3:03:05what's better, how much faster. It's
3:03:06just that when you build on a more
3:03:08conversationalbased uh AI agent platform
3:03:11like agentive which is built on top of
3:03:12the assistance API, it's a very
3:03:14different way of building agents because
3:03:15it's all just based on a prompt and
3:03:17providing the right tools and all the
3:03:18magic kind of happens itself through the
3:03:20prompt. Whereas voice flow gives you a
3:03:22lot more control. So it's really
3:03:23difference between structured AI agent
3:03:25building versus more conversational and
3:03:27open-ended chats through more chat GBT
3:03:30like experience that can just go on and
3:03:31on and on which is what these agents can
3:03:33do. So the purpose is of course fairly
3:03:35similar but the value of this is
3:03:36slightly different in that we are using
3:03:38WhatsApp. So uh many people browsing for
3:03:40services online are hesitant to use
3:03:42website contact forms or other or chat
3:03:44bots that they think are not going to
3:03:45give them access to a real human due to
3:03:47the potential delays that come from it.
3:03:48Right? You land in a website and you're
3:03:50you're shopping around for a different
3:03:51service or product and then there's this
3:03:53this contact form or there's a a web
3:03:55chat widget and you're going to go well
3:03:56I don't really think I'm going to get
3:03:57the help that I really need here at the
3:03:58at the speed that I want. So you might
3:04:00look for a WhatsApp widget and you know
3:04:02that if you click that WhatsApp widget
3:04:03you're going to get to speak directly to
3:04:04someone and this is kind of playing on
3:04:06that fact that if you have the WhatsApp
3:04:08option on your uh website people are
3:04:10much more likely to just click that and
3:04:11go through and try to have a
3:04:12conversation directly to get what they
3:04:14want. So by having a WhatsApp option on
3:04:16a website or other triggers eg you can
3:04:18have a QR code that you could stick on
3:04:20say a real estate sign and you build an
3:04:21agent connect it to your WhatsApp number
3:04:23like we're going to do here and then you
3:04:24create a QR code that people can scan
3:04:26and immediately open WhatsApp and start
3:04:28chatting with it. There's lots of
3:04:29different ways that you can have an
3:04:30access point into a WhatsApp agent like
3:04:32this. But it basically opens up more
3:04:34conversations through a more smartphone
3:04:36native platform. So they can hop on
3:04:38their phone and sort of have a chat away
3:04:39to it rather than being on a website on
3:04:40the computer or a little tiny website
3:04:42chatbot on their phone. Um, in order to
3:04:45essentially engage more prospects or
3:04:47more people interested in the business
3:04:48in conversation, quickly provide value
3:04:51through either the knowledge base and
3:04:52these tools here and real-time quotes.
3:04:54And ultimately, because you're providing
3:04:56that instant value and instant feedback
3:04:57from them, collect their lead
3:04:59information, or better yet, even set
3:05:01appointments through WhatsApp, which you
3:05:02can build on agenda. But that use case
3:05:04is a little bit more advanced and not
3:05:05something I can show within this video,
3:05:07but it definitely is possible. But it's
3:05:08really only a few steps away from the
3:05:10skills that you've learned in this video
3:05:11so far. So keep an eye on that
3:05:12appointment setting use case because if
3:05:14you can do that with AI agent, it's a
3:05:16very, very valuable one. And I've done
3:05:17other videos on the channel here showing
3:05:19you how to do that. So, the usage of
3:05:20this is that they're going to find the
3:05:22uh company's WhatsApp number on their
3:05:24website or elsewhere, maybe a QR code
3:05:26like I said, and they're going to start
3:05:27a conversation on WhatsApp, and then the
3:05:29agent is immediately going to jump in
3:05:30and start responding and be able to
3:05:31answer from the knowledge base, generate
3:05:33quotes, and then capture their lead
3:05:34information. So, without further ado,
3:05:36let's jump into building this agent. So,
3:05:37we can click up here to go to my
3:05:39website, Agentive. You can click on
3:05:41register now. You can just register with
3:05:42your Google account. I'm going to log in
3:05:44with mine.
3:05:48It is free to make an account and we
3:05:50have a free plan so you can just
3:05:51experiment around as much as you need
3:05:52and then you're only going to be charged
3:05:53based on the amount of usage you use. So
3:05:55it's very very cheap and affordable and
3:05:57I wanted to make this platform for you
3:05:58guys to all get on and experiment with
3:06:00building AI agents without coding. That
3:06:02was really the the core of why we
3:06:03started this whole thing. So we've got
3:06:05the dashboard here which will load in my
3:06:06data in a
3:06:07second. So you can see here what the
3:06:09dashboard will look like when you've got
3:06:10your own agents running. We are running
3:06:12the Agentive customer support chatbot
3:06:14through this uh through this account
3:06:16that I'm I'm using right now. So you can
3:06:18see usage costs very cheap sessions etc.
3:06:20So it's really cool when you go into
3:06:21analytics and you can use agentive to
3:06:23see how people are using your agents but
3:06:25that's obviously something for a little
3:06:26bit later once you put these into
3:06:27production. Uh now what we're going to
3:06:29do is of course we can go to agents or
3:06:30we can just create an agent from here
3:06:33and I call this um Connor's cleaning
3:06:36WhatsApp agent. Oh we got a little
3:06:39description as well. um answers
3:06:41questions from KB provides real
3:06:48time so answers questions from the
3:06:50knowledge base provides real-time
3:06:52cleaning quotes and can capture leads to
3:06:54air table. So you're going to see the
3:06:55setup is a lot faster than some of these
3:06:57other platforms. Again, I'm not not
3:06:59trying to gas myself up here. It's just
3:07:01a different way of approaching u
3:07:02building agents. So it's a lot more fast
3:07:04and and rapid prototyping and easy to
3:07:06get things up and running. Of course, if
3:07:08you need much more advanced
3:07:08functionality, you do need to go the
3:07:10extra mile and go on to platforms like
3:07:12Voice Flow. But in this case, we have a
3:07:13prompt, very easy. We have a knowledge
3:07:15and we have tools. So, remember when we
3:07:16went back to being a a chef and the
3:07:18three ingredients concept, these are
3:07:20your three ingredients, right? The
3:07:21prompt that you get to provide as the
3:07:23instructions, the knowledge that you
3:07:24provide as the external knowledge base
3:07:26and the tools that we can connect to it
3:07:28as well. And we can select the model
3:07:29here. So, I think I want a nice and
3:07:31snappy response time because this is
3:07:33going to be on WhatsApp and customerf
3:07:34facing. So I'll go to GPT4 mini. So it's
3:07:38nice and quick. Now we can just put a
3:07:39test test test in there or you are a
3:07:44help. Just put that in there for now as
3:07:46the prompt. The knowledge base we can
3:07:48turn this on. We can create a new
3:07:50knowledge base. Call
3:07:55this. We're going to click here and
3:07:57we're going to upload that same file
3:07:58that we used on voice flow. The same
3:08:00document that will be available in the
3:08:01resources for this uh for this
3:08:03particular guide. Give that a second to
3:08:05process. Once this goes green, we're
3:08:06good to upload it. And you can add
3:08:07multiple files in here. We allowed five
3:08:09files at a time, but you can add dozens
3:08:11and dozens of files. So, you have a
3:08:12massive knowledge base to work
3:08:15with. And just like that, we have
3:08:17connected our knowledge base. And the
3:08:19cool thing about Agent is because we're
3:08:20built on the assistance API, this is
3:08:22actually an independent knowledge base.
3:08:24So, you can create a knowledge base and
3:08:25connect it to multiple different agents.
3:08:27The the knowledge is not restricted to
3:08:28the agent that you build it within. So,
3:08:30I can go and create a new agent and
3:08:31connect this exact same knowledge base.
3:08:32And I have all of these other ones here.
3:08:34And then when we go into the tools
3:08:35section, we are going to have two tools
3:08:36for this. Well, the knowledge base, if
3:08:38we go back to our Figma here,
3:08:40technically the knowledge base is a form
3:08:41of tool that the agent is using. But on
3:08:44platforms like the assistance API and
3:08:45and many platforms, you'll see knowledge
3:08:47treated as its own thing, but
3:08:49essentially it is just another tool that
3:08:50the agent is using at the right time
3:08:52when it needs to pull in knowledge to
3:08:54answer questions. So, OpenAI separates
3:08:56it out into its own thing here. And so
3:08:58do we because we built on top of it. So
3:09:00we do have three different tools but
3:09:02knowledge is its own tool that gets set
3:09:03up through this knowledgebased um
3:09:05connection that we just made before.
3:09:07Then we have the tools and here we have
3:09:09our instant quote from relevance and we
3:09:11have our capture lead information. So we
3:09:12know the process of going on to
3:09:14relevance. So we can just go create a
3:09:15new tool here and this is going to show
3:09:17you a schema. Remember back to when we
3:09:19talked about schemas it explains to the
3:09:21agent how to use the tool. So to add a
3:09:23tool to this agent we need to add a
3:09:25schema to it. And thankfully our buddies
3:09:27at relevance AI provide a very very easy
3:09:29way to create schemas to import into
3:09:31agents like on aentive. So here I can
3:09:33grab that same cost estimate tool for
3:09:35the instant quotation for cleaning
3:09:36services that we've used previously.
3:09:38Again this will be linked. You can just
3:09:39clone this if you haven't got it
3:09:40already. I will provide a link for you
3:09:42to clone this into your relevant account
3:09:43that will be in the resources for this
3:09:45video. And it's just like the previous
3:09:47tutorial that we did where it's got
3:09:48property type square footage and an LLM
3:09:50step here to calculate it. It's going to
3:09:52spit that back and we're going to turn
3:09:54that into a nice response uh with our
3:09:56agent on Agent. We can make sure that
3:09:57we've saved this. So, the cool thing
3:09:58here is that in order to get this
3:10:00connected to uh Agent and our agent over
3:10:02there, we can just go to custom actions
3:10:04here on the tools page. As you can see
3:10:06here, it's mainly intended for use with
3:10:08OpenAI's custom GPTs, which you can get
3:10:10access through chat GPT. And I highly
3:10:12recommend you do check out the OpenAI
3:10:14GPTs because it's a super simple way to
3:10:15spin up your own agents um on the chat
3:10:17GPT site. And so, we can select our tool
3:10:19here and we can get a schema for it. But
3:10:21what I've just realized is that I
3:10:22actually do have an air table lead
3:10:24capture tool here that I've already
3:10:25created on relevance. And it's actually
3:10:26going to be easier for us to set it up
3:10:28here on relevance than to have to do it
3:10:30all separately. So let's just quickly
3:10:31set that up. Now if we go air table,
3:10:35let's just get a simple one that
3:10:36captures the name and phone. I'll
3:10:38provide the template for this tool so
3:10:39you can clone it in. But basically it
3:10:41takes an input of the name of the lead
3:10:42and the email address of the lead and
3:10:44the phone as well. So, it's capturing
3:10:46all three of these as lead information
3:10:48and then it's sending it over to Air
3:10:49Table which we're going to set up just
3:10:51now and it's using a post request to
3:10:53push that data that we collected here
3:10:55and we will collect through WhatsApp
3:10:56eventually and it's pushing it into the
3:10:58Air Table database. So, let's get that
3:11:00set up quickly. If you go to Air
3:11:05Table and I'm just going to use a dummy
3:11:07CRM that I use for all of these
3:11:09tutorials and you guys are going to be
3:11:10able to clone this if you want. In the
3:11:11resources for this video, there will be
3:11:12a link like this. So, if I share this um
3:11:15share
3:11:18publicly, you guys will get something
3:11:20that looks a bit like this and this
3:11:22button up here says copy base. That will
3:11:24copy it into your account. So, all you
3:11:26need to do to copy this air table base
3:11:27is to create an air table account and
3:11:29then click on this copy base and it will
3:11:30copy it over. So, you can get this with
3:11:32the column source preset up. It is
3:11:34fairly easy to set up these fields
3:11:35yourself, but I want to make it easier
3:11:36for you guys. So, you can just copy
3:11:37this. It'll be included in the
3:11:38resources. But here we have the fields
3:11:40that we're looking for. So now all we
3:11:42need to do to send data into this
3:11:43database through our WhatsApp based
3:11:45agentive agent. So when someone provides
3:11:47their details that it gets shot into
3:11:48here is we need to go and see our
3:11:50details for the air tableable web API.
3:11:52So air tableable has their own API which
3:11:54allows us to interact with our databases
3:11:56like this programmatically. So all we
3:11:58need to do is go up to the right hand
3:11:59corner here go to builder hub and if we
3:12:02go to the developer docs here and scroll
3:12:04down to the web
3:12:06API this is a reference for the air
3:12:08table web API. So this documentation is
3:12:10essentially going to tell us how to
3:12:11interact with our air table
3:12:13programmatically through our agents and
3:12:14through voice flow and through relevance
3:12:16and and through agentive as well. So any
3:12:18way you want to interact with it, you
3:12:19can now take this knowledge that you've
3:12:21gained in this video look through this
3:12:22and model what we're going to do here in
3:12:24relevance. You can take that same idea
3:12:25and put it into say voice flow and you
3:12:27can build an air table integration
3:12:28within voice flow yourself where you can
3:12:30send and pull data. So these skills all
3:12:32stack on top of each other and it really
3:12:33centers around understanding how APIs
3:12:35work and that comes down to reading
3:12:37documentation as well. So in this case,
3:12:39if you go back to our relevance tool
3:12:41here, we need to get our URL, which is
3:12:44our endpoint, which we've talked about
3:12:45before. This is the address that we are
3:12:46sending the request to and agentive
3:12:49where we're building the agent. It's
3:12:50going to be using relevance to call air
3:12:52table. It's a bit of a a roundabout way
3:12:53of doing things. But to get all of this
3:12:55information, the easiest way is to go
3:12:57back to this documentation. And the
3:12:59easiest way for us to find the
3:13:00information that allows us to interact
3:13:02with our own Air Table base that we're
3:13:03setting up, is to come down here and
3:13:05find the base that you've just cloned
3:13:07into your account, which will likely be
3:13:08Smith Solar CRM. Don't worry about the
3:13:10name. And Air Table does a really,
3:13:11really good job of making this super
3:13:13easy. And that we can just come here to
3:13:14the leads table. So, if we go back to um
3:13:17Air Table here, you see we are on the
3:13:20leads table. We have these different
3:13:21tabs. You can just ignore these are just
3:13:22different projects that I've done on
3:13:24YouTube. Um they're all kind of there in
3:13:25case people are also cloning this into
3:13:27their account. But we're looking at the
3:13:28leads tab here. So, we go to the leads
3:13:30here and we want to create records. And
3:13:33then it gives us all of the information
3:13:34we need here in order to create records.
3:13:36So, um you can see that it's a post
3:13:38request and it's got
3:13:41HTTPS and all of this information. So,
3:13:44this is the endpoint. We want to copy
3:13:45all of this all the way down to leads.
3:13:47Copy this. Go back to
3:13:50relevance and paste this in. Oh, maybe
3:13:53that was already there. And paste that
3:13:55in there. And then we need to add in two
3:13:57headers. So we have authorization and
3:13:59content type. Now remembering what we
3:14:00learned before, you can see we have H
3:14:02and this H tag means that there's a
3:14:04header. And so the header is
3:14:06authorization. And then the value is
3:14:08going to be bearer and then our token.
3:14:10This is something that tripped me up
3:14:11when I was first learning this uh using
3:14:13APIs. Is that you need to add this
3:14:16bearer word and then a space and then
3:14:18your API key. It's a weird way of doing
3:14:19things. I don't really know why uh why
3:14:21it's like that but sometimes when you're
3:14:23doing these authorizations you need to
3:14:25add in bearer space and then your API
3:14:27key. So it's it's there for a reason um
3:14:29is what I'm saying. And then we have the
3:14:30content type being application JSON. So
3:14:33we're already familiar with that. So
3:14:34going back to relevance we have the
3:14:36header of authorization and content type
3:14:38here. Application JSON. And now we need
3:14:40to add in our Air Table API key so that
3:14:42we are authenticated and we have
3:14:44permission to send an API request. So
3:14:46they're not going to let anyone use this
3:14:47details and and start sending data to
3:14:49our our database, right? They need to be
3:14:51authenticated and that's what API keys
3:14:53do. So to get our Air Table API key, of
3:14:55course, we go to Air Table. We can come
3:14:57up to the top right here, go back to our
3:14:59builder
3:15:00hub, go to personal access
3:15:03tokens, and we can create a new token.
3:15:06Call this
3:15:07YouTube. We can go um add the base. This
3:15:11will be um Smith's Solar CRM. We can add
3:15:15a scope
3:15:18read write and sometimes I find it handy
3:15:20to have the schemas read in there as
3:15:22well. So basically what we're doing here
3:15:23is saying that I give this API key that
3:15:25we're creating permission to interact
3:15:27with this uh this air table and I give
3:15:29it permissions to do these things like
3:15:31read what's in the database write to the
3:15:33databases and create new things and also
3:15:35to see the overall structure of the base
3:15:36and the field types. So we can add that
3:15:39and create the token. We get this token,
3:15:41head back to relevance. And again, this
3:15:43template so that you can clone it into
3:15:45your account is going to be on the uh on
3:15:47the resources. So, if you're following
3:15:48along, you should just clone it into
3:15:49your account and then come down here and
3:15:51make the changes as I do them. So, we
3:15:52can add a make sure we have a space
3:15:54after bearer and then paste our key
3:15:56because if we go back to the web API
3:15:57docs, we can see we have authorization
3:15:59bearer space your API key content type
3:16:02and then application/json. Then we have
3:16:03the data as the payload. Remember, like
3:16:05this is what's inside the envelope. Then
3:16:08we have records and we have the fields
3:16:10name, phone, and status. And then it's
3:16:11provided us an example of how we would
3:16:13send data into that which we don't need
3:16:14to worry too much about because I've
3:16:16already got this fitted in here. It can
3:16:17be quite fiddly. In fact, for this I'm
3:16:19actually going to add another field in
3:16:21here. This is one thing about relevance
3:16:23I'm not a huge fan of. This can feel
3:16:25super fiddly
3:16:27sometimes. So if we go
3:16:30email, this will already be in the
3:16:32template that I give you, by the way.
3:16:34Email. Why would that delete that?
3:16:43Okay, so we have the URL has been
3:16:45updated. The method is post. That's
3:16:47correct. We have the authentication. We
3:16:48have the content type. We have the body
3:16:50all set up. We've added in our fields of
3:16:54name, phone, and email, which we have
3:16:56name, email, and phone. So, we can give
3:16:59it a spin here. If we say
3:17:01Liam, there we go.
3:17:06and we give it a spin. Run the
3:17:09tool. We'll see if it accepts
3:17:18us. Yep. And there we go. If we go back
3:17:21to Air
3:17:22Table, open the base
3:17:24up. There we have it. Liam phone email.
3:17:27So, we can take this tool and we can
3:17:29take the instant quote generator. we can
3:17:31put those into uh Agent and before you
3:17:33know it, we're going to have our agent
3:17:35ready to go. So, let's head back to
3:17:36relevance here. We'll save this tool and
3:17:39we'll change this to name,
3:17:41email, and phone. Um, and just quickly
3:17:44before we do that integration, this is
3:17:46when the description comes into play
3:17:48here. Remember those natural language
3:17:49descriptions of what the tool does, what
3:17:52each of the parameters and inputs are.
3:17:53It's really important to get these right
3:17:55in relevance, and I see a lot of people
3:17:56skipping over this step, but this is
3:17:58what's going to be put into that schema,
3:17:59right? So when relevance generates a
3:18:01schema for us, that onepage manual on
3:18:03how to use this tool and use the API in
3:18:05order to interact with this
3:18:07functionality when we give it into
3:18:09agentive, it's going to be reading over
3:18:11everything in there. And it's going to
3:18:12be those little descriptions around what
3:18:13the tool does and what it's supposed to
3:18:15take in. And and these parts here in
3:18:17relevance is where we get to set that
3:18:18up. A proper description is needed
3:18:20before we do this integration. So this
3:18:22tool captures lead information, stores
3:18:23in Air Table CRM, requires lead's name,
3:18:26phone, and email. Name, phone, and
3:18:27email. The name is name of the lead.
3:18:29Yep. Email, email address of the lead
3:18:31and phone is the phone number of the
3:18:32lead. So that's all good there and ready
3:18:35to
3:18:37integrate. Might even do a quick check
3:18:39on the sparkly cost estimate as
3:18:43well. Yep. Type of property, square
3:18:47footage of the property, and we are
3:18:49ready to go. So now we can click on the
3:18:50custom action step here. Scroll down and
3:18:53click on both of these. Bam. Bam.
3:18:59Scroll down. We're going to change this
3:19:00to custom orth. We're going to generate
3:19:03an API
3:19:05key. There we go. And we're going to
3:19:07generate our open API, not open AI, open
3:19:10API. It's essentially a type of API and
3:19:12a way of describing how the API works.
3:19:14And it gives us all of this information
3:19:16here. I will actually expand it out so
3:19:17you guys can see at least some of it.
3:19:20It's probably easier over on agentive
3:19:21actually. And then we head back to
3:19:23agentive.
3:19:25What we can do is paste in the
3:19:28schema. And now if we scroll through
3:19:30this quickly, I just want you to see
3:19:31what a schema looks like under the hood
3:19:32because we have some important parts uh
3:19:34that's going to really connect the dots
3:19:35for you after everything that we've
3:19:36learned in this video. So I'll zoom in a
3:19:38bit here. Um we have the title of the
3:19:41tool. So we have a few key things in
3:19:42here that we can break down. Basically
3:19:44the paths. We have two paths in here.
3:19:46This is one of them and this is the
3:19:48other. These represent the two tools
3:19:50that we are integrating. You can see one
3:19:52here is the operation ID is basically
3:19:54the name of the tool and that is taken
3:19:56from relevance directly the air table
3:19:57lead capture and the summary here is the
3:19:59actual name or the title of the tool
3:20:00that we had in relevance. This is just a
3:20:02a version where they put in um
3:20:04underscores to connect the uh the gaps
3:20:06and the description here you can see
3:20:07it's the same as a description that we
3:20:09set up over I don't want to go back on
3:20:11there but that was a description that we
3:20:12put under the name to describe what the
3:20:14tool does and then as for the inputs
3:20:16relevance has made it a little bit more
3:20:18complicated by putting a schema in here.
3:20:19Um, so we'll cover that in a second, but
3:20:21basically here's the second tool.
3:20:22Sparkly cost estimate. This tool does
3:20:24this. This about estimating the cost of
3:20:26an apartment. Then down here we have the
3:20:28schemas for the inputs. So we have
3:20:30things like the name. This is for the
3:20:32lead capture tool. The name um this is
3:20:35one of the fields. It's going to be in
3:20:36type string and it's required. We have
3:20:38the email which is type string which is
3:20:40required. And we have the email which is
3:20:42a description
3:20:43here. And of course you can see all the
3:20:45descriptions that we put in on relevance
3:20:47showing up here. then the phone number
3:20:49of the lead, the email of the lead, etc.
3:20:50And here it's specifying how the AI
3:20:52agent should be sending inputs into
3:20:54that. So that's probably the most
3:20:55difficult technical part of this whole
3:20:56video, but I did want to give you a bit
3:20:58of context on how that kind of fits
3:20:59together. This is quite a complex
3:21:01schema. Relevance puts it together in a
3:21:03little bit more complex way um by using
3:21:05these uh these schemas for the inputs
3:21:07down here. But long story short, if we
3:21:09then go to the add or button, we need to
3:21:11set up our authentication, which we can
3:21:14do by coming back to relevance and
3:21:15copying this. We go back to here, paste
3:21:18this in. We go custom orth and we go
3:21:22authorize a
3:21:24orization with a
3:21:27zed. Oh, I need to create the tool.
3:21:30Sorry. So, we can just click create
3:21:32tool. So, the tool has been created
3:21:34successfully. And there we go. We have
3:21:36both of the tools added in because we
3:21:38did them both in one bundle on
3:21:39relevance. And then if we go edit off,
3:21:42we can then put in API keys for both of
3:21:44these so that we have permission to use
3:21:55them. And we can do the same for this
3:22:03one. And there we go. Now we have our
3:22:05knowledge set up, our two tools set up,
3:22:08and you can see that we're pretty darn
3:22:09close to completing this build. We have
3:22:11all of these three done. Might as well
3:22:13make them green for the sake of it. And
3:22:16now the only thing left to do is to
3:22:17write a prompt that connects this all
3:22:19together. And that's really the glue
3:22:20that holds it together. My go-to method
3:22:21of rapidly creating prompts for AI
3:22:23agents is using a relevance tool. Um,
3:22:28perfect that I've created, and I I said
3:22:30I'd give this to you guys for free as
3:22:31well. That's going to be included in the
3:22:33resources. But if I go to use
3:22:35here, it's a prompt writer that includes
3:22:38all of the information from how we do
3:22:40prompting at Morningside, which is based
3:22:41on research and includes all the key
3:22:43things like RO, task, specifics,
3:22:45context, um, explaining how to use the
3:22:46tools that it's been provided as well.
3:22:48So, I'm just going to fill this out
3:22:49quickly here and then get a prompt. And
3:22:50you guys can steal my prompt from the
3:22:52resources or you can use this as well to
3:22:54create your own. But, it's a pretty good
3:22:56exercise because you can see here we
3:22:57have the agent name being
3:23:04Um, and what I like to do here is, so
3:23:06this is just a quick rundown of what the
3:23:07agent does, where it is being deployed,
3:23:09and
3:23:18why. So you can pause the video and look
3:23:20at that there. But just a bit of context
3:23:21on what the agent does, where it is
3:23:23being deployed, and why. Conversions
3:23:25context, we can say
3:23:36knowledge context. We can say with a
3:23:43contain. Then we get to the tools
3:23:45available section and we can
3:23:56say and then we have the other tool
3:23:58which is uh
3:24:07And then for the ideal input and output
3:24:09examples, I'm just going to say none to
3:24:13provide friendly helpful
3:24:18assistant. And so just like that, in
3:24:20maybe a few minutes, I've typed in all
3:24:22of this information about the agent and
3:24:23what it does. And now I can just click
3:24:25run tool here. And it's going to take
3:24:27all of this information, run it through
3:24:29the prompt that I've written that bakes
3:24:31in the best prompting practices for AI
3:24:33agents from my experience and from the
3:24:34projects that we do at Morningside AI
3:24:36and also all the research that we've
3:24:37used to make those prompting practices.
3:24:39And it's going to spit us out an AI
3:24:41agent prompt that we can throw straight
3:24:42into Agent and it'll just glue
3:24:44everything that we've done together,
3:24:45tell it who it is and what it's trying
3:24:46to do, tell it how it's supposed to use
3:24:48the knowledge base, and tell it how and
3:24:50when it's supposed to use those tools in
3:24:51order to reach its objective of
3:24:52capturing those leads for us via
3:24:54WhatsApp. And there we go. So, if we
3:24:56scroll down, we can see it's spit out
3:24:58this entire prompt. I'm going to change
3:24:59it to the raw text so we get all this
3:25:01markdown formatting included. We can
3:25:04view all here. I'm going to copy it all
3:25:07and we're going to take it over to the
3:25:10here and paste this in. And there we go.
3:25:12Act as corner cleaning WhatsApp support
3:25:14and lead generation agent. Engage with
3:25:16potential customers on WhatsApp to
3:25:17provide potential information about our
3:25:19cleaning services. Answer FAQs. answer
3:25:21off instant quotes
3:25:23ra when pricing inquiries arise use the
3:25:26instant quote generator tool tools you
3:25:28have this tool and this tool examples
3:25:30I'm just going to cut that out for now
3:25:33and then notes ra so that should be all
3:25:36good we can start to give this a spin
3:25:37here I am going to zoom out a bit right
3:25:38all right so I'm just going to publish
3:25:39this and make sure that everything is
3:25:41baked in the second and now we can chat
3:25:44to it here um hey how's it
3:25:47going actually slide this across
3:25:51Um, I want to know where you guys are
3:25:55located. There you go. Connor's cleaning
3:25:57is located at XYZ. So, it's obviously
3:25:59using the knowledge base correctly and
3:26:00say, "What what services do you
3:26:03provide?" And we're not asking about
3:26:06pricing. So, it shouldn't go for the
3:26:09quote or it might try to do it at the
3:26:11end. Yes. So, see, it's asking if you
3:26:13have any specific requirements, need a
3:26:14quote, just let me know. Yeah, sure. I'd
3:26:16like like a quote. Boom. I need the
3:26:19property type and square footage. It's a
3:26:21house and it's 1,00 square ft. So now
3:26:24the agent is trying to trigger that tool
3:26:26by taking the house and taking the 1,00
3:26:28and then putting them into the relevance
3:26:30tool based off what the schema has told
3:26:31it how to use the API. It's going to go
3:26:33grab that from relevance, send it back
3:26:35to us and there we go. Here are the
3:26:37quotes for you. Ra, if you're interested
3:26:39in any specific further service and need
3:26:41assistant, just let me know. I can also
3:26:42help you with booking. Now here I would
3:26:44probably change the prompt and make it a
3:26:45bit more forceful and say send me like
3:26:48like let's go to the next step right
3:26:49now. But for now it's good enough. Um we
3:26:51can say uh sure I'd like to book a deep
3:26:54clean. Now it should ask me for my lead
3:26:58information. Okay. Huge
3:27:01Jackman is the
3:27:03name is the phone and huge Jack Jackman
3:27:08is email.
3:27:10We should be able to see if we go back
3:27:12to our handy dandy air table
3:27:14here. Huge popping up here.
3:27:26Oh, bang. And huge Jackman is in the CRM
3:27:31here. It does say that it's booked. I
3:27:33would play around with the prompt a
3:27:34little bit more to be like, hey, look,
3:27:35this is just setting up the next step
3:27:37for someone to call them and book in the
3:27:38service. But you can also do appointment
3:27:40setting through agentive as well. Again,
3:27:41like I said, it's a little bit more
3:27:42advanced than what we want to do here.
3:27:44But as you can see, this is a very
3:27:46different way of approaching building
3:27:48agents because you tell it, you
3:27:50basically provide all of the ingredients
3:27:51and you use that kind of chef's
3:27:52approach. The knowledge and the tools
3:27:54and you connect it all up and you make
3:27:55sure the tools have well described
3:27:56schemas so they know how to use it and
3:27:58they know when to trigger them. The
3:27:59knowledge base has been included in the
3:28:01prompt and also the tools as well have
3:28:02been included in the prompt um telling
3:28:04it how and when to use it. It's really a
3:28:06much faster way of building agents from
3:28:07the highle prompting and then people are
3:28:10just asking and having sort of a free
3:28:11flowing conversation with it. Okay. And
3:28:12just quickly before we go to the step of
3:28:14putting it onto WhatsApp which won't
3:28:15take long at all. I do want to show you
3:28:17how you can debug and when you're
3:28:18working in agentive um it's helpful to
3:28:20know when tools are being triggered and
3:28:22why. So for example, if we go into the
3:28:24transcripts here and we look at this big
3:28:26transcript here with 14 messages that we
3:28:27just had. Hey, how's it going? Ra. We
3:28:31can see here it's using the tools and we
3:28:33can hover over it and we can see it's
3:28:34calling the tool with the URL. It's a
3:28:36post method and we can see the data
3:28:38here. I'll just zoom in on that. The
3:28:40property type and the square footage
3:28:42that was sent away to relevance are
3:28:44here. So if you're having issues with
3:28:45your tools or it's giving weird
3:28:46responses, um you can either come in
3:28:48here to the transcripts after the fact.
3:28:49So say maybe this is on WhatsApp um and
3:28:52something's going wrong or customers are
3:28:54getting upset. You can come into the
3:28:56transcripts here and pick through and
3:28:57see what's going wrong with the tools.
3:28:58And just like down here as well, the
3:29:00lead capture, we can see the name,
3:29:02email, and phone were all put into this
3:29:03request and sent away to relevance AI.
3:29:05And then onto Air Table as a second
3:29:07step. Um, and you can also see the
3:29:08output as well. So the output of the
3:29:10tool is all in here. It's basically just
3:29:13giving us a confirmation back from Air
3:29:14Table that, yep, everything went well.
3:29:16And up here, you can see the
3:29:19output as uh the response with the deep
3:29:22cleaning estimates and stuff like that.
3:29:23You can see it a lot more easily if we
3:29:25say, "Hey, can I get a quote?
3:29:30Okay. And if you give this a second once
3:29:32it's finished generating agentive will
3:29:34then pop up this show usage and bang
3:29:36there in the editor here. You can then
3:29:37debug. Okay. How many tokens are being
3:29:39used? How much is this costing? What was
3:29:40the model? Um etc. And then you can see
3:29:43the tools input here. Apartment 500 ft
3:29:46etc. And the output as well. So it's
3:29:47really easy to debug those tools while
3:29:49you're in Agent. Let's make sure that
3:29:50we've published
3:29:51this. I'm going to publish it again. In
3:29:54Agent, we do have version history. So,
3:29:55if you do publish it and you want to
3:29:57roll back or look at how you had it set
3:29:58up previously, you can now see that I've
3:30:00got two versions, V1 here, and I just
3:30:02took away this little full stop here and
3:30:04you can see that that's I've changed the
3:30:06prompt. So you can update it over time.
3:30:07You can make edits within Agentive here
3:30:09and test test test. And then when you're
3:30:11ready to push that to production and
3:30:12basically if we had this on a WhatsApp
3:30:14agent and say I published this,
3:30:16connected to WhatsApp and it was working
3:30:18and then I looked through the the
3:30:19transcripts and something wasn't quite
3:30:20how I liked it, I could come in here and
3:30:22make edits and then test test and then
3:30:24when I was ready to publish it, I click
3:30:26publish and then it's going to push
3:30:27those live to the agent. So you're not
3:30:28going to mess things up by playing
3:30:30around with things on here. So the final
3:30:31step is of course to deploy it to
3:30:32WhatsApp.
3:30:36Go to the deploy tab here. You can then
3:30:38click connect
3:30:41WhatsApp. I'm going to click
3:30:43continue, get started. You will of
3:30:45course need a Facebook business manager
3:30:47to set up this integration fully. That's
3:30:48free with every Facebook account. So, if
3:30:50you haven't got one already, I'll leave
3:30:51a link in the description so that you
3:30:52can set it up. Takes a few clicks. Then
3:30:54you will see this page here. And you can
3:30:55select the business manager you've
3:30:56created. In this case, I'll be using
3:30:58this testing one. And then you'll be
3:30:59able to set up a new WhatsApp business
3:31:00account, which I can click here. I'll go
3:31:02next. set up a business account
3:31:05name. And then this is the display name
3:31:08for the business. And we're going to
3:31:10call this a a retail business. Now, you
3:31:14need to provide the phone number that
3:31:15you want to connect your agent to. Um,
3:31:16unfortunately, you can't have your own
3:31:18personal WhatsApp number and also have a
3:31:20business account running through it. So,
3:31:21you need to either buy another SIM card
3:31:23or borrow a friend's number who doesn't
3:31:24have WhatsApp, etc. In this case, I'll
3:31:26be using a spare number that I have.
3:31:28Then, they're going to send you a
3:31:29verification code to your number, which
3:31:30you have to enter in. And then you
3:31:32should see the screen once you've
3:31:33successfully passed that verification.
3:31:35So when we continue, it's verifying our
3:31:37information for a second. And now our
3:31:39agent is connected to that phone number
3:31:40and we're ready to give it a test. So if
3:31:42we go finish here, there's one more
3:31:44thing that we need to do on Agentive,
3:31:45which is to click this. Yours may say
3:31:47not registered. Don't worry, you can
3:31:48just click this check box here and click
3:31:50confirm. Give it a second to connect.
3:31:53Now we've successfully connected our
3:31:54agent to that WhatsApp number. Now thing
3:31:56here is this interval. If you're not
3:31:58sure what the interval is, you can read
3:31:59this tool tip here. And if you're done
3:32:01with the deployment and you want to
3:32:02remove it from that number, you can
3:32:03always come back and click deactivate
3:32:04deployment here. But all that's left to
3:32:06do now is to test our functionality.
3:32:07Right. So I have it connected to my
3:32:09phone here. So I'm just going to show
3:32:10you a little bit of a on screen here of
3:32:12me creating this contact and having a
3:32:13message with it. So the number that I
3:32:15set up, I can create a new contact and
3:32:17then I can send them a message.
3:32:22Hey, and you can see on screen here it
3:32:24says this is the business using a secure
3:32:26service from Meta. So, this means this
3:32:27is a business account um as we've
3:32:29connected it through our WhatsApp uh
3:32:30business profile that we set up
3:32:33before. And there we go. We get a
3:32:35message back. Hello, thank you for
3:32:36sharing your information. How can I
3:32:37assist you today? Um if you have any
3:32:39questions about cleaning service or need
3:32:40a quote, feel free to let me know. So, I
3:32:42can say um yes, a quote. Let's ask a
3:32:45question to the knowledge base. Where
3:32:47are you
3:32:51based? There we go. We are based in the
3:32:53greater Boston area. It's giving me the
3:32:55uh the correct location there. So, we
3:32:57can go for the lead capture now. So, if
3:32:58we say um I'd like a
3:33:06quote. So, property type uh it's a house
3:33:10that is
3:33:12about 800 square
3:33:18ft. There we go. We're getting the uh
3:33:21estimations and our quote back. Um, it's
3:33:23asking if we're interested in any of
3:33:24these services. I'd say yes, I'd like
3:33:27the deep cleaning,
3:33:31please. Now, it should ask me for my
3:33:33contact
3:33:35information. There we go. So, Liam, I
3:33:38mean, Liam at mail.
3:33:41Cool. And then we should see it appear
3:33:43over here on our Smith Solar
3:33:46CRM. And boom, there it is. So, we've
3:33:49got everything done. That is just one
3:33:50run through of using this WhatsApp
3:33:52agent. But as you can see, uh the the
3:33:54messages don't come back instantly. So
3:33:56it it feels like it is like it's
3:33:58actually could be a real human there
3:33:59applying and it's giving just clear
3:34:01information right through WhatsApp.
3:34:03Imagine you are reaching out to maybe
3:34:04book an accommodation or you're reaching
3:34:06out to a a cleaning service like this or
3:34:08you're reaching out to any kind of
3:34:09business and you want some real
3:34:10information directly from what feels
3:34:11like a person. And then you also have
3:34:13the functions of getting a real quote of
3:34:15I mean a great use case for this kind of
3:34:17thing is like barbers. Like I say, if
3:34:18you you message a barber on WhatsApp,
3:34:20maybe you're in in Europe somewhere,
3:34:22you're in South America or you're in
3:34:23Central America or and and you want to
3:34:25go to a barber and this is a common
3:34:27issue that I've run into when I'm
3:34:28traveling. It's like, I want to message
3:34:29this barber, but I might not speak the
3:34:32language that well. And then if you
3:34:33message them in English, it will be able
3:34:34to handle that in in English as well as
3:34:36in Spanish or in Portuguese or wherever
3:34:38you are. So, this kind of functionality
3:34:39built through WhatsApp is a really
3:34:40really great use case for you guys to
3:34:41pick up, which is why I wanted to teach
3:34:43you guys it. And we can also go back to
3:34:45agentive here. And if we go to our
3:34:47transcripts for this agent, we can see
3:34:50the one for today is here. So 12
3:34:51messages. You can go through the entire
3:34:53transcript and you can see it's calling
3:34:54the quote tool here. We see all the
3:34:56information that went in and out of it.
3:34:57And then we see the air table lead
3:34:59capture information as well. Input and
3:35:00output. So that is how you use a genty
3:35:02my software for building these WhatsApp
3:35:04based and also other deployments as
3:35:05well. So if we go to studio and we go to
3:35:07deploy, we have Instagram as well. So
3:35:09via our mini chat template. You can hook
3:35:10into Instagram and do appointment
3:35:12settings and things on Instagram. You
3:35:13can go through Messenger if you want to
3:35:14run some Facebook lead ads to Messenger
3:35:17through voice flow as well. Telegram,
3:35:18Discord, we have integrations with
3:35:20everything you need as well. So that's
3:35:22the end of this build. I hope you
3:35:23enjoyed and uh this is a super handy use
3:35:25case um and and deployment really for
3:35:27your AI
3:35:32agents. So now that you understand how
3:35:34AI agents works and can build them for
3:35:36yourself, let's talk about the most
3:35:38important part of this, which is
3:35:39actually making money with these skills.
3:35:41But first, we need to destroy a huge
3:35:43misconception and that you don't need to
3:35:45build the next chat GBT or create some
3:35:47revolutionary AI startup in order to
3:35:49make money in the AI space. The real
3:35:51opportunity is much much simpler. It's
3:35:53just helping businesses to understand
3:35:55and implement AI. This is how I
3:35:57monetized my AI agent skills and it has
3:35:59been the most explosive growth I've ever
3:36:01experienced in my career. And the good
3:36:03news is, if you've made it this far in
3:36:04the video, you are so much closer to
3:36:06being able to tap into this starving
3:36:08market for AI services than you think.
3:36:10But don't take my word for it. I'm just
3:36:11some guy on the internet after all.
3:36:13Maybe you should listen to some of the
3:36:14world's most famous businessmen saying
3:36:15that this is the opportunity to get into
3:36:17right now. If I was 25 years old today
3:36:19in 2024, what would I do? What's a good
3:36:22sector to get involved in? What business
3:36:24would I get involved in? I think
3:36:26everything is looking at AI now in a
3:36:28different way. And I think AI growth is
3:36:30going to be exponential. So, anything to
3:36:32do with AI now, what could that be? In
3:36:34the simplest form is helping people use
3:36:36the technology. there's going to be a
3:36:38massive amount of people wanting to use
3:36:40it that don't know how to and they're
3:36:41willing to pay to solve that pain point.
3:36:44So, is that consulting? Not really. It's
3:36:47implementation and execution. And so,
3:36:50helping a business do that transfer into
3:36:52a world where they're controlling their
3:36:54data and getting information from it.
3:36:56Now, the majority of businesses in
3:36:58America, for example, are between 5 and
3:37:00500 employees. So, they're small
3:37:02businesses. They create 62% of the jobs.
3:37:05They want to use AI. you should help
3:37:07them solve for that and they'll pay you.
3:37:09Even another shark, Mark Cuban, is
3:37:11saying the exact same thing that the
3:37:13biggest opportunity right now is helping
3:37:14these small to mediumsiz businesses who
3:37:16don't understand AI yet, but desperately
3:37:19need it to keep up. And they're
3:37:20absolutely right. If we look at the
3:37:22data, it's pretty obvious. According to
3:37:24recent data, there's 1.7 million
3:37:26businesses in the US alone that are
3:37:27making between $500,000 and $10 million
3:37:30per year. These are small businesses,
3:37:32which, as Kevin Oer says, make up 62% of
3:37:35the jobs in the USA. They create 62% of
3:37:37the jobs. They want to use AI. You
3:37:40should help them solve for that and
3:37:41they'll pay you. These businesses know
3:37:43they need AI to stay competitive, but
3:37:45they don't have the time to learn it
3:37:46themselves. And there's basically no one
3:37:48there to help them. All of the big
3:37:50consulting firms are looking at other
3:37:52big businesses and just leaving these
3:37:53smaller businesses completely ignored,
3:37:55but they still make lots of money and
3:37:57they still have a lot of money to invest
3:37:58in these kinds of services. Basically,
3:38:00all small businesses are starving for
3:38:01some kind of AI services, either
3:38:03education services to help them
3:38:05understand what AI is in the first place
3:38:06and why they might need it. There's the
3:38:08huge need for consulting services where
3:38:10you help them to identify where AI can
3:38:12help with them most in their particular
3:38:13business. And of course, there's
3:38:15implementation services where you help
3:38:16them to build and maintain the AI
3:38:18systems like the AI agents we've just
3:38:20built. And right now, based on the data
3:38:21collected in my community, and we are
3:38:23the largest AI business community on the
3:38:25planet right now, for every person or
3:38:26agency that is currently offering AI
3:38:28services, there are over 1,100
3:38:30businesses in the USA alone that need
3:38:32help. So, that's a 1 to 1,100 ratio,
3:38:35which means this is a completely
3:38:37untapped market. And that's where people
3:38:38like you and I come in, helping these
3:38:40hardworking small business owners to
3:38:41understand AI and implement it so that
3:38:44they have a chance to keep up. And
3:38:45that's really what drives me and the
3:38:46team at Morningside because our company
3:38:48vision is a world where the benefits of
3:38:50generative AI are distributed as fairly
3:38:52as possible and they make it to people
3:38:53like me and you and the small business
3:38:56owners rather than just all going to
3:38:57these giants at the top. And this whole
3:38:58concept of selling services around an
3:39:00emerging technology is nothing new. And
3:39:02we saw the exact same pattern when the
3:39:04internet first came out. Companies that
3:39:05helped businesses to adapt to the web
3:39:07and sort of get online made fortunes.
3:39:09You know, agency.com, Razerfish, etc.
3:39:11And I spotted this opportunity within
3:39:13the AI space in 2023 when it wasn't
3:39:15anywhere near as clear as it was now. No
3:39:17one really knew how to make money out of
3:39:18this stuff. And then I started Morning
3:39:20Side AI. And since then, we've generated
3:39:22over $5 million in selling these kinds
3:39:24of AI services of education, consulting,
3:39:26and implementation. And we're literally
3:39:28still only just getting started. And the
3:39:30best part out of all of this is that as
3:39:31we've proved in this video already, you
3:39:33don't need to be a technical genius to
3:39:35understand AI and even to build AI
3:39:37agents. You just need to be one step
3:39:39ahead of the businesses that you're
3:39:41going to be helping. So, let me show you
3:39:42the three specific ways that you can
3:39:44start making money with your AI agent
3:39:47skills. So, as I said, there's basically
3:39:49three types of services that you can
3:39:51provide to businesses in order to
3:39:52monetize your skills. Firstly, there's
3:39:54education, and this is teaching
3:39:56businesses about AI, running workshops,
3:39:58and doing presentations, training their
3:40:00teams, and creating courses for them to
3:40:01watch. Businesses are desperate for
3:40:03someone who can explain what this stuff
3:40:05is in simple terms and more importantly
3:40:07what it can do for them. After watching
3:40:08this video and probably my other huge
3:40:11video that I did on AI agents, which
3:40:12I'll link down below, um you will know
3:40:14more than enough to start educating
3:40:16businesses on AI and AI agents. Secondly
3:40:18is consulting. And this is where you
3:40:20analyze a business's operations and you
3:40:21show them where AI can help them save
3:40:23time or make more money. You're
3:40:25essentially being their AI strategist.
3:40:27For example, you could go into a
3:40:29business and then recommend something
3:40:30like the sales co-pilot system that we
3:40:32just made in order to help their
3:40:34struggling sales department. And third
3:40:35is implementation. So this is where you
3:40:37actually build and deploy AI solutions
3:40:39for businesses. Or better yet, like my
3:40:41agency, you can do all three of these,
3:40:43but it did take us 2 years to get here.
3:40:45So there is really no rush. You just
3:40:46pick where you want to enter and work
3:40:48your way up to doing more and more if it
3:40:49makes sense. Believe it or not, there
3:40:51are people with only a few months
3:40:52experience in the AI space selling all
3:40:54of these right now. And the demand from
3:40:56businesses is increasing insanely fast
3:40:58right now. I we're seeing this at
3:40:59Morningside. Just so many more
3:41:00businesses reaching out. But here's the
3:41:02thing. You have one small problem, and
3:41:04that's that you don't quite know enough
3:41:06to start moving on this. You are close,
3:41:08but you're not quite there. The way to
3:41:09make money in the AI space or with any
3:41:11services really is to create a knowledge
3:41:13gap between yourself and the people that
3:41:15you're helping. Your knowledge gap is
3:41:17your money maker, and businesses will
3:41:18pay you in proportion to how much more
3:41:20you know about AI agents and their
3:41:22business applications than they do. Now,
3:41:24while this video has taught you a lot,
3:41:26your knowledge gap is still small. But
3:41:28we can fix
3:41:30that. So, let me break down exactly what
3:41:33you need to do next in order to extend
3:41:34your knowledge gap to the point where
3:41:35you can start making money. We can call
3:41:37this video as step one. So, as long as
3:41:39you've taken notes and followed all the
3:41:40tutorials and built the agents alongside
3:41:42me, you're already ahead of most people
3:41:44who have no idea about what agents are,
3:41:46how they work, or how to build them. So,
3:41:47it's a big step forward with this video.
3:41:49But step two is building even more
3:41:52experience building AI agents. So you
3:41:54are more familiar with the platforms and
3:41:55better understand the different ways
3:41:57that they can be used to deliver
3:41:58different kinds of AI agent use cases or
3:42:00even just AI tools in general. I've only
3:42:03really given you a taster here, but I
3:42:04tried to make it as as diverse as
3:42:06possible as you could probably tell. In
3:42:07order to do this second step of
3:42:09extending your knowledge gap further and
3:42:10building more experience, you can go to
3:42:12my free course on school where you'll be
3:42:14able to build another 5 to 10 agents
3:42:15following the tutorials that are in
3:42:17there for you. So the link to join my
3:42:18free school will be in the description.
3:42:20So, if you blast through all those
3:42:21tutorials in there, this is going to
3:42:22further expand your knowledge gap. And
3:42:24remember that the more that you know
3:42:26compared to the businesses that you're
3:42:27trying to help, the more they're going
3:42:28to pay you. So, step two is building a
3:42:30few more agents out and getting a bit
3:42:32more experience on the tools, seeing
3:42:33different use cases, etc. And once
3:42:35you've done that, you'll have what I
3:42:37call foundational knowledge. So, you
3:42:38understand the core AI concepts that
3:42:40we've been through in this video. You
3:42:42can build basic solutions on these
3:42:43platforms. You know what's possible for
3:42:45businesses with agents right now. And
3:42:46then comes the big decision. Do you want
3:42:49to go deeper technically on this
3:42:50building side of getting your hands
3:42:52dirty or do you want to start monetizing
3:42:54what you already know? As we've covered,
3:42:56the building and implementing of the AI
3:42:58systems is only one of the services that
3:42:59you can sell. So naturally, the
3:43:02technical skills needed in order to make
3:43:03money in the implementation services.
3:43:06Actually building these systems and
3:43:07businesses is much more greater than
3:43:09just having a foundation. But with a
3:43:10good foundation, you're basically ready
3:43:12to start having a crack at the other two
3:43:14services of AI education and AI
3:43:16consulting. So, the decision of what to
3:43:18do next comes down to really knowing who
3:43:19you are and what you are really
3:43:21interested in. And this sounds all woo
3:43:22woo and like, oh, you got to know
3:43:24yourself and stuff, but this is I mean
3:43:27it very very seriously in that if I use
3:43:29myself an example, I've always loved
3:43:31making things, right? I used to build
3:43:32block houses with kids. I used to like
3:43:34brew beer with my grandpa. I've always
3:43:36loved tinkering with engines. So when I
3:43:37hit this foundational level that you
3:43:39guys will be at after completing those
3:43:40extra builds in the free course, I kind
3:43:42of naturally just dove deeper into the
3:43:44technical side into building more stuff.
3:43:45I I kept building more and more complex
3:43:47AI systems and building upon those
3:43:49skills that I've I've built already,
3:43:51which led me ultimately to starting
3:43:52Morningside AI where our first service
3:43:54was building AI solutions and systems
3:43:56for clients. But here's the thing, of
3:43:58course, a lot of people aren't like me.
3:43:59They don't get as much of a buzz out of
3:44:01building things. many of you are going
3:44:02to be much better or enjoy more the
3:44:05teaching aspect or working with people
3:44:06and doing the consulting aspect rather
3:44:08than building stuff. So in these cases
3:44:10using the foundational knowledge that
3:44:12you're going to build up to sell AI
3:44:13education to businesses or AI consulting
3:44:16makes a lot more sense. Goes back to the
3:44:17whole Einstein thing about like judging
3:44:19a fish on its ability to climb a tree.
3:44:21If to you the building is like a tree
3:44:22and you feel like a fish and it's not a
3:44:23really good fit, then there's better
3:44:25stuff that you can do and you can find a
3:44:26way to make money in the AI space that
3:44:28leans more into your strengths. like in
3:44:30the case of a fish would be swimming,
3:44:32right? So, by being honest with yourself
3:44:33and saying, "Hey, look, that's not
3:44:34really me. Yeah, sure, I did get it
3:44:36done. I know how that works now, but I
3:44:38don't feel any kind of attraction to
3:44:39doing more of that." While you may see
3:44:41that as a negative and saying like, "Oh,
3:44:42I don't have what it takes." It's
3:44:43actually can be very empowering if you
3:44:45say, "Bang, I'm stopping it here. I'm
3:44:47stopping the learning. I'm stopping the
3:44:48procrastination. Now, I'm going straight
3:44:50into actually monetizing." It's
3:44:51basically putting a stop on when you do
3:44:53this learning big long phase and saying,
3:44:55"No, action starts now. I'm not I'm
3:44:57never going to get there, but with this
3:44:59base, I can do a lot and I'm going to
3:45:00start taking action with it and making
3:45:02money with it today. So, this
3:45:03self-reflection is really what prevents
3:45:04you from getting stuck in an endless
3:45:06learning phase of procrastination when
3:45:08you could be out there making money. So,
3:45:09in summary, the two routes you have and
3:45:11the two options you have from here are
3:45:13if you love building and you kind of
3:45:15naturally feel like you want to learn
3:45:17more like like myself when I was at your
3:45:19stage, then just keep going. go and
3:45:21watch the free course tutorials on my
3:45:22school and then start going and building
3:45:24your own projects and ones for friends
3:45:25and family and whatever you you sort of
3:45:27pulled towards naturally and within two
3:45:29to three months you'll have enough
3:45:30skills and experience to actually start
3:45:32selling implementation properly. But on
3:45:34the other hand, if you haven't fallen in
3:45:35love with the building aspect, then it's
3:45:37probably best that you just go go into
3:45:39the free course, smash out the rest of
3:45:40those tutorials and finish your
3:45:42foundation and then just get started on
3:45:44monetizing your skills either through
3:45:45selling AI education or through selling
3:45:47AI consulting.
3:45:50So once you're clear on what kind of AI
3:45:52services you want to sell, getting your
3:45:54first few clients is actually pretty
3:45:55straightforward. People try to over
3:45:56complicate it, but there's really just
3:45:58two main ways that I'd recommend you do
3:45:59this based off all the success I've seen
3:46:00over thousand thousands of people across
3:46:02my free and paid communities. The first
3:46:04and by far the easiest method is through
3:46:05your warm connections or warm contacts.
3:46:07This means reaching out to people that
3:46:08you already have some kind of
3:46:09relationship with, whether it's friends
3:46:11or family or kind of acquaintances or
3:46:13even friends of friends that you've met
3:46:14once kind of thing. All of these people
3:46:16count as warm connections. So instead of
3:46:17trying to convince complete strangers to
3:46:19trust you with your business, you can
3:46:20start with people that you already know
3:46:22or have some previous relationship with
3:46:23and therefore have an increased level of
3:46:25trust with you through your
3:46:26relationship. And I've covered this
3:46:28many, many times on the channel before.
3:46:30So on the school post for this video, I
3:46:31will add in my complete guides for warm
3:46:33outreach, including resources directly
3:46:35from my AAA accelerator program. The
3:46:36second way is using what I call the
3:46:38community content flywheel. So this is
3:46:40how you can build long-term momentum
3:46:42beyond just warm outreach. So here's how
3:46:43it works. You join the free school
3:46:46community. you start making content
3:46:47about what you're learning at each
3:46:48stage. This could be through YouTube
3:46:49tutorials, which I mean that worked for
3:46:51me. LinkedIn post is another one um or
3:46:53whatever platform you really prefer to
3:46:55create content on. But here's the key.
3:46:57You share that content back into the
3:46:59community. So with over 120,000 members,
3:47:03by posting it into the community, you
3:47:05get an instant audience and people who
3:47:06are really interested in the stuff that
3:47:08you're talking about. So, a perfect
3:47:09example of this is a guy called Rory
3:47:11Ridges, a a young guy from the UK who
3:47:13joined my free community and basically
3:47:14followed this exact process that I've
3:47:16told you in this video so far. So, he
3:47:17took my free course, he learned all the
3:47:19basics, built his foundation, then he
3:47:21started posting simple tutorials on
3:47:23relevance AI, which you've used in this
3:47:25video already, and he literally just
3:47:26started sharing what he'd learned from
3:47:28my videos and making other videos about
3:47:29it. And at the start, he was literally
3:47:31just sharing what he'd learned from my
3:47:32videos. So, he'd watch a video, then go
3:47:34and kind of make his own video on the
3:47:35same sort of topic. And every time he
3:47:37made a tutorial, he'd then share it into
3:47:39the community and the community would
3:47:41watch it. They'd give him feedback, go
3:47:42and subscribe to his channel. This not
3:47:44only helped him grow faster on YouTube,
3:47:45but it also started to position himself
3:47:47as an expert in the community. And he's
3:47:49also building his authority in the AI
3:47:50agency space by getting more momentum on
3:47:52YouTube. Now, his YouTube channel brings
3:47:54him in enough leads to support his
3:47:55growing agency. And I've just seen him
3:47:57recently in the community saying he's
3:47:58hiring. So, that's usually a bloody good
3:48:00sign that he's making some good money
3:48:01off the back of it. He basically started
3:48:02the same flywheel that took me from zero
3:48:04to where I am now. Over $5 million in
3:48:07revenue generated across all my
3:48:08businesses and $450,000 plus subscribers
3:48:11in just two years. And so the community
3:48:13gives you an audience and the content
3:48:15gives you credibility and together this
3:48:17method brings you clients. In the
3:48:18resources for this video on school, I
3:48:20will leave links to my complete guide
3:48:22for creating content to generate leads
3:48:23just like Rory and I have done
3:48:25successfully. And of course, I'll
3:48:26include a link to Rory's channel in the
3:48:28resources in the school community. Now,
3:48:29the really important thing to notice
3:48:30with both of these methods we've just
3:48:32talked about, the warm outreach and the
3:48:34uh community content flywheel is that
3:48:36both of these methods start with giving
3:48:37value first. Whether it's helping your
3:48:40warm connections to understand AI or
3:48:41sharing your knowledge through content,
3:48:43you have to give before you start to
3:48:47get. Now, I know all of this businessy
3:48:49stuff may feel a little bit overwhelming
3:48:50or out of reach for some of you, but
3:48:52you'll seriously be amazed at what baby
3:48:55steps add up to in the AI space. you've
3:48:57already taken the first step by watching
3:48:58this and following through to the end of
3:49:00the video. So, congratulations on that
3:49:02and I seriously seriously want to give
3:49:04you a pat on the back and you should
3:49:05give one to yourself. But all you need
3:49:06to do from now is to keep this momentum
3:49:08going. The next step for all of you is
3:49:11pretty damn clear. You need to jump in
3:49:13the free community. That's by far the
3:49:14best place if you're moderately
3:49:16interested in doing any of this stuff.
3:49:17My community is the number one place to
3:49:18go. It is 100% free to get into. And
3:49:20once you're in there, drop an
3:49:21introduction post saying who you are,
3:49:23what you're about, why you want to do
3:49:24this. Then start working your way
3:49:25through my free course material. I've
3:49:27poured everything I've learned about AI
3:49:29and AI business into videos like these.
3:49:31And on the free course, they're all
3:49:32there in a nice sequence for you to work
3:49:33through on school. And each time you
3:49:35complete a video and you complete the
3:49:37tutorial, you can click the little check
3:49:38box so that you can keep stacking those
3:49:40small wins and those baby steps towards
3:49:41AI literacy and succeeding in the space.
3:49:43All of the resources mentioned in the
3:49:45selling part of this video will be on
3:49:46the school post for this video. So you
3:49:48go into school, you go to the YouTube
3:49:49resources tab, and then this video will
3:49:51be right there. So don't forget to check
3:49:52all those out. And of course, all of the
3:49:54resources for the tutorials, if you
3:49:56haven't already done them, are included
3:49:57on those posts as well. And finally, if
3:49:59you made it this far, could you please
3:50:00do me a favor, leave a like on the
3:50:02video, and drop a comment down below.
3:50:04Let me know what you like the most, what
3:50:05you want to see more of. Click the share
3:50:06button, send it to your friends and
3:50:08family and loved ones so they can start
3:50:09learning these skills, too. All of these
3:50:11actions really help my video to reach
3:50:12more people in the YouTube algorithm.
3:50:13And if you subscribe to the channel,
3:50:15you'll be able to see a lot more content
3:50:16like this helping you to understand AI
3:50:18and more so how to build businesses
3:50:20around this incredible opportunity. And
3:50:22if you're still hungry for more and you
3:50:23want to watch my complete guide to
3:50:24building an AI business, that's going to
3:50:26be linked up here. But that is all for
3:50:28the video. I'm so excited for you to get
3:50:30cracking on this. I sincerely hope
3:50:32you've got something out of this because
3:50:33I put a lot of work into it as did my
3:50:34team. So I'm just really really wishing
3:50:36you all all the all all the best. I'll
3:50:37see you in the next one.

About this transcript

This page contains the full transcript of How to Build & Sell AI Agents: Ultimate Beginner’s Guide by Liam Ottley, generated from the public captions YouTube serves with the video. The transcript has 50,742 words across 7,072 segments, with the original timestamps preserved so you can click any line to jump to that moment in the embedded player.

What you can do with it

Use the transcript to take notes, quote the speaker, build a study guide, generate a summary with ChatGPT or Claude via the YouTube Summary tool, or export it as a timed subtitle file with YouTube to SRT. You can also re-open it in the transcriber to translate the transcript into 100+ languages.

Free YouTube transcript tool

YouTube2Text is a free YouTube transcript generator — no signup, no daily limit. Paste any YouTube link and get the full transcript instantly, with timestamps, click-to-jump, translation to 100+ languages, AI prompts for ChatGPT, Claude, and Gemini, and exports to TXT, SRT, VTT, or Markdown.