YT YouTube2Text

YouTube2Text → Transcripts → Accessibility at an inflection point: Regulation, AI agents, and what comes next I Axe-con 2026

Accessibility at an inflection point: Regulation, AI agents, and what comes next I Axe-con 2026 — Transcript

by Deque Systems · 6,931 words · 1,087 segments · language en · Watch on YouTube

Open in transcriber Summarize with AI Download SRT/VTT

Full transcript

0:00Hello everyone, welcome to Axecon. I am
0:03Priti Kumar, founder and CEO of DQ
0:06Systems and I'm joined by our chief
0:08product officer, Dylan Barrow.
0:11It is a true honor to kick off this
0:14event
0:15and I as I begin these remarks I'm
0:18struck by the sheer sharing of this
0:21conference hence of thousands of you
0:25from every corner of the globe unified
0:29by a single relentless passion for
0:34digital accessibility.
0:37The Axecon experience is a testament to
0:41what happens when a community stops just
0:45wishing for change and starts demanding
0:49it. Your presence here is the heartbeat
0:53of this movement.
1:00Axcon has always looked forward to the
1:04future of accessibility,
1:07but this year
1:09we are truly teaching an inflection
1:11plan.
1:14The way we interact with computers,
1:18the way they understand us is
1:22fundamentally shifting.
1:24And because of this,
1:27accessibility can no longer be a
1:30downstream task or an asterox.
1:34We've all made some progress,
1:37but it must now be woven into the core
1:42infrastructure of how code and content
1:46is born.
1:48Today,
1:50the regulations require it.
1:54And as our virtual and physical world
1:57becomes more and more indistinguishable
2:00from one another and deeply intertwined,
2:04a barrierfree world is no longer just a
2:07goal. It is an absolute imperative.
2:16Change doesn't happen in a vacuum. It
2:20happens because of you.
2:22On the screen, we have a collage of the
2:25champions we've had the privilege to
2:27work with. People who wear their
2:31accessibility pins not just as an award,
2:35but as a promise.
2:39You are the ones who will take these
2:41insights from these two beliefs at gone
2:44and turn them into new momentum required
2:47to lead us through this new era that we
2:49are all living through.
2:52So last year we spoke about the upcoming
2:54changes in digital accessibility.
2:57But today we aren't making predictions
2:59anymore. We are in it. This is the
3:03moment.
3:05We've reached this peak because of your
3:07work and yes also because the world has
3:11caught us. Accessibility regulation is
3:14now through the global from the EAA to
3:18ADA title 2 to FBI. And for the first
3:22time in history, our technology is
3:24actually powerful enough to make digital
3:28equality a reality and not just a goal.
3:32What we've engaged in together is
3:35transformation. And ADSCON is the medium
3:38where knowledge sharing becomes progress
3:41making.
3:42And as we continue with this
3:44presentation, Jill and I will discuss
3:46the opportunities and challenges of us
3:49and the tools that we have to achieve
3:51our goals. You will hear from the titans
3:55leading the industry, Meta, HSBC,
3:58Microsoft, fans and Reuters and more.
4:02You will hear about the expanding impact
4:05of the EAA and ADA title huge and the
4:09culture shifts they are triggering.
4:12You'll hear about the technological
4:14advancements and our own innovations
4:16like ax MCT server and automated
4:19intelligent by def and um that are
4:23designed to help you. And now with over
4:2640 presentations to follow,
4:29you are standing in a ground swell of
4:34energy.
4:35Digital accessibility has momentum. Not
4:39that we added it to our roadmap
4:41momentum. Not that we hired one person
4:44and gave them no budget momentum, but
4:46real momentum and it is growing.
4:51These claimed that EAA is global. Now we
4:56have evidence 2025 was the year that
4:59accessibility moves from a
5:01recommendation to a requirement at an
5:03unprecedented scale. In the first half
5:06of the year, the EAA and Australia's
5:09conformance uh deadlines hit. By July,
5:13sebies uh regulations brought the hammer
5:16down on financial institutions and by
5:18fall or autumn, the pressure was
5:21visible. France demanded grocerers fix
5:24their websites. The Netherlands uh
5:27dispatched non-informance letters to
5:29global giants and Sweden began active
5:34inspections.
5:36On November 12th, we saw the first major
5:39seat of the EAA when the French
5:41disability organizations filed emergency
5:45injunctions against four major grocery
5:48chains. And meanwhile, in Norway, the
5:50health app um Health Si received a
5:54notice of non-compliance with a
5:56staggering daily fine of 50,000 women.
6:00The message of 2025 was clear. The event
6:04the runway is ending.
6:09It's only February 2026, but the wait
6:12and see era is officially dead. On our
6:16most recent visits to Bangalore and
6:19Dublin, the sentiment was unanimous. So
6:22is picking up and the EU working group
6:25is now beginning to share results across
6:28borders. We have the ADA title 2
6:31deadline in April. the HHS section 504
6:35deadline will hit in May. Around the
6:38same time, we expect the new PN301549
6:42standard to be published and then the
6:44FBI remediation deadline arrives in
6:47July. Phew.
6:50Oh, that's a lot. And if you're
6:51receiving score, that's April, May, May,
6:55July. And that's not a compliance
6:57calendar.
6:58That is a countdown. accessibility to be
7:02no longer a project, it's a process.
7:07So the first watershed moment is April
7:1024th, 2026,
7:12which is the deadlines for state and
7:14local governments to make their digital
7:17properties accessible and that is 8
7:20weeks away. We are seeing the closer
7:24inspection of VPADS than ever before.
7:27Yeah, states are rightfully terrified of
7:30the liability and interestingly AI is
7:33acting like a force multiplier.
7:36AI is democratizing the ability to file
7:39the legal challenges
7:41and yet only 14% of the school districts
7:45are prepared. They're scrambling to gain
7:48ground in a race against time.
7:52The second watershed will be the
7:54one-year anniversary of the EAA this
7:56June. I predict we will see escalating
8:00penalties and intense public scrutiny.
8:04Navigating the EAA requires
8:07regional expertise. You cannot treat
8:09your Europe as a monolith. The
8:11requirements in France RGAA are
8:14different from Germany BFSG.
8:18For nonhuman countries, the GDPR effect
8:21is real. And for anyone unfamiliar with
8:24GDPR,
8:26uh it is the EU strict privacy law that
8:29makes every website on Earth ask you
8:32about cookies. The EAA will follow the
8:35same path except you can't select accept
8:38all and move on. If you do business
8:42globally, your one way of and waiting
8:44and seeing is completely wrong.
8:48Now,
8:50I've just spent 10 minutes telling you
8:53the momentum is unstoppable.
8:56Regulations are expanding. Deadlines are
8:59real. Fines are real. I could end the
9:02keynote right here and you walk out
9:04feeling great.
9:07But you didn't come to Axe Con to feel
9:10great. Well, maybe that's you. But you
9:14came because you know there's still work
9:16to do and challenges ahead.
9:22Rana talked about this. We aim reports
9:25that 94.8%
9:28of the top 1 million homepages still
9:30have detectable work as failures.
9:34Mobile is even worse. 90% of the top iOS
9:38and Android apps fail the basic
9:40standards according to Ark Touch. This
9:43is happening despite the fact that
9:45people with disabilities represent $1.9
9:49trillion in annual spending.
9:53Why is this state of digital
9:55accessibility? Why is it so good? Look
9:59at 2026, right? We have self-driving
10:02cars. We have robots that can do surgery
10:06and AI that can write poetry. And most
10:09of the world's login buttons still don't
10:11have accessible names. That's where we
10:14are. Why?
10:19The first reason, our processes are
10:22broken. Accessibility is still too slow
10:25and too manual for development teams. I
10:29spoke with a major bank and they told me
10:31they justified their entire return on
10:34investment for DQ's tools just by
10:38automating the defect creation process.
10:42Not the planning, not the require
10:44requirements analysis, not the
10:46development, etc., etc.
10:50Just the logging.
10:53They were spending so much time writing
10:55tickets about the problem. They didn't
10:57have time to fix the problems. By the
11:00time a bug is manually logged, the
11:03developer has moved on mentally,
11:06emotionally, and sometimes to another
11:08company.
11:09Slow processes create backlog. Backlog
11:13creates debt. And debt left alone long
11:17enough stops being a backlog. It starts
11:20taking over entire releases.
11:27Now you might be thinking well maybe
11:29this is exactly the kind of problem AI
11:32should solve and you may be right except
11:36AI gener generated code has brought a
11:40new problem of its own
11:43AI is prolific
11:45but it is often confidently wrong. I
11:49recently asked an LLM, who is impacted
11:53by lack of keyboard access? Simple
11:55question. It initially listed only
11:59people with motor disabilities. And when
12:01I challenged it, it said, "You're
12:03right." Just like that. No push back, no
12:06questions, just instant agreement. And
12:09in the second response, it recalled the
12:12blind, low vision, and cognitive
12:15disability users it had excluded just 10
12:17seconds ago. It even added power users
12:21and developers to the category.
12:24Understandably so. It wasn't trained on
12:28an accessibility first knowledge base.
12:31It was trained on the internet. And we
12:33just spent five minutes talking about
12:35how broken the internet is for
12:37accessibility. So we trained AI on the
12:40broken web and then we asked it to fix
12:42the broken web. How do you think that's
12:45going to go? Inconsistent AI output
12:49creates a compliance risk that
12:51regulators will not overlook.
12:54If your AI doesn't have the right rules,
12:58the right prompts, and some guard rails,
13:01it's not solving your technical debt.
13:04It's actually just generating it faster.
13:09So, the combination of slow processes
13:12and and the sheer volume of new code
13:14generated by AI
13:17is a recipe for disaster. It causes
13:20defects to be caught late in the
13:23development life cycle, resulting in bug
13:25backlogs that are simply unmanageable.
13:31To solve this, we see companies fall
13:35into the point solution trap. They use
13:38different tools for designers,
13:40developers, testers, and compliance
13:43which inevitably produce disperate
13:46results.
13:48So now you don't have one accessibility
13:51problem, you have four accessibility
13:55problems and they don't agree with each
13:58other. We've seen this before with
14:01screen reader testing.
14:03Teams began tailoring their source code
14:06to specific screen reader quirks rather
14:09than universal assistive technology
14:12standards.
14:14The standards are here for a reason.
14:17They're here to ensure compatibility now
14:21and into the future.
14:24Let's be clear,
14:26technical compliance is the floor, not
14:29the ceiling. There is absolutely a time
14:32and place for manual testing. But that
14:35time should be spent validating
14:37usability and the human experience and
14:40governance of compliance not debugging
14:42why for example NVDA is outputting the
14:46wrong radio button count. Today
14:50we are seeing this pattern of
14:52fragmentation repeat with the EAA. While
14:55the EAA relies on EN301549
14:58the standard as the common ground.
15:01fragmented interpretations by different
15:03member states is introducing complexity
15:05that slows down the ability to scale.
15:09And this reminds me of the landscape a
15:11decade ago when non-standard
15:13interpretations of work made it
15:15difficult for teams to move fast. And
15:16that challenge is exactly by the
15:19open-sourced act score. We knew then
15:23what we know now. Accessibility only
15:26scales when we agree on a single source
15:28of truth. Adhering to universal
15:30standards is the only way to ensure your
15:33quote is compatible with a screen reader
15:35in Berlin, a screen magnifier in Lisbon
15:40or a refreshable braille display in
15:42Bangalore.
15:44And by automating the standards layer,
15:47whether that is code written by a human
15:49or by an AI agent, we finally free up
15:54the accessibility experts to focus on
15:56what they're truly passionate about. the
15:59human experience.
16:03We have a unique opportunity right now.
16:07By having humans and AI agents
16:10collaborate, we can turn accessibility
16:12from a bottleneck into a seamless part
16:15of the digital supply chain. We can
16:18erase finally erase the challenges we've
16:21been discussing for decades. And I mean
16:23decades.
16:25So, what does this look like?
16:28you as the user directing agents as one
16:33of your tools in your daily work.
16:38So we need a multi-layered platform to
16:41make these agents
16:43experts at accessibility.
16:46The user there, humans telling the
16:49agents what to do and providing the
16:52instructions and oversight of the
16:55agentled work and doing the work that
16:57really matters.
17:00The agent player, the autonomous task
17:03doers like Claude, code, open AI codeex
17:08that take the guidance from users and
17:11execute the work.
17:13the integration layer using standards
17:15like MCP to expose the right tools and
17:20data at the right time so that agents
17:23can automate
17:25whole accessibility workflows that cut
17:27across multiple business systems like
17:30Jira, other compliance systems like
17:32security, analytics and more.
17:36the rules and knowledge layer like the
17:39axe platform in DQ University and others
17:43structured machine actionable rules and
17:46knowledge that provides the specialist
17:49knowledge and guardrails and the checks
17:51on the agents work.
17:54This is a new ecosystem where AI
17:57verifies its own work against gold
17:59standard rules. This is what it looks
18:02like. And the human providing
18:04instructions and oversight gets results
18:07and answers at the speed of AI.
18:11So what does this do for you?
18:14Let's give it give you an example. You
18:17can set up a Slack bot to go fetch the
18:19data from apps monitor and reports, do
18:22the analysis and tell you what actions
18:25you need to take, which team do you need
18:28to go talk to and provide further
18:30training to. Dylan will show you this in
18:33action in the presentation very soon.
18:37This isn't the future, it's now.
18:4184% of front-end developers are already
18:45using AI and by the end of last year,
18:48cloud code produced
18:51195 million lines of code weekly for the
18:54for I think 115,000 developers. And this
18:57is just one of the many tools.
19:00Gartner predicts that by 2029
19:0460% of sites and apps will be
19:06architected primarily for AI agent
19:09consumption with humanfacing UX becoming
19:12a secondary layer as users increasingly
19:15delegate tasks to intelligent agents.
19:20Let that sit for a while.
19:23We spent 30 years creating a web that
19:25works for humans. Sort of. Now we have
19:29to create a web that works with humans
19:31and their assistant agent.
19:34And we must ensure that when an agent
19:37talks to an agent or a human talks to a
19:39natural language interface or interacts
19:42with a generative
19:44interface compiled on the fly. It is
19:47accessible by design not retrofitted not
19:50batched by design.
19:53Of
19:54course, technology moves faster than
19:57culture.
19:59Our customers fall into three camps. The
20:01averse, highly regulated banks, for
20:04example, who say, uh, we're not going to
20:07displace critical thinking with AI,
20:10the hesitant. Telecoms, for example,
20:13told me, one telecom told me, well,
20:15we've been told to be efficient, but
20:17fear the risk,
20:19the excited.
20:21a big tech chief accessibility officer
20:24who sees this as the only way to
20:26innovate our way out of the problem. So
20:30no, maybe. And finally, wherever you
20:34sit, the good news is that technology
20:36can meet you where you're at.
20:39We are at a unique inflection point.
20:41You've heard that a couple of times. We
20:43don't have to rely on the bad data AI
20:45was trained on. We can supplement it
20:47with structured machine readable use. We
20:50can use MCP to let AI verify its own
20:53work and iterate until it passes the
20:56tests. We have the tools, we have the
20:58standard, we have the deadlines, plenty
21:01of them. We just need the will to seize
21:04the moment.
21:07Uh we see the world's leading companies
21:10seasing this AI moments with incredible
21:13ambition. JP Martin Chase is deploying
21:15coding assisted agents to build the
21:18future of banking. American Express is
21:21reimagining customer service through AI
21:23power and chat. Walmart is architecting
21:26super agents to revolutionalize a
21:29millions of people's shop. But here's
21:31the hard truth. While these companies
21:33are seizing the AI moment, it is up to
21:36the champions within them, the people
21:38here to ensure accessibility is a part
21:41of that AI in infrastructure from day
21:44one. If we don't make it in now, we
21:46aren't innovating. We are just
21:48automating exclusion.
21:52We have the vision. Dylan has the engine
21:54to show you exactly how we turn this AI
21:56empirical into reality. Here is our
21:58chief product officer, Dylan Barrow.
22:02Thank you, Pryy. Um, as PY already said,
22:05AI is changing the way we all work and
22:07change the change is speeding up too.
22:10But the good news is that AI agents will
22:12prioritize accessibility. Does this
22:14statement surprise you? Before I tell
22:17you how AI agents will prioritize
22:19accessibility, let me impress on you all
22:21that we have a huge opportunity in our
22:23hands right now for accessibility. And
22:26this opportunity is an opportunity for
22:28you too. the opportunity to leverage the
22:31momentum of AI. AI is just about the
22:34only thing getting funding right now and
22:36the floodgates are open. This means that
22:38you can get funding for accessibility in
22:40the context of these AI initiatives. AI
22:43is changing the way we work and the
22:45workflows themselves. In some cases,
22:47these workflows are being completely or
22:49mostly automated. This means that you
22:52can leverage this moment to embed
22:54accessibility into these workflows. If
22:56you've ever been involved in the
22:58software rewrite, you will know it's
22:59easier to write the functionality, if
23:01you know about it from the start, than
23:03it is to retrofit it. Now is the time to
23:06get accessibility built into the AI
23:08workflows.
23:10At the same time, you can position
23:11yourself within the organization as a
23:13force for limiting the risk of these
23:16changed workflows, allowing your
23:18organization to reach the benefits of AI
23:20while staying compliant or even getting
23:22ahead in compliance.
23:24AI agents don't object to doing things.
23:27They just need to be enabled with the
23:29right tools and instructions.
23:34Let me give you a little taste of what's
23:36possible.
23:38I press play here. I've set up a
23:41personal bot. My bot can talk to me on
23:43Telegram and I ask it to stay on top of
23:45some of the things I need to pay
23:46attention to. When it notices something,
23:49it sends me a message. Here it has seen
23:51that a new issue was created in one of
23:53my products that's related to
23:54accessibility. I can tell it to go and
23:57address it.
23:59Once it knows what's going on, it
24:00updates me with its plan. At any point
24:03in this process, I can change my
24:05instructions, ask it to give me more
24:07information, or tell it to stop. In this
24:09case, when it comes back with a plan,
24:12let's give it a little bit of time here.
24:13Oh, there's the message. I can peruse
24:17that plan, and I can see that it's a
24:18good plan. Now I've I've set up my bot
24:21that it has access to all the AXDE tools
24:24MCP server and that means that it also
24:26has access to the analyze and remediate
24:28tools that that provides. I've given it
24:30instructions to use Axe Dev Tools MCP
24:33server when addressing accessibility
24:35issues.
24:36When it's done, it sends me a message
24:38with a link to the pull request so I can
24:40inspect the code changes. Let's have a
24:42look to see if the changes are good.
24:45So we'll open up GitHub here. And the
24:47first thing we'll notice as as we scroll
24:49down is that the fi there are four files
24:51that have changed. Well, let's look at
24:53those files. The first change here is
24:55one where it's changed the table uh what
24:59were previously table cells that should
25:00have been header cells into header
25:02cells. The second change it's made is to
25:05uh fix a color contrast issue by making
25:07the background darker. The third thing
25:09it's done is added an alt attribute to
25:12an image that was missing an an alt
25:14attribute by referencing the name in the
25:16code. And then the fourth thing that
25:18it's done is add some area labels to
25:21some uh icon links so that there's an
25:25accessible name for those too. So
25:27obviously Axe DevTools MCP and tools
25:29like it are one way you can make your
25:31agents better at accessibility.
25:35But how can we really seize this
25:36opportunity together?
25:39Well, the same principles apply here to
25:41the ones that apply for successful
25:43digital accessibility programs, but they
25:45are magnified. So, the degree to which
25:47you make this true will determine the
25:49degree to which your AI workflows will
25:51be good at accessibility.
25:53They need to be efficient. Now, with AI,
25:55this is somewhat diminished because the
25:57AI agent doesn't mind doing work and
25:59it's mostly so fast that it does not
26:01make a big difference in the overall
26:02productivity. But what you do want to
26:05ensure is that you're not giving your
26:06agent instructions that will result in
26:09unnecessary or even counterproductive
26:11work. So the quality of the tools and
26:13the quality of the advice is important.
26:16Secondly, and most importantly, the work
26:18needs to be streamlined. This means that
26:20it needs to fit into the way the
26:22workflow is designed. It needs to speak
26:24AI. Some of you may have seen that
26:27initially there was a lot of talk about
26:29agents browsing the web and doing things
26:30through existing UI. Well, it turns out
26:33that that is not a native AI way of
26:35doing things. So now we have MCP, we
26:38have the universal uh uh commerce
26:41protocol and we have webmc and other
26:44emerging standards that are much more AI
26:46native.
26:48Thirdly, it needs to be scalable. This
26:50means that it needs to be automated as
26:52much as possible. As AI scales, the
26:55accessibility process that's embedded
26:56into it must also scale. This means that
26:59for the majority of the work, it needs
27:00to be automated wherever the AI workflow
27:03needs it to be automated.
27:06So you should pay really pay attention
27:08to two things. Firstly, don't just
27:11accept that these workflows are fixed.
27:13They're not fixed things. They are
27:15highly customizable. Whether it's
27:17through leveraging MCP tools or skills
27:20or plugins or hooks, there are ways you
27:22can make the AI do stuff. And there are
27:25ways you can tell the AI to make the
27:27human pay attention to stuff. So take
27:30charge of the AI workflow at your
27:33company. Create and acquire the best
27:35tools to accessibility your company's AI
27:38agents. Create skills, plugins, and
27:40hooks. See how far you can push the
27:43limits. It may seem daunting at first,
27:46but actually the AI can help you do
27:47this. Just ask it.
27:50Secondly, the most tedious item in the
27:52accessibility work we do is fixing and
27:54testing. Try as much as you can to
27:57automate this tedious work. The more you
28:00automate this, the more AI capable your
28:02agents will be.
28:05At DQ, our goal is pretty simple. We
28:08want to put all the tools and data of
28:09the AX platform in the digital hands of
28:12your AI agents. Whether that is creating
28:15a monitoring scan to monitor a site or
28:18whether that is finding out which of
28:19your portfolio's applications has had
28:22the biggest drop in accessibility, you
28:24need to be able to do it through an
28:25agent. Our principles are simple and you
28:28should adopt the same or similar
28:30principles. Don't toil, get results.
28:34Don't hunt for information, get answers.
28:37This is how AI is changing our
28:39expectations and our lives.
28:42Let you show me what let me show you
28:44what this might look like for a product
28:46manager of an application. Kate Kate has
28:50her co-work agent set up with access to
28:52Axe Monitor and some of her other
28:54corporate tools. She wants to know what
28:56the accessibility of her site is like.
28:58She doesn't have to log in to Axe
29:00Monitor. The Ax MCP server knows where
29:02the data is and can map from her query
29:05about the recipe site to the scan inside
29:08monitor. To her horror, she can see that
29:10the score has been sinking over the last
29:12couple of days. This could be a
29:14disaster. Kate's a visual learner, so
29:16she asked for a bar chart so she can
29:18understand this a bit better. This is
29:20one of the main advantages of an AI
29:22agent with access to your data. It can
29:24simply create visualizations of that
29:26data for you on the fly, giving you
29:28answers immediately.
29:31Inspecting the bar chart, Kate can see
29:32that the biggest increase is in the
29:34serious issues, but there are a lot of
29:36critical issues, too. She needs to take
29:38action.
29:39quarterly reviews in four weeks and this
29:41is on the agenda. She asked for a plan
29:43for her team to get the numbers above
29:45her corporate threshold of 80%. The AI
29:48agent puts together a 3-week plan. Week
29:51one is to fix all the critical issues
29:53and the agent even gives HTML code
29:55markup suggestions for how to address
29:57the issues. She could easily push these
29:59to get over tickets from her agent. The
30:02agent tells her this will get a score up
30:04to 50% because it knows the X monitor
30:06scoring algorithm.
30:08With two more weeks of effort after
30:10that, the agent predicts that the team
30:12could be at 96%.
30:15This would be the highest score ever.
30:18This is the the accessibility management
30:20at the speed of AI. The good news, you
30:23can sign up now for our beta program for
30:25the Ax MCP server for Axe Monitor and
30:27get early access to these powerful
30:29tools. Just look for the links in the
30:31chat or the transcript transcript.
30:36The second thing we're doing is
30:37constantly driving the amount of
30:38everything we can automate down. Sorry,
30:41up
30:43could be down. Last year we released the
30:45AIdriven IGTS and advanced rules. High
30:48quality rules with low to zero false
30:50positives that do more of the testing
30:52for you or for your agent.
30:55Our AI powered features include the XMCP
30:57server, automated IGTS
31:00or intelligent guided uh tests for those
31:02of you who are not familiar with them.
31:04Advanced rules and action assistant.
31:06They help you test faster and at scale.
31:09Find, fix, validate issues as you code.
31:12Avoid agent sprawl by integrating into
31:15your company's AI agents, not creating
31:17new AI agents. Increasing team adoption,
31:21the team adoption of your of the tools
31:23because they make it so easy. Getting on
31:26demand expert back successful knowledge
31:28and guidance immediately.
31:30On top of that, with our Jira
31:32integration, EIA and RGA a standards
31:35alignment, you can more easily and
31:36confidently streamline dev workflows and
31:39tracking, maintain compliance reporting
31:42with global standards. When AI powered
31:44features partner with humanentric
31:46workflows, your accessibility strategy
31:49becomes more powerful than ever. And DQ
31:51is here to help make it all possible.
31:55With that, I want to make you aware of
31:57two presentations. one by Harris
31:59Schneiderman on how to get results
32:01without toiling called shift left
32:03without shifting gears and the second
32:06one on AI automating more testing called
32:09Axe innovations harnessing AI
32:11responsibly and with that I'm going to
32:13hand you back over to Py
32:23we can't hear you
32:25>> uh okay well of course you can't hear me
32:28because I'm on mute. Hello everybody.
32:30Didn't just showed you the engine, the
32:33Axe MCP server, automated IGTS,
32:38the agent to track compliance, and our
32:40grounded trusted rules.
32:44Now, let's talk about how you can use
32:48this to lead
32:50to navigate this new landscape. Your
32:52strategy
32:54must include three core pillars.
32:57Number one, get leadership to buy in.
33:01The EAA and Title 2 aren't just
33:03regulations. They are your leverage.
33:07Use this legal shift to move
33:10accessibility from a nice to have to a
33:13non-negotiable leadership priority.
33:17Two, equip your teams.
33:21Strategy fails without literacy.
33:26move your teams from how do I get
33:29started to having the tools and training
33:32to get going.
33:35Three,
33:37get a baseline.
33:39You can't fix
33:41what you haven't measured. Stop the
33:45guesswork.
33:46Implement the right tech stack to
33:49automate the standards layer so your
33:51human experts can finally focus on what
33:56matters, the human experience.
34:03We are ending the era of the Friday
34:09afternoon nightmare we all know too
34:11well.
34:13No manual only testing loops. No more
34:17missed deadlines. No more audit
34:19graveyards.
34:23No more inconsistent standards. And no
34:27more point solutions that don't scale.
34:33Agentic remediation, as Dylan showed
34:35you, can replace tickets with pull
34:38requests.
34:40And I know how. I know some of you felt
34:44that replacing tickets with pull
34:47requests. Agents find, fix and validate
34:51code in real time,
34:54guided by you, matching your patterns
34:57and work act standards before humans
35:01look at it and make sure it's right.
35:06Grounded and trusted truth.
35:08We are curing what I politely called the
35:13confidently wrong problem, but let's
35:16call it what it is, the confident liar
35:19problem.
35:21We are grounding AI in expert knowledge.
35:25We have eliminated some of the
35:27hallucinations that create technical
35:29debt. And through the validation loop,
35:33we can make sure humans are flagged if
35:36the AI goes off the rails. So no more
35:39guessing
35:41and accessibility at AI dev ops speed.
35:45By moving remediation into the IDE
35:49and the PR stage, we can turn
35:51accessibility into a standard of
35:54quality,
35:56not a downstream bug you are told about
35:59on a Friday afternoon.
36:04As we said right at the beginning,
36:08this is all about transforming knowledge
36:13sharing into progress making.
36:17You have two full days ahead of you,
36:20overflowing with opportunities to gain
36:22and share that knowledge. Use them,
36:26connect with each other. If you haven't
36:29already, join the Axecon Discord.
36:33This community is your greatest
36:35resource. Lean on it.
36:38Take a deep breath.
36:40Check your itineraries
36:43and get ready to immerse yourself.
36:46This is the power of community. This is
36:49energy in action. This is axecon. Let's
36:53get your book.
36:57>> Fantastic. Thank you so much, Py. Uh
36:59Dylan, thank you as well. Um really
37:03fantastic. I'm I'm I'm feeling
37:04energized. Um judging by the volume of
37:08tra uh chat and uh the amount of uh Q&A
37:11coming through, uh we've got an
37:13energized group here in attendance. Um
37:16just a reminder for all of y'all u
37:18please bring your questions to the Q&A
37:20section within the chat module. um and
37:22make use of that upvote that thumbs up
37:25feature uh to help the the the best
37:28questions bubble up to the top so I can
37:30relay them to Dylan and Priy here. Um
37:32Dylan and Priy the first question I've
37:34got there are quite a few questions
37:36around the same theme. Uh this question
37:39is how viable are AI tools for
37:42development when studies have shown that
37:44hallucinations are fundamental to
37:47underlying technology that LLMs are
37:49based on. There are quite a few other
37:50questions about that relative to AI in
37:53accessibility, AI in development. How do
37:55we trust this?
37:58>> Well, I think my answer to that is you
38:00don't trust, you you verify. So, I've
38:03been using AI agents a lot to code and I
38:07find that uh first of all, you know, you
38:10need to ask them to do a very good plan
38:13and you need to inspect that plan in
38:15detail. uh and even if you've given them
38:17uh a plan of what you want them to do,
38:20they'll sometimes sort of forget that
38:22that the plan exists and and so you
38:25constantly have to ensure that the work
38:28that they're doing um in fact meets
38:30whatever requirements you you want it to
38:32meet. And so that same thing goes for
38:34the functional requirements for uh the
38:37implementation sometimes the goals of
38:39what you're trying to achieve. And it's
38:41of course going to apply uh for
38:42accessibility. So I think if there's one
38:46message that Piny and I are trying to
38:48get through in this overall
38:49presentation, it's uh take control of
38:51that. Don't just accept um the the the
38:54out the the out the output that you're
38:56getting because you can make it you can
38:58make the results a lot better just by
39:02the way that you actually use the tools
39:04yourself. If you just give them a sort
39:06of a vague um
39:09command or or instruction, of course
39:12they're going to, you know, you're going
39:14to end up with with something that's far
39:16away from what you intended. You have to
39:17be very precise. You have to structure
39:20the the work. You have to give them very
39:22detailed instructions. You have to make
39:24sure that they put those instructions
39:26into their memory. Remember them,
39:28remember the principles. You have to
39:29sometimes remind them of the principles.
39:31Now, now these things are getting better
39:33and better and better. I mean, they're
39:35already sort of streets ahead now than
39:37they were 3 months ago, and they're
39:39streets ahead of where they were 3
39:40months before that. The underlying LLMs
39:43are actually hugely powerful. And what a
39:45lot of the innovation that's happening
39:47right now is around how to structure and
39:50use that in such a way that we get uh
39:52good good outcomes. And so, Claude code
39:55has made huge um advances here. There's
39:58things like open claw that have been
39:59released. They take different approaches
40:02and and a lot of what these things are
40:04doing is simply structuring uh what they
40:07ask the LLM to do in such a way that
40:09they get better results. Now those same
40:11things are going to apply those same
40:13advances are going to apply to
40:15accessibility
40:17um as long as you have the right
40:20knowledge and tools and and ways of
40:22checking that what the actual agent is
40:24doing is accessible. So, so that's the
40:26second thing that I think Py and I are
40:29are trying to get across is first of all
40:32take charge of this, learn how to use it
40:34properly, customize it, etc. But second
40:36of all, you got to make it an expert at
40:38accessibility. You got to give it the
40:40tools to so it can help itself and then
40:43you've got to have the the human
40:45oversight. You got to check its work.
40:47Um, it's not about trust at all. I don't
40:50think we can trust any of these things.
40:52not as much as we can trust any of our
40:54developers to to not make mistakes or do
40:57exactly what what what uh what we
41:00thought we asked them to do.
41:02>> And the only thing I'll add to that Ryan
41:04before I you know you ask the next
41:06question is
41:09I I've been doing a lot of coding as
41:10well with agents and what I've noticed
41:12is that you have to iterate a lot. So
41:16it's not just one and done. You can't
41:19just say okay this is the output I'm
41:21going to take it for granted. You do
41:22have to validate trust and verify like
41:24Dylan said but it's also a very
41:26iterative process. So you have to uh
41:30remember that you know. Yeah.
41:32>> Thank you. Um follow-up question. Um
41:34actually a two-parter. Um could you
41:37clarify
41:39um what the axe tools are trained on and
41:42then also there are a number of folks
41:44are aware of the no false positives
41:46mantra associated with axe core. Could
41:49you elaborate on whether no false
41:51positives also applies to this this
41:54training?
41:56>> So yeah the the no false positives
41:58mantra of uh Xcore sometimes gets
42:01misunderstood. I'll start with that
42:03because it's misunderstood like of
42:05course there are going to be bugs and of
42:07course there are going to be things that
42:09we get wrong. Uh also some things can
42:12appear as false positives where new
42:14technology has come up and something
42:16that used to work in a very reliable
42:18accurate way no longer works as reliably
42:21and accurately and it has to be updated.
42:23So so when we say no false positives
42:26what do we mean? What we mean by that is
42:29that every time we think about creating
42:31a rule for ax core, we think about first
42:34of all can we uh is it feasible for us
42:38to create a rule which given everything
42:40that we know about the technology etc
42:42has a very high chance of being
42:44extremely accurate right and by that we
42:47mean you know extremely low if we test
42:51thousands of issues say we tested 10,000
42:54issues we expect them sorry elements we
42:57expect the number of uh issues that
42:59occur to be less than one in a sort of a
43:02thousand. That's what we're kind of
43:04aiming for something in that range. And
43:06if you look at the number of false
43:07positives that do occur if you use
43:09Axore, it's extremely low. And that is
43:12partially that is you know because we've
43:15taken that care in asking ourselves
43:17during the design phase of everything
43:19single rule is it possible with the
43:21technology that we have today to do this
43:23in such a way that that that that error
43:25rate is going to be extremely extremely
43:27small. Right? So that's that's what zero
43:30false positives manifesto means to us.
43:33And we take that same zero false
43:35positives manifesto. It's an approach of
43:37of of designing the rules that that that
43:40you're actually going to even try to
43:42write such that they have a that you
43:44have a very high uh confidence that
43:47they're going to be they're going to be
43:48good. They're not going to be noisy. And
43:50we apply it to AI as well. I mean, if
43:52you really want to get into a lot of the
43:54details with this, please attend the the
43:56talk by Wilfierce and No Barrel because
43:59they will go into this into a lot more
44:01detail than I can go into here. But the
44:04principle is first of all we come up
44:06with ideas and then we critically
44:08evaluate them and say given the
44:09technology is there a chance is there a
44:12chance we could write this such that
44:13it'll be high quality and low false uh
44:16low false positive and if there isn't
44:17then we don't do it and then if we think
44:20that there is then the approach from
44:23from that point on is to try to
44:25implement it to test it against live
44:27data and to iterate on that and we won't
44:30release the rule unless we get to a
44:32point where that false positive rate is
44:34low with non-determin deterministic
44:38stuff like LLM in the picture. We also
44:41have started to introduce the concept of
44:43how you know how how noisy are you
44:46willing to allow this to be so that you
44:48can set that to some degree yourself.
44:50But we're also taking other approaches
44:52because some of the non-determinism is
44:55you know fi out of five times it might
44:57get it right three and wrong twi too
44:59there are other techniques that we can
45:00use to then get get you a much more
45:03consistent result all the time by by
45:06applying statistical techniques to that
45:08as well and we try to apply those on top
45:10of it as well. So for us it's to try
45:14when we say zero false positives we
45:15really mean taking that approach to very
45:18very high quality rules. Why do we do
45:20that? Why do we think that's important?
45:21Why don't we just, you know, come up
45:23with a bunch of rules that is um that
45:26are noisy? And the reason is because
45:28false positives are require effort to
45:31determine whether or not they're they're
45:33real or not. And that turns people off,
45:35especially if you're trying to move
45:37faster and faster and faster with AI.
45:39The more false positives you have, the
45:41more noise and the more you're likely to
45:43just not use the tool at all. So, so you
45:46know we we cannot achieve DQ's manifesto
45:50digital equality by creating noisy noisy
45:53tools and and that's the reason that we
45:55have this uh zero false positive um
45:58manifesto in the first place. I I don't
46:00know Ryan have I answered all aspects of
46:02that 2.5
46:04>> I think so. Yeah. Thank you. Thank you
46:05very much.
46:06>> Ryan, one one thing I'll add to that is
46:08that it is a continuous investment and
46:11we believe in that. We've got subject
46:13matter experts who are constantly giving
46:15feedback and uh creating test cases and
46:18creating um websites to test the rules
46:22before we release it. So yeah.
46:24>> Okay. I want to get in one more question
46:26here. We've got about two minutes. Um so
46:29if you could budget your your answer
46:30time to be about two minutes, that'd be
46:32fantastic. Um this is a trickier a
46:34trickier one though. Um even though
46:36legal deadlines like ADA title 2
46:38compliance are approaching, we see a lot
46:40of movement among government and private
46:43corporations to shove and distance
46:45themselves from DEI, right? Um how can
46:49we remain staunch advocates at our
46:52companies to make accessibility a
46:53priority when this is met from fear of
46:57retaliation or association with DEI?
47:03>> That's I'll let you take that. Yeah,
47:05that's that's very sad actually to hear.
47:09Um, we encourage you to use the
47:12regulations to really get your
47:15leadership to pay attention to this. You
47:18know, um, I think the human condition is
47:22such that things happen in waves. They
47:24come and go. What we have to do is we
47:27have to remain steadfast and we have to
47:30stay passionate.
47:33We have to keep the momentum for digital
47:35equality to actually prevail. Um whether
47:40it's AI, whether it's DEI,
47:43um we have to use whatever momentum
47:47builders we can get to get the message
47:49to our leadership and make accessibility
47:52a reality.
47:53>> Thank you, Pretty. I just wanted to add
47:54on to private corporations,
47:57don't forget uh there are other
48:00regulations. We talked about the
48:02European Accessibility Act, for example,
48:03in in this presentation, right? ADA
48:05title two isn't the only game in town.
48:07So,
48:07>> yeah.
48:08>> Um, Dylan and Py, thank you so much. Uh,
48:10I've got to get folks out of here on
48:12time. Really appreciate it. Everyone in
48:15attendance, thank you for the fantastic
48:16everybody. Byebye. Q&A.
48:20>> Yeah, really appreciate your time. Um,
48:22have a great Axecon everybody.
48:23>> Bye.

About this transcript

This page contains the full transcript of Accessibility at an inflection point: Regulation, AI agents, and what comes next I Axe-con 2026 by Deque Systems, generated from the public captions YouTube serves with the video. The transcript has 6,931 words across 1,087 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.