Forward Deployed Engineer (FDE) Explained for Beginners — Transcript

Full transcript

1. 0:00AI gives software engineers superpowers
2. 0:02to 100x their output. And as AI models
3. 0:04become better, the GTM engineer was born
4. 0:07to rethink how companies bring products
5. 0:09to market faster and cheaper. And now
6. 0:11comes the hardest part, getting AI into
7. 0:13production inside real companies with
8. 0:16real workflows and real constraints.
9. 0:18That's where the forward deployed
10. 0:19engineer comes in. I'm Alex Lindahl. I
11. 0:21was one of the first GTM engineers at
12. 0:22Clay, and right now I'm helping to build
13. 0:24Clay's own FDE team. I've spent the last
14. 0:26year inside enterprise companies
15. 0:28building these AI systems firsthand. And
16. 0:30here's what that experience actually
17. 0:32taught me. The model isn't the
18. 0:33bottleneck, the implementation is. So,
19. 0:36in this video, I'm walking you through
20. 0:37what an FDE actually is, where the role
21. 0:39came from, what these engineers do
22. 0:41day-to-day, and what it means for your
23. 0:43go-to-market. And I'll also show you
24. 0:45exactly how we're running this at Clay.
25. 0:46But first, let me explain the core
26. 0:48problem the FDE was built to solve. I
27. 0:50call this the last mile problem, and the
28. 0:53FDE is basically the proven solution for
29. 0:55it. Companies assume that when you buy
30. 0:57an AI product, the hard part is the AI
31. 0:59tech. But from my experience, the hard
32. 1:01part is everything that happens
33. 1:03afterwards, when the AI has to touch
34. 1:05your actual data and your actual
35. 1:07workflows. And a16z put it perfectly,
36. 1:10enterprises buying AI are like your
37. 1:12grandma getting an iPhone. She wants to
38. 1:14use it, she just needs someone to set it
39. 1:16up. So, the FDE is that person, not a
40. 1:18consultant who hands over a slide deck,
41. 1:20not a support ticket that gets resolved
42. 1:21in 72 hours, an engineer living inside
43. 1:24the customer environment who's building,
44. 1:26watching, and fixing in real time. And
45. 1:28here's where it gets interesting,
46. 1:29because this model didn't come from
47. 1:31Silicon Valley. It came from somewhere
48. 1:33far more unusual, and it was built to
49. 1:35solve a specific problem. Back in 2001,
50. 1:38the US intelligence community had more
51. 1:40data than it had ever collected. It had
52. 1:42world-class analysts. What it did not
53. 1:44have was a way to make those two things
54. 1:46talk to each other. The CIA had its
55. 1:48systems, the NSA had its systems, the
56. 1:50FBI had its systems, and none of them
57. 1:53could see what the others were seeing.
58. 1:54So, Palantir was founded to solve
59. 1:56exactly that problem, but they hit a
60. 1:57wall immediately. A wall that shaped
61. 1:59everything they built after it. The
62. 2:01problem was they didn't know any spies
63. 2:03for the agency. And even if they did,
64. 2:05those individuals couldn't disclose the
65. 2:07details of their work. There were no
66. 2:08requirements documents, no user
67. 2:10interviews, no feedback loops.
68. 2:12Traditional product development was
69. 2:14completely useless. So Palantir did
70. 2:16something nobody else was doing. Instead
71. 2:18of asking customers what they wanted,
72. 2:20they put engineers directly inside
73. 2:22customer environments. The engineers
74. 2:24learn by doing, observing,
75. 2:25experimenting, and building in real
76. 2:27time. That decision created the FTE
77. 2:30model, which is wild when you think
78. 2:31about it. And the name actually came
79. 2:33from the military itself, forward
80. 2:34deployed forces. What started as a
81. 2:36necessity inside classified government
82. 2:38agencies has become one of the fastest
83. 2:40growing engineering roles in tech. And
84. 2:43the numbers show exactly that. What
85. 2:44we're going to do is we're going to
86. 2:45start by finding jobs. Because basically
87. 2:47what we want to do is find open jobs for
88. 2:50the forward deployed engineer. So we
89. 2:51want job titles that include, you
90. 2:53guessed it, forward deployed engineer.
91. 2:56So now we have a table with all the
92. 2:586,000 job openings. And when we do this,
93. 3:01what we ultimately want to do is find
94. 3:04the unique number of companies hiring.
95. 3:06So we can go down here and actually,
96. 3:08we'll do the dedupe on the company name.
97. 3:11So now this will eliminate almost 5,000
98. 3:13rows, and the result will be unique
99. 3:15companies that are hiring for the
100. 3:17forward deployed engineer. You have
101. 3:18companies like Workato, C3,
102. 3:21Happy Robot, 10X, Adobe. And we have a
103. 3:24number of different uh data in here
104. 3:26already, location, company name.
105. 3:29But let's actually get additional
106. 3:31information on the company. So we can
107. 3:33break this out into uh different
108. 3:35statistics on what size of companies are
109. 3:37hiring. So what I'm doing here is I'm
110. 3:39enriching the company by the domain, and
111. 3:41then we can find all this other
112. 3:42additional information such as country,
113. 3:45when the company was founded, location,
114. 3:47is it private or public? Now when we run
115. 3:49this, we'll run 10 rows first. I always
116. 3:52run 10 rows first just to make sure the
117. 3:54output is what I expected. So, it's
118. 3:56especially when you're running an AI
119. 3:57agent, you want to see what the output
120. 3:59is on 10 rows before you run, say like
121. 4:011,000 rows. And that helps you prevent
122. 4:03and save uh credits. Okay, so now we
123. 4:05have some more information here. Looks
124. 4:07good, so I'm going to run it for the
125. 4:08rest of the rest of the companies.
126. 4:11We have a little over 1,200 companies
127. 4:13that are hiring right now for a forward
128. 4:14deployed engineer. And by the way, this
129. 4:17same job posting data works for your own
130. 4:19pipeline. I have a free clay template
131. 4:21that runs exactly this. You drop in a
132. 4:22list of the company domains and out pops
133. 4:25all their open jobs and recent news.
134. 4:27Then AI turns those signals into natural
135. 4:29one-line icebreakers for every account.
136. 4:32It's free. The link's in the
137. 4:33description. Go grab it. Have fun with
138. 4:35it. But let's get back to the video. Now
139. 4:37we can go to Sculptor. Now Sculptor is
140. 4:39our agent that can do two things. One is
141. 4:43you can input natural language to tell
142. 4:45it what you want to build, and it'll go
143. 4:46off and build um a table or columns for
144. 4:49you.
145. 4:50But what I like to do is actually go to
146. 4:51the analyze mode, and this allows you to
147. 4:55layer AI on top of this entire data set.
148. 4:58So, we can do things like give me a
149. 5:00statistical
150. 5:02analysis
151. 5:04of the salary and size of companies that
152. 5:08are hiring a forward deployed engineer.
153. 5:11So, now what AI is doing is it's
154. 5:14analyzing all of this data, and then
155. 5:15it's going to give me a analysis of what
156. 5:18it says. Okay, so we just pulled job
157. 5:21openings for forward deployed engineers
158. 5:23at Google, OpenAI, and Salesforce. Okay,
159. 5:27now the next thing we can do is take the
160. 5:29LinkedIn URL, and we can create an agent
161. 5:31that analyzes the job posting for the
162. 5:35jobs to be done. So, one prompting
163. 5:36technique here is is just defining what
164. 5:39the output should look like. So, then we
165. 5:40can use it in a more structured way. So,
166. 5:42now we have a basic prompt, and we're
167. 5:44going to use uh AI to actually generate
168. 5:46a a more complete prompt for this
169. 5:48research. Now, what the agent is doing
170. 5:51is it's going off, going to the website
171. 5:54with the job posting, reading through
172. 5:56that job posting, and then parsing out
173. 5:58what are the jobs to be done and what
174. 6:00are the technologies that are mentioned.
175. 6:03And what we get back here is reasoning
176. 6:06of what it did, kind of like why it
177. 6:07output, and how it came to the
178. 6:09conclusions. It gives us the jobs to be
179. 6:11done, technologies, and then one nice
180. 6:15trick here is that it shows the steps
181. 6:16that the agent took. This helps with uh
182. 6:18debugging if you're not getting the
183. 6:20expected outputs. Obviously, it's just
184. 6:22going to one URL here, but for other
185. 6:25agents, it could have multiple different
186. 6:26steps. Let's uh dive into some of these
187. 6:30um jobs to be done to see what these
188. 6:31companies are looking for. Codify
189. 6:33repeatable deployment patterns and drive
190. 6:35insights to product and engineering.
191. 6:37Re-engineer isolated agents. Improve
192. 6:40engineering velocity and business
193. 6:41outcomes. And then translate ambiguous
194. 6:44business ambitions into standardized
195. 6:46technical roadmaps. So, that is one of
196. 6:47the challenges um with working with
197. 6:49customers and deploying AI is that
198. 6:51sometimes they have an idea of like what
199. 6:53they want um or an outcome that they
200. 6:54want, but it's challenging for them to
201. 6:57actually articulate what that is and how
202. 6:59to get there. So, we help them uh with
203. 7:01our forward deployed engineers to figure
204. 7:03out what outcomes they're looking for,
205. 7:05what are the use cases to kind of map to
206. 7:07achieving those outcomes, and then
207. 7:09building a roadmap that shows how we get
208. 7:11there step-by-step in terms of building
209. 7:13the tables, the workflows, and the
210. 7:15agents that that support it. Okay, let's
211. 7:16take another one here. Develop complex
212. 7:19AI applications. Provide white glove
213. 7:21deployment of those complex AI systems.
214. 7:24And here we're seeing it show up again,
215. 7:25transform field insights into product
216. 7:27roadmap feedback. So, again, the part of
217. 7:29the job of the FD is not only to help
218. 7:32deploy AI successfully into companies,
219. 7:35but it's also to bring all of the
220. 7:36insights from those customer deployments
221. 7:39back into the product to improve the
222. 7:40product uh much faster. And let's get a
223. 7:43breakdown of all the technologies that
224. 7:44are used. So again, we're using sculptor
225. 7:47to run AI across the entire data set,
226. 7:49all of the data and all of the different
227. 7:50columns. Now we have an analysis and
228. 7:53let's open up the doc to see what that
229. 7:55looks like. Cool. So some of these
230. 7:56Python very predominant, obviously
231. 7:58Google because the job is to work at
232. 8:01Google for text AI Gemini, that's
233. 8:03Google's model LangGraph, CrewAI, MCP
234. 8:06shows up quite a bit, rag, so on. Let's
235. 8:10dive a little deeper into the salaries
236. 8:12of what these companies are offering.
237. 8:14And what we should see here is we have
238. 8:17five different bands of company size. So
239. 8:19startup, early growth, mid-market, late
240. 8:21stage, and enterprise. And then we have
241. 8:23the different salary ranges for the
242. 8:25minimum starting point and then also the
243. 8:27max starting point. The minimum really
244. 8:28starts around I would say like 80 to
245. 8:31like 150 and the max ends around
246. 8:36uh like 250. So naturally the enterprise
247. 8:39companies are paying a lot more for
248. 8:41their FTEs compared to the startup. But
249. 8:43what's interesting is that, you know,
250. 8:45it's not only uh startups or early
251. 8:47growth companies that are hiring. We
252. 8:49really have a spectrum of companies that
253. 8:51are hiring this this role. Okay, another
254. 8:53question to ask is is this happening
255. 8:56just in the US or is it happening around
256. 8:58the world? So let's pull another uh
257. 9:01analysis on that. Great. Okay, so we
258. 9:03have some findings. Obviously United
259. 9:04States is close to 60% followed by UK
260. 9:0810% There's a pretty big drop off
261. 9:11between 60 and 10% and then another
262. 9:14pretty big drop off from the other
263. 9:16countries. Yeah, so diving into the data
264. 9:18a little bit more um let's look at the
265. 9:21technical landscape that these job
266. 9:23postings require. So the technical
267. 9:25requirements reveal a standardization
268. 9:27around modern stack fundamentals. Python
269. 9:30TypeScript 55% REST APIs 50% 40% AI LLM
270. 9:36requirement. They're not only asking
271. 9:38engineers to deploy software, they're
272. 9:40asking them to implement retrieval
273. 9:41augmented generation, tune prompts in
274. 9:43production, and debug agentic workflows
275. 9:45at customer sites. And then if we kind
276. 9:47of look at this in terms of the
277. 9:49responsibilities, they kind of cluster
278. 9:50in three different motions. So one is
279. 9:53building customer specific integrations,
280. 9:55that's 40% of openings. Second is
281. 9:58translating business requirements into
282. 9:59technical architecture, 30%, and then
283. 10:02owning end-to-end delivery, 25%. Now
284. 10:05here's the piece that Emergence Capital
285. 10:07identified that I think is the most
286. 10:09underrated part of this entire job. Most
287. 10:11companies think they've documented their
288. 10:12business logic in SOPs and process
289. 10:14guides, but those documents are, and
290. 10:17this is their phrase, not mine, static,
291. 10:19incomplete, and often wildly outdated. I
292. 10:23mean, think about what that means for an
293. 10:24AI system trying to do real work. AI
294. 10:27systems need to internalize how work
295. 10:29actually gets done, not the ideal
296. 10:30version of it. The FD's job is to
297. 10:33extract that real-world knowledge
298. 10:34directly on site by watching how live
299. 10:36production actually works. So here's
300. 10:38what we're actually doing at Clay, and I
301. 10:40want to be specific because I think
302. 10:41people underestimate how hands-on this
303. 10:43is. Basically, our GTM engineers are
304. 10:46doing five things for every customer.
305. 10:48First, they take the customer's business
306. 10:49context, their workflows, their goals,
307. 10:52and turn it into actual working Clay
308. 10:53builds. Second, they're building the GTM
309. 10:55infrastructure directly inside Clay.
310. 10:57Outbound sequences, enrichment
311. 10:59workflows, lead scoring, agentic-powered
312. 11:02research flows, all of it running inside
313. 11:04the product. Third, they integrate Clay
314. 11:06with everything the customer is already
315. 11:07running. Their CRM, their internal
316. 11:09tooling, their data warehouse, whatever.
317. 11:12Clay has it fit into how the company
318. 11:13actually operates, not the other way
319. 11:15around. Fourth, and this one is
320. 11:17underrated, they bring all the field
321. 11:19intelligence back to our product team.
322. 11:21Every deployment is a proving ground.
323. 11:23When something breaks or a customer
324. 11:24finds a workaround we never built for,
325. 11:26that's a product insight. It feeds
326. 11:28directly back into what we ship next.
327. 11:30And fifth, they're building agents and
328. 11:32tuning prompts, writing the prompts that
329. 11:34make Clay's AI columns actually perform,
330. 11:36debugging the outputs, and iterating
331. 11:38until the results are good enough to run
332. 11:40on their own. Now, I live this myself.
333. 11:42In Q1, I was embedded with a couple of
334. 11:44customers, spending roughly an hour a
335. 11:46day with one of them for several weeks
336. 11:47straight. And by the way, that level of
337. 11:49investment changed everything. The
338. 11:51relationship got stronger, the Clay
339. 11:53deployment went deeper. I got introduced
340. 11:55to teams and stakeholders I would have
341. 11:57never reached through a normal CSM call.
342. 11:59And we built a whole road map of use
343. 12:01cases we never would have found
344. 12:03otherwise. That's the model, and it's
345. 12:05the reason we're betting on it at Clay.
346. 12:07Now, let me zoom back out for a second,
347. 12:09because I want to leave you with what I
348. 12:10think all of this actually means. Every
349. 12:12AI product is only as good as its
350. 12:14deployment, and the FDE is the role that
351. 12:16closes the gap between what AI can do
352. 12:19and what it actually does inside a real
353. 12:21company. And if you want more content
354. 12:23like this on GTM engineering, what we're
355. 12:25building at Clay, and where this whole
356. 12:26field is heading, subscribe. I also put
357. 12:28out a weekly newsletter with over 7,000
358. 12:31GTM operators, founders, and investors
359. 12:33already inside. Link is in the
360. 12:35description.