Knowledge Systems: The New GTM Stack — Jeffrey Wang, Exa

AI Engineer · 18 min · 249 sentences · from YouTube's caption track

Each timecode opens YouTube at the start of that sentence. Line anchors (#s42) are the cue ids in the WebVTT, and every line carries its start and end seconds. All transcripts has every talk, and the whole corpus as one file.

  1. 00:01[music]
  2. 00:12Hey everybody, I'm Jeff.
  3. 00:13I guess I was introduced, but I'm the co-founder of Exa, and today going to give a talk on turning go-to-market into an AI engineering problem in the spirit of this
  4. 00:22AI engineering fair.
  5. 00:24And just a quick show of hands just to like understand the audience, like raise your hand if you're a technical.
  6. 00:30Okay, great.
  7. 00:31Okay, so I kind of oriented this talk around like go-to-market as presented to to engineers.
  8. 00:38So, happy that I did that.
  9. 00:40Cool.
  10. 00:41So, first just to like ground the ground like what what Exa is cuz it's sort of relevant inside of this presentation.
  11. 00:47Exa is a Exa is a search engine for agents.
  12. 00:49Think like agents are really smart, but they don't have access to the web.
  13. 00:53We're like this web MCP web tool that agents can access.
  14. 00:55We power Cursor, we power Cognition, we power a lot of the AI ecosystem at this point.
  15. 01:00And before we start, I also just want to like talk about, you know, especially to the technical audience, like why should you even care?
  16. 01:06Like why should you care about go-to-market?
  17. 01:07I guess this audience cares about go-to-market cuz you chose to go to this go-to-market talk, but I think there's this like funny narrative right now, which is like people are like, "Oh, like product is the only thing that matters."
  18. 01:17Or "Distribution is the only thing that matters."
  19. 01:19And there's all sort of like all sorts of like Twitter flame wars like like oh, is Glean going to succeed because they're really good at distribution, but they're like what what the heck is their product?
  20. 01:27And then and other people are like, "Oh, the like the the product needs to be super good cuz agents you know, agents shop for the product, so they'll shop the for the best product."
  21. 01:35And so, my view and my experience in the last few years is that you just kind of have to do both.
  22. 01:40Like I think you have to get product right and you have to get go-to-market right.
  23. 01:44Like you got to build this thing, it's got to be good, and then you got to get it into people's hands.
  24. 01:48If you don't do both things, then you don't have a company.
  25. 01:51So, that's kind of my view on the matter.
  26. 01:54And I I say like a really funny thing also is like as a technical person when you start a company or you start some some sort of project, like very much so the bias is like, "Hey, I'm going to just build the thing.
  27. 02:06I'm going to make it really really freaking good, right?"
  28. 02:08Like that's kind of like the bias you have as like an engineer.
  29. 02:10That's the bias we had when we started X.ai and we were like honestly pretty bad at go-to-market.
  30. 02:15Like we
  31. 02:15[laughter]
  32. 02:16we were not doing enough marketing, we were not doing enough sales.
  33. 02:18Um but I think the cool thing about about um go-to-market particularly in 2026 is you can treat go-to-market like an engineering problem and particularly an AI engineering problem.
  34. 02:29And so I think that's like a super exciting thing.
  35. 02:31Like it's like more fun for engineers than ever to do go-to-market cuz you can automate things.
  36. 02:37You can you can do so much as one person and uh etc. Also um I want to make this interactive.
  37. 02:44If if anybody has questions at any point, please please ask cuz I'm aware there's a lot of talks and I don't want to bore you.
  38. 02:51Cool.
  39. 02:51So um cool.
  40. 02:53So so the hypothesis I have is if you're an engineer or or if you're anyone, you can treat go-to-market like an engineering problem.
  41. 02:58So first, I guess like what does what do go-to-market teams do?
  42. 03:01So I have like a laundry list of things here of things that go-to-market teams do, but here are a few.
  43. 03:06Like one is you got to research like your customer, right?
  44. 03:09You got to research your targets.
  45. 03:10You have to find out information about your about targets.
  46. 03:13You have to find the right people at particular companies.
  47. 03:16You have to build POCs.
  48. 03:18Um there's like a just a ton of stuff you have to do, right?
  49. 03:21Um so you know, I'm not going to I'm not going to list everything here, but like what is the grand unifying theme?
  50. 03:26Well, go-to-market is a data problem, right?
  51. 03:30So you have all you have this like entire world of uh of of what your product does and then and this entire world of like all your potential customers
  52. 03:39and you're just going to like learn and figure out what your world looks like.
  53. 03:45And so this is my this is my uh proposal.
  54. 03:47It's a data problem and we have to solve from a data perspective.
  55. 03:51Cool.
  56. 03:51So okay, so what is data that is relevant?
  57. 03:54Uh I propose that you need basically a live model of your world that agents can act on.
  58. 04:00And so, what does that mean?
  59. 04:01Okay, well, one is you have a ton of internal data, right?
  60. 04:04There's all this information that you know about your customers, about people that are at your company, uh data about how people use the product.
  61. 04:14That's like internal data that you know.
  62. 04:16And then there's all sorts of external data, right?
  63. 04:18Like there's over 60 million companies in the world, and there's like billions of people, like over a billion that are on LinkedIn, for example, and all sorts of stuff are is is like happening every day, right?
  64. 04:28Like there's all this news.
  65. 04:30And so, when you're building like this data go-to-market system, um it's important to keep in mind just like all the different sources that exist and and and uh and are available to your agents.
  66. 04:42And cool.
  67. 04:43So, I'm going to like go through, hopefully pretty fast, just all the different components of what we've built at Exa.
  68. 04:51And just for like context, uh I've been really passionate about this for a long time.
  69. 04:54So, like Exa was launched in uh the middle of 2023, and so we were post-GPT-4.
  70. 05:01And GPT-4 was really incredible, cuz it could actually, even then, even though it's way worse than like Fable or whatever, like it could actually just automate entire parts of go-to-market.
  71. 05:12And so, from the beginning, I've been thinking about our go-to-market from from a very, very uh AI agent-first perspective.
  72. 05:17And so, we're going to go over two interfaces that we have that help us, and then two agents.
  73. 05:22Cool.
  74. 05:26Cool.
  75. 05:27Okay, the first is what we call our ICP dashboard.
  76. 05:30And the ICP dashboard is a product that we have internally that answers the question, like what is our world?
  77. 05:36Like what is the world of customers and use cases that we care about?
  78. 05:41And what we actually do is we go ahead and use Exa, and again, Exa is this like uh arbitrarily powerful search engine for AIs essentially.
  79. 05:51And we just classify basically like every possible company that is inside of our total addressable market.
  80. 05:57And I kind of blurred out some of the details on like how much money we make from each category and stuff like that.
  81. 06:03But yeah, we have like categories like model providers, AI coding platforms like say Cursor, go-to-market intelligence tools.
  82. 06:09And this makes up our TAM and we have an understanding of literally like almost every company within those segments.
  83. 06:16And then for each of those companies we can deep dive, right?
  84. 06:18So here's the example of SpaceX.
  85. 06:20We can see how much annual spend we could anticipate them to have then all this like metadata about the company.
  86. 06:25So we have a list of all the companies and then a ton of data about each company.
  87. 06:29How do we do this?
  88. 06:31Again, we're able to do this because Exa is this search engine.
  89. 06:34We take the internet, we crawl it, we train we train embeddings to do web search really well.
  90. 06:39And so basically from a technical perspective you can think about Exa as like embeddings over the internet.
  91. 06:44And when you have embeddings over the internet you have this like arbitrarily powerful semantic filtering and slicing and dicing of any type of data that you want.
  92. 06:52And so we use that to generate this this like gigantic list of potential ICPs.
  93. 06:59Cool.
  94. 07:00Next, we have a tool we call Request Lens.
  95. 07:03Request Lens, what is Request Lens?
  96. 07:04Well, it's basically a system where anytime something significant happens with any of our customers, we're alerted.
  97. 07:10Someone signed up, someone used a ton of searches, someone stopped using searches, someone showed up that we really really care about.
  98. 07:17All these things are signals that we are notified about and that our team can act on.
  99. 07:28Cool.
  100. 07:29So those are the two interfaces that we have and then I'll go over two types of agents that we have.
  101. 07:33So one is coding agents.
  102. 07:36So our go-to-market team is crazy crazy crazy deep on agents.
  103. 07:42So like like our our our engineering team uses a lot of agents, but our go-to-market team is like like you could look you could look at some of their like devin spend and like other agent spend.
  104. 07:49It's really freaking high.
  105. 07:50And that's because everybody on our go-to-market team is constantly asking agents about our customers.
  106. 07:57Uh we have like like account executives that build demos for our customers.
  107. 08:02Like it's just this crazy ecosystem where we have like maybe a dozen different agents inside of our Slack and anybody can use any of them.
  108. 08:09They all have access to tons and tons of our internal data.
  109. 08:12And uh yeah.
  110. 08:13Anytime we want to dig deeper on account, anytime we want to make a demo, etc., we depend heavily on agents.
  111. 08:23Cool.
  112. 08:23And then I want to talk about another really cool agent that I'm pretty proud of.
  113. 08:25We call it Jeff Bots.
  114. 08:27Uh or I call it Jeff Bots.
  115. 08:28Uh Jeff Bots is an AI clone of myself.
  116. 08:32Uh as much as possible.
  117. 08:33So, what is it?
  118. 08:33Well, basically this winter break uh I'm sure a lot of you spent that break playing with Opus 4.5. And I was no different.
  119. 08:42So, I was in Puerto No, I was in Mexico.
  120. 08:45I was in Mexico and I would I had a week off.
  121. 08:48And so, my goal with that week and with Opus 4.5 was to uh just try to make a digital clone of myself.
  122. 08:53And so, I did things like analyze like 760 of my emails to figure out what my email voice is.
  123. 09:00Like, oh, I use 18 words on average per email and I like to end emails with best and not sincerely.
  124. 09:05Like all all that type of stuff, right?
  125. 09:06So, I made like a like a voice for myself.
  126. 09:10And then I also made a decision-making framework.
  127. 09:13So, I made like a decision-making framework where I analyzed hundreds of decisions I've made in the past.
  128. 09:19And I I analyzed them and I created evals.
  129. 09:21So, I actually created evals from those decisions and calibrated this agent system to behave like myself.
  130. 09:29And then finally I gave it like read and write access to all the data that I personally have.
  131. 09:33And there's a cool advantage to this because like I basically have access to every single system at the company uh cuz I'm in the the the nice seat of of having that and so like
  132. 09:44um yeah, this thing has access to like everything.
  133. 09:47And basically what happens is anybody at the company can use Jeffbot to create drafts of Slack messages that are basically like answers or decisions that are made.
  134. 09:56And this is a huge great thing.
  135. 09:57Like our go-to-market team uses it to like draft emails, for example.
  136. 10:03Cool.
  137. 10:04Um all right, so those those are the systems that uh that we have at at Exa.
  138. 10:10It works pretty well.
  139. 10:10Our go-to-market team is very lean, but very productive.
  140. 10:14Um and so yeah, I just want to cover like lastly just a few principles um principles I have around what it means to be an agent-first company.
  141. 10:25So firstly, to be agent-first you must be API-first, right?
  142. 10:28So like all these systems that we built, whether they were those whether it was those agents or whether it was those GUIs that we have, like if there did not exist really good APIs on top of any internal and external data
  143. 10:40we'd be we'd be out of luck, right?
  144. 10:42Like you need to create really good APIs.
  145. 10:44If you don't have really good APIs your agents are not going to be able to have data access.
  146. 10:49So you can think about this as MCP, CLI, whatever, right?
  147. 10:52Like it doesn't really matter.
  148. 10:53Uh you just need some interface that's programmatic.
  149. 10:57Secondly is like I I think there's like still this mistake in the agent world which is made that's like hey, does everything need to be a chatbot?
  150. 11:06Uh I think the answer is no. Like I think I think both GUIs and chatbots are are both super useful and have their own benefits.
  151. 11:16Like uh I don't know how many people in this room have thought about dynamic user interfaces but like yes, dynamic user interfaces are amazing.
  152. 11:24Like yes, technically AI can just produce a new UI for any use case that you have.
  153. 11:29Like just to answer a question, it could produce like an HTML markdown file, right?
  154. 11:33But I think there is something really nice about being able to visit the same consistent UX for the same use cases over time so that you can like learn how to use some tool.
  155. 11:42Um so yeah, I think like having crystallized UIs and then also arbitrarily powerful flexible chat agents are both important components of being agent first.
  156. 11:53And then finally, uh you know, there's this question like, "Hey, should you like shop for like Salesforce or should you like build your own CRM or something, right?"
  157. 12:04I actually think this is like a false dichotomy.
  158. 12:06It's like like there it's not a choice.
  159. 12:10Like we don't live in a world where the choice is between purchasing SaaS and building things yourself.
  160. 12:14Like the way I like to think about it is like you should just be using something that is arbitrarily customizable, right?
  161. 12:22Like whether you like obviously if you build something yourself, then it's arbitrarily customizable cuz you can write code and make it better at any given point.
  162. 12:29But also if you procure SaaS, um if you can make that SaaS work on your behalf and be arbitrarily customizable, then that works too, right?
  163. 12:39Like you don't need to build this like GUI and like have a proactive roadmap as to like what features would make really great sense inside of some system.
  164. 12:47Like if you can arbitrarily customize the system, even if it's a system you've purchased, then you're like pretty good, right?
  165. 12:52So like for example, we use Salesforce.
  166. 12:54Like we use Salesforce at X.ai and uh it's great because uh it's a really good good database.
  167. 12:59It's made a lot of amazing choices around what sales should look like, choices that we don't want to make ourselves.
  168. 13:05And then it exposes MCP.
  169. 13:07So all of our agents have access to Salesforce MCP.
  170. 13:09Works really well.
  171. 13:10Our team uses it every day.
  172. 13:11And so yeah, I think infinite customizability um is is really the highest order bit.
  173. 13:19Cool.
  174. 13:20Um that's that's all I had.
  175. 13:23Uh Yeah, does anyone have any questions?
  176. 13:26Oh, we have time for a few questions.
  177. 13:28Okay, coming.
  178. 13:36Hey, um so you said you uh took all your past decisions.
  179. 13:40Can you elaborate a bit about that?
  180. 13:43What What artifacts are those?
  181. 13:45Usually people don't save like their decisions.
  182. 13:47Is it Slack?
  183. 13:48Is it email?
  184. 13:49Is it other other artifacts?
  185. 13:52Yeah, good question.
  186. 13:52I looked at decisions I made within Slack and email.
  187. 13:56I mean, a surprisingly large amount of everything that goes on a company is is is on Slack, right?
  188. 14:02So like if you just read like a ton of Slack history, like you can definitely find hundreds of decisions that you made in the past.
  189. 14:09Awesome.
  190. 14:10Quick question uh over here.
  191. 14:11Uh so your go-to-market team, what's the split between uh are they just all like AI cracked or do they also have like the domain expertise, too?
  192. 14:21What's the split between technical and non-technical?
  193. 14:24Because obviously you need them to like know how to do marketing, sales, etc. But then do they also are they also upskilling in terms of using AI systems?
  194. 14:31Are you handing them tools or they building their own?
  195. 14:35That's a very good question.
  196. 14:36So our go-to-market team is comprised of like there's there's account executives which like run the deals.
  197. 14:45There are like sales like SDRs that help with uh demand generation.
  198. 14:51And then there are separately they're separate like a FDE org.
  199. 14:55So forward deployed engineering organization.
  200. 14:58And what I'll say is that like everyone that's Okay, everyone that's uh not not in FDE is like has learned how to use AI really well.
  201. 15:11So like the answer is like they're not vibe coding.
  202. 15:14They're not generally with you know, in some there's some exceptions.
  203. 15:17They're not generally vibe coding these interfaces that we have.
  204. 15:20Um but they're using the tools really really well.
  205. 15:22And like we make sure that we have training sessions and like just make sure that people really understand how to use these tools.
  206. 15:28And then this funny we have this funny thing which is like our four deployed engineering organization is actually the one that like does a lot of the maintenance and feature building
  207. 15:37uh of these AI systems.
  208. 15:38And so they're both running deals and like like supporting deals um but then also making everything smoother by like doing sales but then also building the sales system.
  209. 15:53Like it's it's kind of it's kind of a funky thing we have going on.
  210. 15:55Yeah.
  211. 15:58How do you think about uh different security?
  212. 16:02Oh.
  213. 16:02Hey.
  214. 16:03How do you think about different uh security boundaries within your enterprise?
  215. 16:06What do you What you said suggested that you've got Jeffbot which had runs with all of your full privileges and then it's available to everybody which suggests that there's one security level and everyone can see everything all the time.
  216. 16:17Is that what you're going with or is there some uh other guardrails in place?
  217. 16:22Yeah, that's a good question.
  218. 16:23We we pay pretty special we we pay pretty careful attention to guardrails.
  219. 16:27So for example um in the case of Jeffbot um when I use Jeffbot and I call Jeffbot has access to a ton of systems and it can for example do reads and writes.
  220. 16:39However, when anybody else calls Jeffbot all can do is draft messages and also I don't give Jeffbot permissions to all of our MCPs and tools in the case where other people call it.
  221. 16:51And so in short it's like pretty it's pretty well defined or we we we do pay some care to the security.
  222. 16:58Yeah.
  223. 16:59Okay, last question.
  224. 17:04Um, can you can you share the origin story of the FDE team?
  225. 17:09Did that just happen organically or did you intentionally do it?
  226. 17:12I'm just really curious like how that came to exist.
  227. 17:16Yeah, for sure.
  228. 17:17I mean, uh my my philos- my my hypothesis on this is like, once upon a time the FDE role didn't really exist.
  229. 17:25Like, Palantir started calling some people FDEs, but it that was really it.
  230. 17:29And what tech companies had was like solutions and sales engineers.
  231. 17:34And then, like account executives.
  232. 17:37I was a solutions engineer.
  233. 17:38Got it.
  234. 17:39Yeah, yeah.
  235. 17:39The thing The thing that I think has changed is that um because of AI, as like en- as a technical person that is supporting revenue generation, you can actually not only support the revenue generation, but then very easily build the tooling
  236. 17:55to smooth everything over.
  237. 17:58And make your own life easier, make the lives of AEs easier.
  238. 18:01Like, because of AI, this is just possible now.
  239. 18:03Like, that's like two Before that was like two jobs, and now it's like one job.
  240. 18:07Um in theory.
  241. 18:08Like, now when our team grows, like right now it's about eight or nine FDEs, like what will will it scale such that everyone does everything?
  242. 18:15Probably not.
  243. 18:15But, at least right now that's what we have, and I think that's a really good working model to get pretty far.
  244. 18:20Eight out of how many?
  245. 18:22Uh eight Oh, eight of like how big is our go-to-market org?
  246. 18:24Or eight You have eight FDEs, and the size of the company right now is how many?
  247. 18:28Oh, the We're about 115 people.
  248. 18:31Okay.
  249. 18:31Yeah.