AI Engineer World's Fair 2026

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

Jeffrey Wang· exa18:49

Read the talk

Knowledge Systems: The New GTM Stack

Jeffrey Wang explains how Exa turns market discovery, customer activity, company knowledge, and past decisions into a go-to-market system that people and agents can use—then examines the interfaces, permissions, and organizational roles needed to operate it.

From a talk by Jeffrey Wang

At a glance

Ideas worth remembering

  • Go-to-market can be modeled as a data system connecting product knowledge, internal customer behavior, and changing external market information.

  • Exa pairs stable interfaces with flexible agents: the ICP dashboard maps potential customers, Request Lens exposes behavioral signals, and Slack agents investigate accounts and build demos.

  • Jeffbot treats personal replication as three problems: writing style from 760 emails, judgment calibration from hundreds of past decisions, and controlled access to company systems.

  • Agent-first systems need programmatic data access, but stable graphical interfaces remain useful for repeated tasks that benefit from a familiar user experience.

  • Buy versus build is better evaluated through customizability: Exa retains Salesforce’s existing sales model while giving agents programmatic access through MCP.

  • Agent permissions should depend on the caller. Wang’s Jeffbot invocation can read and write broadly, while other employees receive drafting capabilities and a reduced tool set.

  • FDEs can shorten the loop between customer work and internal tooling, but Wang expects specialization may become necessary as Exa grows beyond its current model.

Treat go-to-market as a live data problem

Engineers naturally want to keep improving the product. Wang says Exa followed that instinct early on and consequently did too little marketing and sales. His correction is blunt: a company must both build something good and get it into people’s hands. AI makes the second job more approachable to engineers because research, account preparation, and other go-to-market work can now be designed as software systems rather than handled only through manual processes.

Introduces the live model of the world that agents can act on, including internal business data.
Introduces the live model of the world that agents can act on, including internal business data.

Customer research, target identification, contact discovery, and proofs of concept look like separate sales activities, but they share a dependency: the team needs information about what the product does and which prospective customers might need it. Wang therefore frames go-to-market as a data problem. The goal is to maintain a live model of the company’s world that agents can act on, rather than repeatedly assembling account context from scratch.

That model combines two broad classes of information:

  • Internal data: customer records, employee knowledge, and product-usage behavior.
  • External data: companies, people, news, and other events occurring outside the business.

Wang cites more than 60 million companies worldwide and more than a billion people on LinkedIn to convey the scale of the external side. These figures motivate broad discovery infrastructure; they do not establish that Exa’s internal model covers every company or person.

Exa began approaching its own go-to-market operation from an agent-first perspective after launching in the middle of 2023. Wang organizes the resulting system around two interfaces and two kinds of agents. The interfaces make recurring information easy to inspect; the agents handle less predictable investigation and production work.

How it fits togetherFrom scattered data to go-to-market action

Customer knowledge, company information, and product usage.

Internal and external information feed a live view of the market and customers. Interfaces support repeated inspection, while agents answer flexible questions and help produce account-specific work.

1:121:42
Suggest correction

This note stays in this page until you copy or download it. Nothing is submitted; reloading clears the draft.

0:12 · section reference included

Map the addressable market with the ICP dashboard

The first interface is Exa’s internal ICP dashboard. It answers a foundational question: which customers and use cases make up the company’s world? Exa uses its search system to classify almost every company it identifies within its total addressable market. Wang names model providers, AI coding platforms, and go-to-market intelligence tools as example segments.

Supports the technical explanation of Exa as semantic search built from embeddings over web data.
Supports the technical explanation of Exa as semantic search built from embeddings over web data.

The dashboard supports both breadth and depth. At the top level, it groups potential customers into market segments. At the company level, it exposes metadata and an estimate of anticipated annual spend; Wang uses SpaceX as the example. He intentionally hides some category revenue information. The talk also supplies no calculation method, error rate, refresh schedule, or coverage audit, so anticipated spend should be understood as a planning estimate rather than a verified forecast.

The discovery mechanism begins with Exa crawling the web and training embeddings for web search. Wang’s shorthand is “embeddings over the internet”: web information is represented so the system can retrieve and group it by semantic meaning rather than exact keyword overlap alone. Exa uses that semantic filtering to generate a large candidate list for the dashboard. The talk explains this retrieval layer but does not specify the downstream classifier, validation process, or rules that decide whether a company belongs in the market.

5:275:57
Suggest correction

This note stays in this page until you copy or download it. Nothing is submitted; reloading clears the draft.

5:27 · section reference included

Turn customer changes into signals, then investigate in Slack

A market map changes slowly; customer behavior can change by the minute. Request Lens handles that second timescale. It alerts the team when someone signs up, runs many searches, stops searching, or appears from an account Exa especially cares about. These events turn raw activity into prompts for attention. Wang says the team can act on them, but he does not describe the significance thresholds or claim that an alert automatically triggers outreach.

Captures the discussion of roughly a dozen Slack agents used to investigate accounts and build demos.
Captures the discussion of roughly a dozen Slack agents used to investigate accounts and build demos.

The go-to-market team then uses perhaps a dozen agents inside Slack to investigate accounts and produce customer-specific work. These agents can access substantial internal data. Account executives use them to build demos, and the wider team relies on them when it needs a deeper account analysis. The important operating choice is placement: the tools live in the collaboration environment the team already uses, rather than requiring every salesperson to operate a separate developer interface.

Wang describes agent spending by the go-to-market team as unusually high, even relative to engineering, but gives neither a dollar amount nor a measured return. The stronger evidence in the talk is behavioral: agents are routinely used for account questions and demo construction. Productivity is asserted rather than quantified.

6:597:28
Suggest correction

This note stays in this page until you copy or download it. Nothing is submitted; reloading clears the draft.

6:59 · section reference included

Separate a personal agent’s voice, judgment, and access

Wang’s most personal example is Jeffbot, an attempt to build a digital version of himself during a week off in Mexico. The system has three distinct ingredients: a writing profile, a decision framework, and access to company systems. Keeping those ingredients separate matters because sounding like someone, deciding like them, and acting with their privileges are different engineering problems.

Introduces Jeffbot as Wang’s experiment in building a digital clone of himself.
Introduces Jeffbot as Wang’s experiment in building a digital clone of himself.

For voice, Wang analyzed 760 of his emails. The resulting profile included concrete habits: his emails averaged 18 words, and he tended to close with “Best” rather than “Sincerely.” This replaces a vague instruction to imitate his tone with observations derived from his actual writing.

For judgment, he analyzed hundreds of previous decisions and converted them into evaluations. Those examples became a calibration target: given comparable situations, the system should produce decisions resembling his past choices. The source material came from ordinary work records rather than a formal decision journal; in the Q&A, Wang identifies Slack and email as the artifacts containing those decisions.

The talk does not report evaluation scores, holdout methodology, grader design, or failure categories. Jeffbot therefore demonstrates a useful construction pattern—derive style from communications and behavioral tests from past decisions—but does not establish how faithfully it reproduces Wang’s judgment, especially in new situations unlike those records.

The final ingredient is tool and data access. Wang can invoke Jeffbot with broad read and write access to systems he can personally use. Other employees use it to create drafts of Slack messages, answers, decisions, and emails. That distinction becomes crucial when the audience later asks whether sharing the assistant also shares the cofounder’s privileges.

8:238:52
Suggest correction

This note stays in this page until you copy or download it. Nothing is submitted; reloading clears the draft.

8:23 · section reference included

Build programmatic access without turning everything into chat

Wang extracts three design principles from these systems. First, agent-first requires API-first. Dashboards and chat agents can only use internal and external information if that information has a programmatic interface. He treats MCP, a CLI, and an API as possible implementations of the same requirement: agents need a callable route to the data and operations involved in the work.

Introduces Wang’s claim that buy versus build is a false dichotomy.
Introduces Wang’s claim that buy versus build is a false dichotomy.

Second, agent-first does not mean chatbot-only. A model can generate a fresh interface for a one-off request, but repeated tasks benefit from a stable interface whose layout and behavior people can learn. Exa’s design keeps both:

  • Crystallized interfaces for recurring views such as the ICP dashboard and Request Lens.
  • Flexible chat agents for account questions and tasks the interface did not anticipate.

The shared programmatic data layer supports both forms rather than forcing the team to choose one.

Third, Wang reframes buy versus build around customizability. Software built internally can be changed directly, but purchased software can also work if it exposes enough of its data and behavior for the company’s agents to adapt it. The criterion is whether the system can participate in the company’s workflows, not whether the company wrote every screen itself.

Salesforce is his example. Exa keeps the database and sales conventions Salesforce already provides instead of recreating those choices, then lets agents access it through MCP. Wang says the team uses that connection daily. This is a pragmatic middle path: buy an established system of record, then build agent-driven behavior around its programmatic interface. It does not mean every SaaS product is infinitely adaptable; the approach depends on what the purchased system actually exposes.

10:0310:33
Suggest correction

This note stays in this page until you copy or download it. Nothing is submitted; reloading clears the draft.

10:03 · section reference included

Give tool users training—and give agents caller-dependent permissions

The Q&A clarifies that Exa does not expect every salesperson to build these systems. Account executives run deals, SDRs support demand generation, and a separate forward deployed engineering organization builds and maintains much of the tooling. Non-FDE staff generally use rather than create the interfaces, with training sessions helping them become proficient with the available AI tools.

Forward deployed engineers connect both sides of the loop. They run or support deals, encounter friction in the sales process, and then add features or maintain the systems used to do that work. The people closest to customer-specific failures can improve the shared tooling rather than sending every problem through a distant product queue.

An audience member then identifies the dangerous implication of Jeffbot’s broad access: if everyone can call an assistant carrying Wang’s permissions, has the company effectively collapsed into one security level? Wang’s answer is that permissions depend on the caller. When he invokes Jeffbot, it can read and write across many systems. When another employee invokes it, the system can draft messages but does not receive all the MCP and tool permissions available in Wang’s own use.

This establishes two separate controls: tool availability and allowed actions. Restricting both is stronger than merely telling the model not to use a capability. The talk remains at the policy level, however; it does not explain identity verification, authorization enforcement, audit logging, or which data may enter an employee-visible draft.

Compare the ideasJeffbot’s capability depends on who invokes it

The invocation can access many systems and perform reads and writes.

The shared assistant does not run with one universal permission set. Wang describes broader capabilities for his own calls and a drafting-only path with fewer tools for other employees.

14:0914:35
Suggest correction

This note stays in this page until you copy or download it. Nothing is submitted; reloading clears the draft.

13:36 · section reference included

Combine deal support and system building—until scale forces specialization

The final question asks how Exa’s FDE organization emerged. Wang contrasts the role with the older separation among solutions engineers, sales engineers, and account executives. His hypothesis is that AI lets a technical person supporting revenue also build the tools that make that support easier. Work that previously resembled two jobs—helping a deal and improving the sales system—can temporarily fit into one role.

That arrangement creates a short feedback loop: support a customer, observe repetitive work, build an improvement, and make the next deal easier for both the FDE and the account executive. It also explains why Exa assigns maintenance of its go-to-market AI systems to people participating directly in customer work.

Wang closes with an important limit. Exa has about eight or nine FDEs in a company of about 115 people, and he doubts the model will scale indefinitely with everyone doing everything. At the company’s current size, he considers the combined role a productive way to get far. Growth may eventually require more specialization. The organizational lesson is therefore a working experiment, not a universal end state.

17:0417:16
Suggest correction

This note stays in this page until you copy or download it. Nothing is submitted; reloading clears the draft.

17:04 · section reference included

Resources

From the talk

  • AI EngineerReference

    Conference information and the broader catalog of talks on agents, MCP, evaluation, interfaces, and AI engineering.

  • Agents for Everything Else — swyx

    A complementary account of using coding agents for design implementation, data synchronization, research, and other knowledge work outside conventional software development.

  • Forward Deployed Engineering 101

    Expands the organizational model Wang introduces at the end, including how FDE work differs from bespoke services and conventional sales engineering.

  • AI in GTM at Notion — Flora Liu

    Offers another go-to-market architecture built around shared customer context, behavioral signals, agent drafting, human approval, and outcome feedback.

Read the complete timestamped transcript
  1. 0:01

    [music]

  2. 0:12

    >> Hey everybody, I'm Jeff. I guess I was

  3. 0:13

    introduced, but I'm the co-founder of

  4. 0:15

    Exa, and today going to give a talk on

  5. 0:19

    turning go-to-market into an AI

  6. 0:20

    engineering problem in the spirit of

  7. 0:21

    this

  8. 0:22

    AI engineering fair. And just a quick

  9. 0:24

    show of hands just to like understand

  10. 0:26

    the audience, like raise your hand if

  11. 0:27

    you're a technical.

  12. 0:30

    Okay, great. Okay, so I kind of

  13. 0:32

    oriented this talk around like

  14. 0:34

    go-to-market

  15. 0:36

    as presented to to engineers. So, happy

  16. 0:39

    that I did that.

  17. 0:40

    Cool. So, first just to like ground the

  18. 0:43

    ground like what what Exa is cuz it's

  19. 0:44

    sort of relevant inside of this

  20. 0:46

    presentation. Exa is a Exa is a search

  21. 0:48

    engine for agents.

  22. 0:49

    Think like agents are really smart, but

  23. 0:52

    they don't have access to the web. We're

  24. 0:53

    like this web MCP web tool that agents

  25. 0:55

    can access. We power Cursor, we power

  26. 0:56

    Cognition, we power a lot of the AI

  27. 0:58

    ecosystem at this point.

  28. 1:00

    And

  29. 1:01

    before we start, I also just want to

  30. 1:02

    like talk about, you know, especially to

  31. 1:03

    the technical audience, like why should

  32. 1:05

    you even care?

  33. 1:06

    Like why should you care about

  34. 1:07

    go-to-market? I guess this audience

  35. 1:08

    cares about go-to-market cuz you chose

  36. 1:09

    to go to this

  37. 1:11

    go-to-market talk, but I think there's

  38. 1:12

    this like funny narrative right now,

  39. 1:14

    which is like people are like, "Oh, like

  40. 1:15

    product is the only thing that matters."

  41. 1:17

    Or "Distribution is the only thing that

  42. 1:19

    matters." And there's all sort of like

  43. 1:20

    all sorts of like Twitter flame wars

  44. 1:22

    like like oh, is Glean going to succeed

  45. 1:24

    because they're really good at

  46. 1:25

    distribution, but they're like what what

  47. 1:26

    the heck is their product? And then and

  48. 1:28

    other people are like, "Oh, the like the

  49. 1:29

    the product needs to be super good cuz

  50. 1:31

    agents

  51. 1:32

    you know, agents shop for the product,

  52. 1:33

    so they'll shop the for the best

  53. 1:34

    product." And so, my view and my

  54. 1:36

    experience in the last few years is that

  55. 1:39

    you just kind of have to do both. Like I

  56. 1:41

    think you have to get product right and

  57. 1:42

    you have to get go-to-market right. Like

  58. 1:44

    you got to build this thing, it's got to

  59. 1:45

    be good, and then you got to get it into

  60. 1:47

    people's hands. If you don't do both

  61. 1:49

    things, then you don't have a company.

  62. 1:51

    So, that's kind of my view on the

  63. 1:53

    matter. And I I say like a really funny

  64. 1:56

    thing also is like as a technical person

  65. 1:58

    when you start a company or you start

  66. 2:00

    some some sort of project, like very

  67. 2:02

    much so the bias is like, "Hey, I'm

  68. 2:04

    going to just build the thing. I'm going

  69. 2:06

    to make it really really freaking good,

  70. 2:08

    right?" Like that's kind of like the

  71. 2:09

    bias you have as like an engineer.

  72. 2:10

    That's the bias we had when we started

  73. 2:12

    X.ai and we were like honestly pretty

  74. 2:14

    bad at go-to-market. Like we

  75. 2:15

    >> [laughter]

  76. 2:16

    >> we were not doing enough marketing, we

  77. 2:17

    were not doing enough sales.

  78. 2:18

    Um but I think the cool thing about

  79. 2:21

    about um go-to-market particularly in

  80. 2:23

    2026 is you can treat go-to-market like

  81. 2:26

    an engineering problem and particularly

  82. 2:28

    an AI engineering problem. And so I

  83. 2:30

    think that's like a super exciting

  84. 2:31

    thing. Like it's like more fun for

  85. 2:33

    engineers than ever to do go-to-market

  86. 2:36

    cuz you can automate things. You can you

  87. 2:38

    can do so much as one person and uh

  88. 2:41

    etc. Also um I want to make this

  89. 2:44

    interactive. If if anybody has questions

  90. 2:45

    at any point, please please ask cuz I'm

  91. 2:47

    aware there's a lot of talks and

  92. 2:49

    I don't want to bore you.

  93. 2:51

    Cool. So um cool. So so the hypothesis I

  94. 2:54

    have is if you're an engineer or or if

  95. 2:56

    you're anyone, you can treat

  96. 2:57

    go-to-market like an engineering

  97. 2:58

    problem. So first, I guess like what

  98. 2:59

    does what do go-to-market teams do? So I

  99. 3:01

    have like a laundry list of things here

  100. 3:04

    of things that go-to-market teams do,

  101. 3:05

    but here are a few. Like one is you got

  102. 3:06

    to research

  103. 3:08

    like your customer, right? You got to

  104. 3:09

    research your targets. You have to find

  105. 3:11

    out information about your about

  106. 3:13

    targets. You have to find the right

  107. 3:15

    people at particular companies. You have

  108. 3:17

    to build POCs. Um there's like a just a

  109. 3:19

    ton of stuff you have to do, right? Um

  110. 3:22

    so you know, I'm not going to I'm not

  111. 3:24

    going to list everything here, but like

  112. 3:25

    what is the grand unifying theme? Well,

  113. 3:27

    go-to-market is a data problem, right?

  114. 3:30

    So you have all you have this like

  115. 3:31

    entire world of uh of of what your

  116. 3:35

    product does and then and this entire

  117. 3:37

    world of like all your potential

  118. 3:38

    customers and you're just going to like

  119. 3:40

    learn and figure out what your world

  120. 3:44

    looks like. And so this is my this is my

  121. 3:46

    uh proposal. It's a data problem and we

  122. 3:48

    have to solve from a data perspective.

  123. 3:51

    Cool. So okay, so what is data that is

  124. 3:53

    relevant? Uh I propose that you need

  125. 3:56

    basically a live model of your world

  126. 3:58

    that agents can act on. And so, what

  127. 4:01

    does that mean? Okay, well, one is you

  128. 4:02

    have a ton of internal data, right?

  129. 4:04

    There's all this information that you

  130. 4:05

    know about your customers, about people

  131. 4:08

    that are at your company,

  132. 4:11

    uh

  133. 4:12

    data about how people use the product.

  134. 4:14

    That's like internal data that you know.

  135. 4:16

    And then there's all sorts of external

  136. 4:17

    data, right? Like there's over 60

  137. 4:19

    million companies in the world, and

  138. 4:21

    there's like billions of people, like

  139. 4:23

    over a billion that are on LinkedIn, for

  140. 4:24

    example, and all sorts of stuff are is

  141. 4:27

    is like happening every day, right? Like

  142. 4:28

    there's all this news. And so, when

  143. 4:31

    you're building like this data

  144. 4:33

    go-to-market system, um it's important

  145. 4:35

    to keep in mind just like all the

  146. 4:37

    different sources that exist and and and

  147. 4:39

    uh

  148. 4:40

    and are available to your agents.

  149. 4:42

    And cool. So, I'm going to like go

  150. 4:44

    through, hopefully pretty fast, just all

  151. 4:47

    the different components of what we've

  152. 4:49

    built at Exa. And just for like context,

  153. 4:53

    uh I've been really passionate about

  154. 4:54

    this for a long time. So, like Exa was

  155. 4:56

    launched in

  156. 4:58

    uh the middle of 2023, and so we were

  157. 5:00

    post-GPT-4. And

  158. 5:02

    GPT-4 was really incredible, cuz it

  159. 5:04

    could actually, even then, even though

  160. 5:06

    it's way worse than like Fable or

  161. 5:08

    whatever, like it could actually just

  162. 5:09

    automate entire

  163. 5:11

    parts of go-to-market. And so, from the

  164. 5:12

    beginning, I've been thinking about our

  165. 5:13

    go-to-market from

  166. 5:14

    from a very, very uh AI agent-first

  167. 5:16

    perspective. And so, we're going to go

  168. 5:18

    over two interfaces that we have that

  169. 5:19

    help us, and then two agents.

  170. 5:22

    Cool.

  171. 5:26

    Cool. Okay, the first is what we call

  172. 5:28

    our ICP dashboard. And the ICP dashboard

  173. 5:32

    is a product that we have internally

  174. 5:33

    that answers the question, like what is

  175. 5:35

    our world? Like what is the world of

  176. 5:38

    customers and use cases that we care

  177. 5:40

    about? And what we actually do is we go

  178. 5:44

    ahead and use Exa, and again, Exa is

  179. 5:47

    this like uh

  180. 5:49

    arbitrarily powerful search engine for

  181. 5:50

    AIs essentially. And we just classify

  182. 5:53

    basically like every possible company

  183. 5:54

    that is inside of our total addressable

  184. 5:56

    market. And I kind of blurred out some

  185. 5:59

    of the details on like how much money we

  186. 6:00

    make from each category and stuff like

  187. 6:03

    that. But yeah, we have like categories

  188. 6:04

    like model providers, AI coding

  189. 6:06

    platforms like say Cursor, go-to-market

  190. 6:07

    intelligence tools.

  191. 6:09

    And this makes up our TAM and we have an

  192. 6:11

    understanding of literally like

  193. 6:13

    almost every company within those

  194. 6:14

    segments.

  195. 6:16

    And then for each of those companies we

  196. 6:18

    can deep dive, right? So here's the

  197. 6:19

    example of SpaceX. We can see how much

  198. 6:22

    annual spend we could anticipate them to

  199. 6:23

    have then all this like metadata about

  200. 6:25

    the company. So we have a list of all

  201. 6:26

    the companies and then a ton of data

  202. 6:28

    about each company.

  203. 6:29

    How do we do this? Again, we're able to

  204. 6:32

    do this because Exa is this search

  205. 6:33

    engine. We take the internet, we crawl

  206. 6:35

    it, we train we train embeddings to do

  207. 6:38

    web search really well. And so basically

  208. 6:40

    from a technical perspective you can

  209. 6:41

    think about Exa as like embeddings over

  210. 6:43

    the internet. And when you have

  211. 6:45

    embeddings over the internet you have

  212. 6:46

    this like arbitrarily powerful semantic

  213. 6:49

    filtering and slicing and dicing of any

  214. 6:51

    type of data that you want. And so we

  215. 6:53

    use that to generate this this like

  216. 6:54

    gigantic list of potential ICPs.

  217. 6:59

    Cool. Next, we have a tool we call

  218. 7:02

    Request Lens. Request Lens, what is

  219. 7:04

    Request Lens? Well, it's basically a

  220. 7:05

    system where anytime something

  221. 7:07

    significant happens with any of our

  222. 7:08

    customers, we're alerted.

  223. 7:10

    Someone signed up, someone used a ton of

  224. 7:13

    searches, someone stopped using

  225. 7:14

    searches, someone showed up that we

  226. 7:16

    really really care about. All these

  227. 7:18

    things are signals that we are notified

  228. 7:20

    about and that our team can act on.

  229. 7:28

    Cool.

  230. 7:29

    So those are the two interfaces that we

  231. 7:31

    have and then I'll go over two types of

  232. 7:33

    agents that we have. So

  233. 7:34

    one is coding agents. So our

  234. 7:37

    go-to-market team

  235. 7:38

    is crazy crazy crazy deep on agents. So

  236. 7:42

    like like our our our engineering team

  237. 7:44

    uses a lot of agents, but our

  238. 7:45

    go-to-market team is like like you could

  239. 7:46

    look you could look at some of their

  240. 7:47

    like devin spend and like other agent

  241. 7:49

    spend. It's really freaking high. And

  242. 7:50

    that's because everybody on our

  243. 7:52

    go-to-market team is constantly asking

  244. 7:54

    agents about our customers.

  245. 7:57

    Uh we have like

  246. 7:58

    like account executives that build demos

  247. 8:01

    for our customers. Like it's just this

  248. 8:03

    crazy ecosystem where we have like maybe

  249. 8:05

    a dozen different agents inside of our

  250. 8:07

    Slack and anybody can use any of them.

  251. 8:09

    They all have access to tons and tons of

  252. 8:10

    our internal data. And uh yeah. Anytime

  253. 8:13

    we want to dig deeper on account,

  254. 8:15

    anytime we want to make a demo, etc., we

  255. 8:17

    depend heavily on agents.

  256. 8:23

    Cool. And then I want to talk about

  257. 8:24

    another really cool agent that I'm

  258. 8:25

    pretty proud of. We call it Jeff Bots.

  259. 8:27

    Uh or I call it Jeff Bots. Uh Jeff Bots

  260. 8:29

    is an AI clone of myself. Uh as much as

  261. 8:32

    possible. So, what is it? Well,

  262. 8:34

    basically

  263. 8:35

    this winter break uh

  264. 8:38

    I'm sure a lot of you spent that break

  265. 8:39

    playing with Opus 4.5. And I was no

  266. 8:42

    different. So, I was in Puerto No, I was

  267. 8:44

    in Mexico. I was in Mexico and I would I

  268. 8:46

    had a week off. And so, my goal with

  269. 8:49

    that week and with Opus 4.5 was to uh

  270. 8:52

    just try to make a digital clone of

  271. 8:53

    myself. And so, I did things like

  272. 8:54

    analyze

  273. 8:55

    like 760 of my emails to figure out what

  274. 8:59

    my email voice is. Like, oh, I use 18

  275. 9:01

    words on average per email and I like to

  276. 9:03

    end emails with best and not sincerely.

  277. 9:05

    Like all all that type of stuff, right?

  278. 9:06

    So, I made like a like a voice for

  279. 9:09

    myself.

  280. 9:10

    And then I also made a decision-making

  281. 9:13

    framework. So, I made like a

  282. 9:14

    decision-making framework where I

  283. 9:16

    analyzed hundreds of decisions I've made

  284. 9:17

    in the past.

  285. 9:19

    And I I analyzed them and I created

  286. 9:21

    evals. So, I actually created evals from

  287. 9:23

    those decisions and calibrated this

  288. 9:25

    agent system to behave like myself.

  289. 9:29

    And then finally I gave it like read and

  290. 9:30

    write access to all the data that I

  291. 9:32

    personally have. And there's a cool

  292. 9:34

    advantage to this because like I

  293. 9:35

    basically have access to every single

  294. 9:37

    system at the company

  295. 9:39

    uh cuz I'm in the the the nice seat

  296. 9:41

    of of having that and so like um yeah,

  297. 9:45

    this thing has access to like

  298. 9:46

    everything. And basically what happens

  299. 9:48

    is anybody at the company can use

  300. 9:49

    Jeffbot to create drafts of Slack

  301. 9:52

    messages that are basically like answers

  302. 9:54

    or decisions that are made. And this is

  303. 9:56

    a huge great thing. Like our

  304. 9:57

    go-to-market team uses it to like draft

  305. 9:59

    emails, for example.

  306. 10:03

    Cool. Um

  307. 10:05

    all right, so those those are the

  308. 10:06

    systems that uh

  309. 10:08

    that we have at at Exa. It works pretty

  310. 10:10

    well. Our go-to-market team is very

  311. 10:11

    lean, but very productive.

  312. 10:14

    Um and so yeah, I just want to cover

  313. 10:15

    like

  314. 10:16

    lastly just a few principles

  315. 10:19

    um

  316. 10:20

    principles I have

  317. 10:22

    around what it means to be an

  318. 10:23

    agent-first company.

  319. 10:25

    So firstly, to be agent-first you must

  320. 10:27

    be API-first, right? So like all these

  321. 10:29

    systems that we built, whether they were

  322. 10:30

    those whether it was those agents or

  323. 10:32

    whether it was those GUIs that we have,

  324. 10:35

    like if there did not exist really good

  325. 10:37

    APIs on top of any internal and external

  326. 10:39

    data

  327. 10:40

    we'd be we'd be out of luck, right? Like

  328. 10:42

    you need to create really good APIs. If

  329. 10:44

    you don't have really good APIs your

  330. 10:46

    agents are not going to be able to have

  331. 10:48

    data access. So you can think about this

  332. 10:50

    as MCP, CLI, whatever, right? Like it

  333. 10:52

    doesn't really matter. Uh you just need

  334. 10:53

    some interface that's programmatic.

  335. 10:57

    Secondly is like I I think there's like

  336. 10:58

    still this mistake in

  337. 11:00

    the agent world which is made that's

  338. 11:02

    like hey, does everything

  339. 11:04

    need to be a chatbot?

  340. 11:06

    Uh

  341. 11:07

    I think the answer is no. Like I think

  342. 11:09

    I think both GUIs and chatbots are are

  343. 11:12

    both super useful and have their own

  344. 11:16

    benefits. Like uh I don't know how many

  345. 11:19

    people in this room have thought about

  346. 11:19

    dynamic user interfaces

  347. 11:21

    but like yes, dynamic user interfaces

  348. 11:24

    are amazing. Like yes, technically AI

  349. 11:26

    can just produce a new UI for any use

  350. 11:28

    case that you have. Like

  351. 11:30

    just to answer a question, it could

  352. 11:31

    produce like an HTML markdown file,

  353. 11:32

    right? But I think there is something

  354. 11:34

    really nice about being able to visit

  355. 11:36

    the same consistent UX for the same use

  356. 11:39

    cases over time so that you can like

  357. 11:41

    learn how to use some tool. Um so yeah,

  358. 11:43

    I think like having crystallized UIs and

  359. 11:45

    then also arbitrarily powerful flexible

  360. 11:47

    chat agents are both important

  361. 11:49

    components of being agent first.

  362. 11:53

    And then finally,

  363. 11:55

    uh

  364. 11:56

    you know, there's this question like,

  365. 11:57

    "Hey, should you like shop for like

  366. 11:59

    Salesforce

  367. 12:00

    or should you like build your own CRM or

  368. 12:03

    something, right?" I actually think this

  369. 12:05

    is like a false dichotomy. It's like

  370. 12:08

    like there it's not a choice. Like we

  371. 12:10

    don't live in a world where the choice

  372. 12:11

    is between purchasing SaaS and building

  373. 12:13

    things yourself. Like the way I like to

  374. 12:15

    think about it is like

  375. 12:18

    you should just be using something that

  376. 12:20

    is arbitrarily customizable, right? Like

  377. 12:22

    whether you like obviously if you build

  378. 12:24

    something yourself, then it's

  379. 12:25

    arbitrarily customizable cuz you can

  380. 12:26

    write code and make it better at any

  381. 12:28

    given point.

  382. 12:29

    But also if you procure SaaS, um if you

  383. 12:33

    can make that SaaS work on your behalf

  384. 12:35

    and

  385. 12:36

    be arbitrarily customizable, then that

  386. 12:38

    works too, right? Like you don't need to

  387. 12:39

    build this like

  388. 12:41

    GUI and like have a proactive roadmap as

  389. 12:44

    to like what features would make really

  390. 12:45

    great sense inside of some system. Like

  391. 12:47

    if you can arbitrarily customize the

  392. 12:49

    system, even if it's a system you've

  393. 12:50

    purchased, then you're like pretty good,

  394. 12:52

    right? So like for example, we use

  395. 12:53

    Salesforce. Like we use Salesforce at

  396. 12:55

    X.ai and uh it's great because uh it's a

  397. 12:58

    really good good database. It's made a

  398. 13:00

    lot of amazing choices around what sales

  399. 13:03

    should look like, choices that we don't

  400. 13:04

    want to make ourselves. And then it

  401. 13:06

    exposes MCP. So all of our agents have

  402. 13:08

    access to Salesforce MCP. Works really

  403. 13:10

    well. Our team uses it every day.

  404. 13:11

    And so yeah, I think infinite

  405. 13:13

    customizability um is is really the

  406. 13:15

    highest order bit.

  407. 13:19

    Cool. Um

  408. 13:21

    that's that's all I had.

  409. 13:23

    Uh

  410. 13:24

    Yeah, does anyone have any questions?

  411. 13:26

    >> Oh, we have time for a few questions.

  412. 13:28

    Okay, coming.

  413. 13:36

    >> Hey, um so you said you uh took all your

  414. 13:39

    past decisions. Can you elaborate a bit

  415. 13:41

    about that?

  416. 13:43

    What What artifacts are those?

  417. 13:45

    Usually people don't save like their

  418. 13:47

    decisions. Is it Slack? Is it email? Is

  419. 13:49

    it other other artifacts?

  420. 13:52

    >> Yeah, good question. I looked at

  421. 13:53

    decisions I made within Slack and email.

  422. 13:56

    I mean, a surprisingly large amount of

  423. 13:59

    everything that goes on a company is

  424. 14:01

    is is on Slack, right? So like if you

  425. 14:03

    just read like a ton of Slack history,

  426. 14:04

    like you can definitely find hundreds of

  427. 14:07

    decisions that you made in the past.

  428. 14:09

    >> Awesome. Quick question uh over here. Uh

  429. 14:12

    so your go-to-market team, what's the

  430. 14:13

    split between uh are they just all like

  431. 14:17

    AI cracked or do they also have like the

  432. 14:19

    domain expertise, too? What's the split

  433. 14:21

    between

  434. 14:22

    technical and non-technical? Because

  435. 14:24

    obviously you need them to like know how

  436. 14:26

    to do marketing, sales, etc. But then do

  437. 14:28

    they also are they also upskilling in

  438. 14:30

    terms of using AI systems? Are you

  439. 14:31

    handing them tools or they building

  440. 14:33

    their own?

  441. 14:35

    >> That's a very good question. So our

  442. 14:37

    go-to-market team

  443. 14:39

    is comprised of like there's there's

  444. 14:42

    account executives which like run the

  445. 14:44

    deals.

  446. 14:45

    There are like sales like SDRs that help

  447. 14:49

    with uh demand generation.

  448. 14:51

    And then there are separately they're

  449. 14:54

    separate like a FDE org. So forward

  450. 14:56

    deployed engineering organization.

  451. 14:58

    And

  452. 14:59

    what I'll say is that like

  453. 15:02

    everyone that's Okay, everyone that's uh

  454. 15:04

    not not in FDE

  455. 15:06

    is like

  456. 15:08

    has learned how to use AI really well.

  457. 15:11

    So like

  458. 15:12

    the answer is like they're not vibe

  459. 15:13

    coding. They're not generally with you

  460. 15:16

    know, in some there's some exceptions.

  461. 15:17

    They're not generally vibe coding these

  462. 15:18

    interfaces that we have. Um but they're

  463. 15:20

    using the tools really really well. And

  464. 15:22

    like we make sure that we have training

  465. 15:24

    sessions and like just make sure that

  466. 15:25

    people

  467. 15:26

    really understand how to use these

  468. 15:28

    tools. And then this funny we have this

  469. 15:30

    funny thing which is like our four

  470. 15:31

    deployed engineering organization is

  471. 15:33

    actually the one that like does a lot of

  472. 15:35

    the maintenance and feature building

  473. 15:37

    uh of these AI systems. And so they're

  474. 15:39

    both running deals and like like

  475. 15:42

    supporting deals

  476. 15:43

    um but then also making

  477. 15:46

    everything smoother by like

  478. 15:48

    doing sales but then also building the

  479. 15:51

    sales system. Like it's it's kind of

  480. 15:54

    it's kind of a funky thing we have going

  481. 15:55

    on. Yeah.

  482. 15:58

    >> How do you think about uh

  483. 16:00

    different security?

  484. 16:02

    >> Oh.

  485. 16:02

    >> Hey. How do you think about different uh

  486. 16:04

    security boundaries within your

  487. 16:05

    enterprise? What do you What you said

  488. 16:07

    suggested that you've got Jeffbot which

  489. 16:09

    had runs with all of your full

  490. 16:11

    privileges and then it's available to

  491. 16:13

    everybody which suggests that there's

  492. 16:14

    one security level and everyone can see

  493. 16:16

    everything all the time. Is that what

  494. 16:18

    you're going with or

  495. 16:19

    is there some uh other guardrails in

  496. 16:21

    place?

  497. 16:22

    >> Yeah, that's a good question. We we pay

  498. 16:24

    pretty special we we pay pretty careful

  499. 16:26

    attention to guardrails. So for example

  500. 16:29

    um in the case of Jeffbot

  501. 16:31

    um

  502. 16:32

    when I use Jeffbot and I call Jeffbot

  503. 16:34

    has access to

  504. 16:36

    a ton of systems and it can for example

  505. 16:38

    do reads and writes. However, when

  506. 16:40

    anybody else calls Jeffbot all can do is

  507. 16:42

    draft messages and also I don't give

  508. 16:45

    Jeffbot

  509. 16:46

    permissions to

  510. 16:48

    all of our MCPs and tools in the case

  511. 16:50

    where other people call it. And so in

  512. 16:52

    short it's like pretty

  513. 16:54

    it's pretty well defined or we we we do

  514. 16:56

    pay some care to the security.

  515. 16:58

    Yeah.

  516. 16:59

    >> Okay, last question.

  517. 17:04

    >> Um, can you can you share the origin

  518. 17:07

    story of the FDE team? Did that just

  519. 17:10

    happen organically or did you

  520. 17:11

    intentionally do it? I'm just really

  521. 17:13

    curious like how that came to exist.

  522. 17:16

    >> Yeah, for sure. I mean,

  523. 17:18

    uh

  524. 17:19

    my

  525. 17:20

    my philos- my my hypothesis on this is

  526. 17:22

    like, once upon a time the FDE role

  527. 17:24

    didn't really exist. Like, Palantir

  528. 17:26

    started calling some people FDEs, but it

  529. 17:28

    that was really it. And what tech

  530. 17:30

    companies had was like solutions and

  531. 17:32

    sales engineers.

  532. 17:34

    And then, like account executives.

  533. 17:37

    >> I was a solutions engineer.

  534. 17:38

    >> Got it. Yeah, yeah. The thing The thing

  535. 17:40

    that I think has changed is that um

  536. 17:42

    because of AI, as like

  537. 17:46

    en- as a technical person that is

  538. 17:48

    supporting revenue generation,

  539. 17:51

    you can actually not only support the

  540. 17:52

    revenue generation, but then very easily

  541. 17:54

    build the tooling

  542. 17:55

    to smooth everything over.

  543. 17:58

    And make your own life easier, make the

  544. 18:00

    lives of AEs easier. Like, because of

  545. 18:01

    AI, this is just possible now. Like,

  546. 18:03

    that's like two

  547. 18:04

    Before that was like two jobs, and now

  548. 18:05

    it's like one job.

  549. 18:07

    Um in theory. Like, now when our team

  550. 18:09

    grows, like right now it's about eight

  551. 18:11

    or nine FDEs, like what will will it

  552. 18:13

    scale such that everyone does

  553. 18:14

    everything? Probably not. But, at least

  554. 18:16

    right now that's what we have, and I

  555. 18:17

    think that's a really good working model

  556. 18:18

    to get pretty far.

  557. 18:20

    >> Eight out of how many?

  558. 18:22

    >> Uh eight Oh, eight of like how big is

  559. 18:23

    our go-to-market org?

  560. 18:24

    >> Or eight You have eight FDEs, and the

  561. 18:26

    size of the company right now is how

  562. 18:28

    many?

  563. 18:28

    >> Oh, the We're about 115 people.

  564. 18:31

    >> Okay.

  565. 18:31

    >> Yeah.