AI Engineer World's Fair 2026
Knowledge Systems: The New GTM Stack — Jeffrey Wang, Exa
Loading slides…
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.
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.
From scattered data to go-to-market action
Read the diagram as text
- Internal data. inputs — Customer knowledge, company information, and product usage.
- External data. inputs — Companies, people, news, and changing events.
- Live model of the world. knowledge — Connects product knowledge with potential and existing customers.
- Interfaces. access — The ICP dashboard and Request Lens expose recurring views and signals.
- Agents. access — Slack agents investigate accounts, build demos, and draft work.
- Go-to-market team. action — People interpret the information and act on customer opportunities.
- Internal data → Live model of the world: supplies company context.
- External data → Live model of the world: supplies market context.
- Live model of the world → Interfaces: supports recurring views.
- Live model of the world → Agents: supplies actionable context.
- Interfaces → Go-to-market team: surfaces companies and signals.
- Agents → Go-to-market team: returns research, demos, and drafts.
Suggest correction
This note stays in this page until you copy or download it. Nothing is submitted; reloading clears the draft.
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.
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.
Suggest correction
This note stays in this page until you copy or download it. Nothing is submitted; reloading clears the draft.
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.
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.
Suggest correction
This note stays in this page until you copy or download it. Nothing is submitted; reloading clears the draft.
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.
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.
Suggest correction
This note stays in this page until you copy or download it. Nothing is submitted; reloading clears the draft.
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.
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.
Suggest correction
This note stays in this page until you copy or download it. Nothing is submitted; reloading clears the draft.
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.
Jeffbot’s capability depends on who invokes it
Read the diagram as text
- Wang invokes Jeffbot. privileged — The invocation can access many systems and perform reads and writes.
- Another employee invokes Jeffbot. restricted — The invocation receives fewer MCP and tool permissions.
- Read and write actions. privileged — Available in Wang’s described use of the assistant.
- Draft messages. restricted — The employee-facing use prepares content without receiving the full tool set.
- Wang invokes Jeffbot → Read and write actions: broader permissions.
- Another employee invokes Jeffbot → Draft messages: restricted permissions.
Suggest correction
This note stays in this page until you copy or download it. Nothing is submitted; reloading clears the draft.
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.
Suggest correction
This note stays in this page until you copy or download it. Nothing is submitted; reloading clears the draft.
Resources
From the talk
The recording, corrected transcript, chapter links, and official reading version of Wang’s presentation.
- Jeffrey Wang on XReference
Wang’s supplied public profile, identifying him as an Exa cofounder.
Further reading
- AI EngineerReference
Conference information and the broader catalog of talks on agents, MCP, evaluation, interfaces, and AI engineering.
Related talks
- 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.
Full transcript
Loading transcript…