AI Engineer World's Fair 2026
Build the AI GTM Agent That Knows the Buyer Before the First Message
Loading slides…
Read the talk
Build a GTM Agent That Knows the Buyer Before the First Message
Connect visitor identity, CRM history and social signals to buyer intelligence, then use that context to coordinate chat, sales alerts and outreach without restarting discovery.
From a talk by Dr. Sajjan Kanukolanu
Before you start: Familiarity with CRM records, B2B sales workflows and basic data pipelines will help; no specific agent framework is required.
The buyer arrives informed
By the time a B2B buyer contacts a vendor, they may already have researched the market, compared alternatives and assembled a shortlist. The seller enters a conversation that has been underway elsewhere, without knowing what the buyer has learned or which competitors remain in contention. The architectural problem is how to recover enough context to make that first interaction useful.
Dr. Sajjan Kanukolanu approaches this problem from his work at Position², where he introduces himself as VP of Global Operations and Strategy, leading services and the company’s AI transformation through deployment. He brings more than twenty years across product, technology and marketing. In his June 2026 month-to-date update, Kanukolanu reports more than 75 client AI agents, more than 18 vertical knowledge bases and more than 800 runs per month. These describe the implementation practice behind the architecture, rather than a measured improvement in sales outcomes.
Suggest correction
This note stays in this page until you copy or download it. Nothing is submitted; reloading clears the draft.
Why adding AI to the existing stack is insufficient
Adding an AI interface to an existing go-to-market workflow leaves three connected problems to solve:
| Problem | What remains missing |
|---|---|
| Buyer understanding | The buyer’s role, history and intent |
| Integration | A joined view of intent signals, CRM records and newly captured context |
| Architecture | A system that uses that context throughout its decisions and actions |
A model cannot infer a relationship history that the surrounding system never supplies. Connecting a few data sources helps, but if the workflow still treats each interaction as an isolated event, the architecture remains the constraint. Buyer understanding, integration and architecture have to be solved together.
Suggest correction
This note stays in this page until you copy or download it. Nothing is submitted; reloading clears the draft.
Do not restart discovery at the chat window
The buyer’s research increasingly happens beyond the seller’s visibility. A generative AI conversation can conceal which content the buyer consumed, which vendors they investigated and whether the seller made the shortlist. Kanukolanu cites a 94% figure for GenAI as a primary research source; the linked Forrester account supports the narrower finding that 94% use AI during the buying process, drawing on its 2025 survey. The distinction matters: adoption does not establish that AI is every respondent’s primary research channel.
Kanukolanu also cites a 67% preference for a rep-free buying experience. That is a preference, not a claim that buyers never want human interaction. He then describes 80% of deals as going to vendors on a pre-contact list; the related 6sense research distinguishes roughly 80% wins for the pre-contact favorite from 95% for vendors on the initial shortlist, so favorite status and shortlist membership should not be treated as the same measure. Finally, he cites 17% of buying time spent talking with potential vendors, without specifying the underlying study population or period. The practical point is that vendor conversations occupy only part of the buyer’s decision process.
Now the informed buyer arrives on the website and opens its AI chat. The system asks, “How can I help you?” It follows with another broad discovery question, asking what the visitor wants to know. The seller has gained a conversational interface but still knows nothing about where to begin. Kanukolanu describes buyers abandoning these exchanges because the questions move them backward. The system needs to establish who the visitor is and what they need, then continue from that point.
Suggest correction
This note stays in this page until you copy or download it. Nothing is submitted; reloading clears the draft.
Signals, buyer intelligence and action
The architecture begins with a signals layer. CRM records contribute deals, contacts, account owners, sales representatives and existing conversations. Enrichment adds information about anonymous website visitors. Its quality is a dependency for everything downstream: incorrect identity or company information produces incorrect context and, eventually, inappropriate communication.
Social activity adds signals that a CRM alone may miss. LinkedIn engagement can reveal interest; job changes can reveal a new account opportunity. Consider an executive sponsor at an existing customer who moves to another company. The relationship already exists, but the target account has changed. Capturing that move systematically lets the team add the destination company to its account list and pursue the existing champion relationship there.
The buyer intelligence layer turns those inputs into decisions. Its knowledge base holds product information, buyer definitions, ideal customer profiles (ICPs), personas, titles, playbooks and success criteria. The buyer database also holds identified visitor records. Qualification cannot depend on an undefined impression of a good lead: the system must codify what constitutes fit, what indicates a buying stage and what looks like research. A context builder connects the signals, and routing logic decides who should receive which message, when and through which channel. LinkedIn engagement might lead to LinkedIn outreach, an email or an advertisement; the originating channel does not determine the response by itself.
The action layer uses those decisions in several places:
- Website chat: Address an identified visitor by name, resume prior activity or offer content selected from intent signals.
- Sales alerts: Give representatives enough actionable information without overwhelming them.
- CRM updates: Record account context produced by the first two layers, beyond a log of sales conversations.
- Outreach sequences: Choose the outreach type, recipient and timing.
These are coordinated outputs of the same intelligence. If chat uses context but CRM updates and outbound sequences do not, the buyer can still encounter a fragmented conversation.
Suggest correction
This note stays in this page until you copy or download it. Nothing is submitted; reloading clears the draft.
Connect people, accounts and deals
The context graph is the connected record behind those decisions. It links a buyer’s persona, account and deal to signals, actions and touchpoints. Person-level activity belongs to an account, and an account can contain several people. Connecting those records makes a buying committee visible instead of leaving its members scattered across unrelated contact entries.
Deal context adds another essential question: has sales or marketing already contacted this person, or other members of the committee? Combined with the activity history, this lets the team prioritize high-intent accounts and select the relevant people within them. The graph’s value is not simply that it stores more facts; it lets an account-level decision take account of multiple people and prior outreach.
Suggest correction
This note stays in this page until you copy or download it. Nothing is submitted; reloading clears the draft.
From an anonymous visit to an outreach decision
Position² implements the layers as connected jobs and agents. The initial visit does not arrive with a known person attached. Recognition is work the system must perform:
- Identify the visitor. The visit triggers an identification job. Multiple sources run together because no single tool catches every visitor.
- Consolidate and deduplicate. Combine the identification outputs before passing them into enrichment.
- Enrich the contact. Add the information needed by the intelligence layer. Individual providers can change without changing this overall architecture.
- Apply the ICP filter. An agent consults knowledge-base criteria for industry, geography, company size and persona. Nonmatching contacts and their information are deleted from this flow; qualifying records enter a clean table of qualified accounts.
This ordering matters: qualification operates on consolidated, enriched records and a defined policy, rather than on whichever identification result arrived first.
The qualification boundary can be expressed as a small TypeScript function. It receives enriched records and the criteria maintained in the knowledge base, then produces the qualified collection. The particular industries, geographies, sizes and personas remain policy data rather than being embedded in the function.
typescript
type EnrichedContact = {
personId: string;
accountId: string;
industry: string;
geography: string;
companySize: number;
persona: string;
};
type IcpPolicy = {
industries: ReadonlySet<string>;
geographies: ReadonlySet<string>;
minCompanySize: number;
maxCompanySize: number;
personas: ReadonlySet<string>;
};
function qualifyContacts(
contacts: readonly EnrichedContact[],
policy: IcpPolicy,
): EnrichedContact[] {
return contacts.filter((contact) =>
policy.industries.has(contact.industry) &&
policy.geographies.has(contact.geography) &&
contact.companySize >= policy.minCompanySize &&
contact.companySize <= policy.maxCompanySize &&
policy.personas.has(contact.persona)
);
}
The returned collection excludes nonmatches; it does not itself delete stored records. In the described workflow, deletion and persistence of the clean table are subsequent storage operations.
Next, the match agent checks qualified people against CRM history: has the team talked to them, and are they warm or hot? The action agent then decides what happens per person, including coordinated decisions across an account’s buying committee. Its outputs can include a Slack alert with context and a draft email for a representative, a decision to send an email directly, or LinkedIn outreach. A draft supplied to a representative and an automatically sent message are different action paths; selecting an action does not make every draft a completed send.
Suggest correction
This note stays in this page until you copy or download it. Nothing is submitted; reloading clears the draft.
What the visitor and GTM team see
The product demonstration begins with Position²’s system, Intelligence. On arrival, Kanukolanu receives a personalized chat message that refers to an earlier conversation and asks whether he wants to continue it. He explains that accepting produces a link to that specific conversation. The visible experience is a continuation of known activity, replacing the generic opening question with a relevant next step.
The internal interface separates GTM and organic sections. Within GTM, Kanukolanu opens the anonymous-visitor dashboard to show how the identification workflow organizes visitors. The demonstrated dashboard contains approximately 3,000 visitors from approximately 280 accounts, with no reporting window stated. Information technology appears as a leading visiting industry and is also one of Position²’s target verticals.
The dashboard combines an industry breakdown with hot contacts and accounts, using signals that include prior HubSpot conversations. Kanukolanu has changed the displayed names for the demonstration while retaining accurate titles. Contact records expose company, location and industry information; the team can sort by industry, seniority and engagement, or inspect the companies visiting the site. These views turn the underlying records into choices about which accounts and people to pursue.
The LinkedIn intelligence view provides a second route into the same kind of account analysis. For the preceding few days, Kanukolanu reports about 100 people arriving through LinkedIn from 73 companies and engaging with eight recent posts. The view lists people, interaction types and company affiliations. Seniority filters separate C-suite executives, VPs, directors, managers and individual contributors; clicking a record exposes more detail about its overall engagement.
Suggest correction
This note stays in this page until you copy or download it. Nothing is submitted; reloading clears the draft.
Where the system breaks
ICP drift makes a correct filter stale. A team might initially target companies with $20 million in annual recurring revenue in one vertical, then win customers with $50 million or $100 million in another. Those outcomes can change which accounts deserve attention, but only if the knowledge base changes with them. Kanukolanu recommends quarterly retraining of models and agents using closed-won and closed-lost opportunities. He does not specify whether that involves model-weight updates or changes to knowledge and agent configuration; the concrete requirement is that new customer evidence reaches future qualification decisions.
Alert fatigue breaks the connection between intelligence and action. When every account is hot, representatives become overwhelmed, stop acting and lose trust in the system. The knowledge base and context graph should therefore restrict alerts to accounts and contacts worth a salesperson’s attention. Alert volume is not evidence that the agent is helping.
Identity limits constrain how much personalization the pipeline can support. Kanukolanu estimates roughly 70% or slightly higher company identification accuracy, versus approximately 15–20% for individuals. He describes these as structural limitations of available tools, but provides no evaluation defining whether the percentages measure correctness, coverage or match rate. Those measures are different: identifying a fraction of visitors does not establish that the resulting identities are correct. His operational advice is to test alternative systems against existing results and plan outreach around the identification limits actually observed.
Human approval friction can stop an otherwise functioning pipeline. A queue that feels like a chore becomes backlogged and ignored. Drafts should need only minor edits and make sending easy, ideally a single click after review. Position² uses more than 30 seconds of email editing as a warning threshold: Kanukolanu says representatives then prefer writing the message themselves. This is an adoption heuristic, tying the quality of generated drafts to the effort required to use them.
Suggest correction
This note stays in this page until you copy or download it. Nothing is submitted; reloading clears the draft.
Make the policy controllable and the feedback continuous
Identity is the starting dependency, but it must participate in a feedback loop: visitor information should flow into the knowledge base, and that knowledge should inform subsequent decisions. Score fit and intent separately. Fit asks whether this is the kind of buyer the business serves; intent asks what the available signals say about the buyer’s readiness. Combining the two into an undifferentiated judgment can send the wrong message to the wrong person.
An auditable, adjustable policy engine gives the GTM team control over those decisions. When an output is wrong, an operator needs to locate the responsible agent or workflow stage and adjust it, rather than depending on a developer for every correction. The qualification criteria, routing choices and action decisions need to remain inspectable parts of the system.
Finally, the workflow must capture what happens after it acts. Sends, replies and closed deals create new evidence; wins, losses and deferred sales should return to the knowledge base so agents can use them in later decisions. That closes the architectural loop: the system does not merely accumulate visitor signals and produce messages. It uses the outcomes of those messages to improve whom it recognizes as relevant, when it engages and what it proposes next.
Suggest correction
This note stays in this page until you copy or download it. Nothing is submitted; reloading clears the draft.
Resources
Further reading
Forrester explains its 2025 survey findings on business buyers' AI adoption and information sources.
A vendor's explanation of company versus person matching, traffic-dependent coverage, and the separate need to validate match correctness.
Full transcript
Loading transcript…