AI Engineer World's Fair 2026
Build the AI GTM Agent That Knows the Buyer Before the First Message
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.
Read the complete timestamped transcript
- 0:01
Hello. We know that the modern B2B selling has evolved significantly. The uncomfortable truth is that by the time the buyer reaches you, the decision is mostly made. They've researched, compared, and probably shortlisted vendors or partners that they want to work with.
- 0:20
That may include you or your competitors. This talk is about the architecture that lets GTM teams know the buyer more than they today.
- 0:32
My name is Sajjan Kanukolanu. I have twenty-plus years' experience across product, technology, and marketing. At Position² as the VP of Global Operations and Strategy, I lead the services teams and our AI native transformation from vision to positionings to deployment, where I work closely with our technology team.
- 0:54
We have launched seventy-five-plus AI agents specifically for our clients, eighteen-plus vertical knowledge bases that power these agents, and we've had eight hundred-plus runs per month, and this is data as of month to date June twenty twenty-six.
- 1:11
By the time we get to the end of the year, I expect these numbers to increase significantly.
- 1:18
The interesting thing about GTM and AI is that many GTM teams try to bolt AI onto their existing stack or existing processes and workflows, and therein lies the problem.
- 1:33
That approach itself is hindrance to scale. The interesting thing, though, is there are three different problems that GTM leaders and GTM teams have to solve in order to be successful in this AI era.
- 1:50
And these three are, number one, AI. By just bolting AI onto your existing systems, AI doesn't really and cannot really find out
- 2:01
who your buyer is. It doesn't understand their role. It doesn't understand their history. It doesn't understand their intent standalone.
- 2:10
That's number one. Number two is the integration itself. The old GTM stacks may not necessarily be best positioned to capture all the intent signals and the data, the CRM context, context that the AI captures on its own and bring that all together.
- 2:27
So the systems today, as they stand, may be broken. By putting AI on top of it is not gonna help solve a problem. The third is the architecture.
- 2:38
The underlying architecture has to evolve with AI at the core, and I'll talk about what I mean by that in a little bit. But the fundamental concept here is that with AI bolted on to your existing GTM systems, processes, and workflows, you will not be able to scale.
- 3:00
And most GTM deployments do not solve for all these three problems simultaneously. You could accomplish one and two, but if three is not accomplished or you don't solve for number three, then you still have a problem.
- 3:14
So it is important that all these three problems are solved together, and I will show you an architecture diagram in just a bit that's gonna walk you through how we do it at Position².
- 3:24
Another problem that we face is that with the buyer. The buyers are constantly evolving,
- 3:32
and what I mean by that is by the time the buyer reaches out to you in today's world, they are not early stage. They're actually late stage.
- 3:42
Here are some stats to prove that point. Ninety-four percent of buyers use GenAI as a primary research,
- 3:49
and these platforms are almost like a black box where we don't know what they consumed, what they read about, who they have investigated, if your organization is on top of that list or not.
- 4:02
And that is information that is from the Forrester twenty twenty-six research. The second stat is about sixty-seven percent of B2B buyers prefer a rep-free experience. They don't want to talk to people as they make their decision.
- 4:18
Eighty percent of deals, very interesting statistic. Eighty percent of deals go to the buyers who are part of the pre-contact list, so essentially buyers have already made up their mind.
- 4:31
And the fourth and very vital stat here is only seventeen percent, only seventeen percent of the total buying time is spent talking with potential vendors. What all these numbers tell us is the buyers do their research thoroughly.
- 4:46
They use various tools, including the GenAI platforms, and by the time they come talk with you or come to your website, they have made up their mind.
- 4:55
Now imagine a situation where the buyer does that research, comes to your website, and you have a chat window for them to engage with so they can converse with your very intelligent AI system that is now responsible for telling them about what you have to offer and how they can buy your services or product.
- 5:15
But many GTM systems that are bolted with AI would essentially start off with a conversation that would be something like this: "How can I help you?
- 5:27
What is it that you would like to... What is it that you would like to know today?"
- 5:32
But these are questions that actually move you backwards. They don't move you forward as an organization. The buyer has done their research, and when some basic questions are asked to them, they're actually walking out of the system.
- 5:46
They're not continuing their conversation. So it is very important that as a seller, you know what the buyer's needs are, you know who the buyer is, you know who the buyer is, who-- what their buying needs are, and you're able to cater to those as they engage on your site.
- 6:04
So what does this really mean when it comes to implementing an architecture? How does one accomplish an architecture or a system where the buyer information is known as soon as they come to your site?
- 6:19
There are essentially three architectural layers behind an intelligent system and running an intelligent GTM. The first is signals. This includes all the communication and information coming from your CRM systems, which includes your information about the deals, your contacts, who the owner or the sales rep is for that, who are you conversing with.
- 6:43
It also includes an enrichment system where you have the right information pulled out of these anonymous visitors that are hitting your site.
- 6:53
This system is going to work as long as that information is accurate. If you're off on the enrichment, your data is off and so is your communication.
- 7:04
And one of the things that we find really helpful and really useful, very important, is signals coming out of social media platforms, especially LinkedIn. What we find is the engagement, the job changes are huge signals for you to capture as an organization, which we do as ourselves, and I'll show you what, what I mean by that.
- 7:25
And these, these signals really help identify how and who you should target and when. So, for example, if you have an exec sponsor at your current account and that individual is leaving the organization and moving into another account, and that company is not part of your list, given that you have this exec sponsor currently as your contact,
- 7:50
you ought to make sure that the company that they're moving to next is part of your account list, and they become your new exec sponsor, your champion in that organization.
- 8:00
That's the power of social signals, and it's very important that these are captured, and not many companies leverage this in a automated, systematic way, and I'll talk about how to, how to do that in just a second.
- 8:12
The second layer is that of buyer intelligence.
- 8:15
This is about making sure you have the right information about the buyer. It all starts with, for us, at least at Position² and we believe, uh, this is how it should be for a lot of organizations,
- 8:28
is through a knowledge base. A knowledge base is a repository of context that you have about your product, your buyers, the ICPs, the personas, the different titles, different playbooks, different criterias of success.
- 8:46
Everything that makes you successful sits in a knowledge base. It also has visitor identity information, de-anonymized information sitting in your buyer intelligence database. The ICP scoring. What defines a score?
- 9:01
Uh, uh, what score defines a fit, and w-what score defines a buying stage versus someone who's a research mode? All these are things you have to codify within your buyer intelligence system.
- 9:13
Then you have a context builder. All the signals that you have need to form a context graph that I'll talk about next.
- 9:22
And then you need to have a routing logic. This is where you would define what message goes to who when. So if someone were to e-engage with you on LinkedIn, do they get a LinkedIn outreach or do they get a email or should they be shown a ad?
- 9:41
Um, these are the insights or inputs that actually go as part of your routing logic.
- 9:47
The third layer is that of action. Once you have the signals and the information and the intelligence from the first two layers, now it's time for action, which basically means now you have...
- 9:58
when user lands on your, uh, page or your website, the chat window that popped up earlier that I talked about shouldn't ask generic questions like, "What can I help you with?
- 10:09
What are you looking for? What's your name?" That's not the communication at this point. The communication is personalized, and I'll show you an example of what I mean by that.
- 10:16
It needs to identify the user, address them by name,
- 10:21
and talk about where they potentially left off last time they were on your site. Or based on the intent signals, the chat needs to offer them specific content.
- 10:31
The action also based on these signals should be about sending the right alerts and information to the rep. Too many, you lose the, the rep, the sales rep, and too few, you're not helping them hit their, hit their numbers, um, and, and get revenue for the organization.
- 10:52
The third piece here is that of updating the CRM. The CRM system cannot just be an update about the contacts that the sales had. It needs to be about the context on-- A-about the account that the first two layers drive for you.
- 11:09
And finally, the sequence trigger. The actions need to determine what type of outreach sequence need to be sent out at what time. What outreach is triggered when to who.
- 11:22
Without all these three layers running in tandem,
- 11:27
the GTM model is still obsolete, is old school, and it's not set for the AI era that we are talking about.
- 11:39
And here's an example of context graph. All the signals we just talked about roll into this connected record, which is per buyer. The persona, account, and deal, these are all signals and actions that are logged for every touch point across the account for every single person.
- 12:03
A person's signals link to their account The accounts have multiple people in them, so the buying committee information surfaces for the account based on the signals coming by person and for each account.
- 12:21
Then you have deal-level information here, which is very specific to sales and marketing activities. So has sales and has marketing already reached out to this individual or people, the buying committee at this account?
- 12:38
And all these together, with all those signals and data points and touch points, become a context graph.
- 12:46
The reason why a context graph is important is because without it, you don't know which account is a priority and which one is not. And therefore, when you have this information, you're able to prioritize the high-intent accounts and contacts within those accounts for your sales team and marketing to pursue.
- 13:08
Now, having covered the architecture and the importance of a con-- a context graph, I want to walk you through
- 13:17
how we implement this architecture and what that flow looks like, followed by a quick demo of what our system shows us when it comes to all the details that I just talked about, and how our system helps us identify key accounts and key stakeholders that we need to reach out to.
- 13:34
This is how the three-layered architecture I just talked about looks like as connected agents. When a user lands on our site, at that moment, we don't know who they are.
- 13:45
They're anonymous. And that is when that visit kicks off the first job, which is the identification job.
- 13:54
And as you'll notice, I'm using more than one source here running together, and that's because no single tool catches every visitor.
- 14:02
We pull the information provided by these tools, consolidate them, and dedupe them, and send it over to the next layer, which is the enrichment phase.
- 14:14
These tools change all the time for us, but not necessarily the architecture as such.
- 14:20
Once these contacts are enriched, the next layer, which is the intelligence layer, kicks in.
- 14:32
Here you'll notice that our ICP filter agent kicks in, which is referencing information from the knowledge base. What this does at this stage is it's looking back at the knowledge base and making sure that the contacts that come in are part of the criteria we've defined in our knowledge base.
- 14:54
Anything that doesn't fit, such as wrong industry or wrong geo, wrong size, wrong persona, are weeded out.
- 15:04
Those contacts and information are, are deleted. What fits our criteria and what's relevant
- 15:12
moves into a clean table of qualified accounts.
- 15:17
After this, the match agent kicks in. It checks for people against our CRM. Have we talked to them? Are they warm? Are they hot? That's the evaluation that happens at this stage.
- 15:32
Finally, the action agent kicks in. This agent decides what happens next per person.
- 15:42
And this could include one person from an account, or if there's a buying committee, the actions are decided for the entire buying committee per account.
- 15:56
Here, the action agent decides different alerts that the sales reps have to get on Slack with the right context and a draft of an email that they can possibly send.
- 16:10
This agent also decides if an email has to be fired up and sent over to the contacts we have. It also decides if we need to use LinkedIn for an outreach.
- 16:21
And this is how the workflow works, from the three-layered architecture to how we implement GTM agents today at Position2.
- 16:34
Now let me walk you through how the system actually looks in the back end. So we saw the three-layered architecture. We saw the agents working together. Now what does the output look like is what I'll show next.
- 16:50
Here's a view of our system at Position2, which is called Intelligence.
- 16:57
What you see here is as soon as I land, the system is able to understand who I am, and I see a personalized message
- 17:06
coming up here as part of our chat interface.
- 17:09
Instead of just asking me generic questions, it's actually pointing me to one of the conversations that I had with it earlier and prompting me or asking me if I want to continue there.
- 17:17
And if I were to say yes, it pulls up a link or gives me a link here, which I can click on and go to the specific conversation that I was, uh, part of earlier.
- 17:26
What you see here is a culmination of all the signals, information, and all the data that we collected as part of the user that is reflected here in a chat.
- 17:38
So as opposed to having a generic greeting for me, this is more personalized.
- 17:44
Let me now take you to our internal dashboard
- 17:49
and show you how this data actually is projected to us when we look at the accounts that we
- 17:58
are pursuing. Here's a view of the interface that gets us into a dashboard. What you have here is two key sections that we have built. One is for GTM, the other is organic.
- 18:09
For our conversation today, I'll be walking you through what happens when we go to our GTM dashboard.
- 18:16
Once you enter the dashboard, what you see are multiple cards. And for the conversation today, I will be focusing on the anonymous visitors and the LinkedIn intelligence sections. What you see here is when I click on the anonymous visitors, it takes us to a dashboard where
- 18:32
the flow of agents and the architecture that I showed you earlier is de-anonymizing the users. It's identifying the users that come onto our website, and it's categorizing them into multiple sections.
- 18:42
What you see here is we have-- Our system has identified about three thousand odd people or visitors coming in from about two hundred and eighty different accounts. And the system tells us that information technology is one of the key verticals or key industries that's visiting our site.
- 19:01
And this is, in fact, one of the verticals that we focus on. What you also see is the industry breakdown as well as hot signals, and this is information of people
- 19:10
from different companies that are being categorized as hot contacts or hot accounts at this point. This is a combination of different signals, including the conversations that have happened within HubSpot.
- 19:22
For the sake of the conversation here, I have changed the names, although the titles are accurate. If I were to scroll down, I have more information about these users, about the contacts, the companies they belong to, the location, as well as the industry type.
- 19:43
All this information is, uh, or, or can be sorted. All this information can be sorted
- 19:53
based on industry, seniority, different types of engagements they've had with us,
- 19:59
and we can also look at the companies that visit our site.
- 20:05
The second part to this is if I were to go to the LinkedIn intelligence, and this is a part that many GTM teams miss. LinkedIn has a wealth of information that is very useful.
- 20:17
What you see here is over the last few days, about a hundred people visited our site through LinkedIn from seventy-three different companies, and they've engaged with eight different posts that we put out recently.
- 20:30
It has a list of all the folks that have engaged and the type of engagement they've had with us. We can actually sort them out by C-suite, VPs, directors, managers, and individual contributors.
- 20:41
Those are the categories that we use, and you have a list of companies where they belong to. By clicking on each of these, we can actually go into very specific details about their overall engagement with us.
- 20:56
This level of detail really helps us bring in the right level of intelligence into our system. Now that we have looked at the architecture, let me walk you through what does tend to break when it comes to building these types of systems, and I want to focus on four key areas.
- 21:15
The first is the ICP drift, and this is real. Your ICPs evolve over time. For example, today you focus on a twenty million dollar ARR organization from a sort of vertical, and if you were to close a fifty million dollar or a hundred million dollar ARR firm from a different vertical, then you might realize that that is
- 21:31
a new area for you to focus on and for you to pursue. And that information is critical to enter into your knowledge base. Without that, your agents are working off of the old information, always giving you wrong in- wrong data that you might not want.
- 21:45
What this means is you need to retrain your models, to retrain your agents every quarter
- 21:53
with your closed won or closed lost opportunities, and this is critical. Without this, your agents are looking at wrong information, pointing you to the wrong accounts and the wrong people.
- 22:07
The second point is alert fatigue. If everything is flagged as hot to a sales rep, they would stop acting because it's, it just gets overwhelming for them.
- 22:21
And at that moment, they stop trusting the system, and the system's dead.
- 22:26
So it's important to use the knowledge base and the context graph to cut through the noise and focus on the alerts.
- 22:33
Focus on alerts from the accounts and contacts that do actually matter for the salespeople to act upon.
- 22:43
The third item is identity ceiling, and this is a limit that exists with tools today. A lot of these platforms, when it comes to identifying the anonymous visitors, they're almost seventy percent, a little more than seventy percent accurate when it comes to company identification.
- 23:01
They're only about fifteen to twenty percent accurate when it comes to individual identification, and this is a structural limit and a structural issue with these systems. The only way to truly maximize this is to keep testing different systems till you hit on one that gives you a much better numbers than what you already have.
- 23:22
But most important thing here is to keep these metrics in mind as you plan your GTM outreach.
- 23:30
And finally, the human bottleneck. The minute the approval queue feels like a chore, it's going to become backlogged. It's going to be ignored, and the system is not gonna be trusted anymore.
- 23:45
What this means is you need to make sure that you shrink the friction between,
- 23:52
let's say, a draft email that's ready to be sent to that being sent. And what I mean is if your draft email is in a spot where it's almost ready to go with very minor edit,
- 24:05
and you're able to kick it out with just one click, then that's a system that's gonna sustain. But if the time taken to edit an email that the AI writes
- 24:16
is, let's say, longer than... We look at thirty seconds. If it's longer than thirty seconds- then this is a dead initiative because
- 24:27
the sales would then rather focus on drafting their own email as opposed to trusting what the system gives them. Finally, I want to leave you with four key takeaways.
- 24:37
One, you have to start with the identity.
- 24:40
Not being able to identify visitors who s- who land on your site
- 24:45
is as good as a dead GTM project. What you need is a system that's put in place, just like what I showed you earlier or something similar to that, that is robust enough to identify the visitors coming to your website
- 24:58
and has a feedback mechanism where that information is able to go back into your knowledge base and vice versa. The second is the score fit and intent needs to be looked at separately.
- 25:09
Conflating them is going to send the wrong message to the wrong person.
- 25:14
The third is you need to have a policy engine that is auditable and adjustable. What it means is you need to have a system where if it doesn't work, you have a way to audit which agent or what part of the workflow, for example, like the one that I just showed you, is broken and you have a
- 25:32
way to go fix it, as opposed to going after a developer to help you come fix it. So you need to build a system where, as a GTM leader or a GTM team member, you are able to control the system and fix it when needed, because things do break quite frequently when it comes to AI and managing
- 25:50
the output from AI. Finally, you need to let the flywheel compound. What I mean by that is every send, every reply, and every closed deal should make the model smarter, should make your system smarter.
- 26:07
What it means is any information about wins, losses, deferred sales, all that needs to go back into your knowledge base, and your agents need to be able to capture that information to make the right decision for the future.
- 26:23
That's all I have for you today. Thank you very much for your time