An AI Mission for Lead Qualification

A lead score is trained on two things at once: what the prospect did, and what you did to make them do it. Only one of those tells you something you did not already know.

The sales queue opens on a Monday and the top of it looks reassuring. The three highest-scoring accounts share a familiar shape — someone opened the last four emails, a contact clicked through from a paid placement, another spent real time on the pricing page, and a form came back with a job title from the buying committee the playbook describes. The reps work them in order, and two go quiet after a first call that never quite finds a reason to exist while the third turns out to be a curious analyst who will never buy anything. Meanwhile, several screens down and beneath a threshold nobody revisits, sits an account that has done almost nothing measurable and has just published a set of job openings for a function that only gets staffed once an organisation has committed budget to precisely the problem the product solves. Nothing in the scoring model has any way to see that, and nobody will find out until a competitor closes the account.

This is not a tuning failure, and adding more behavioural signals will not fix it, because the problem sits in what an engagement-weighted score is actually made of. Every click in that score is co-authored: the prospect supplied the action, and the company supplied the prompt, the placement, the cadence, the retargeting budget, and the decision to put that account in the segment at all. What the score measures, in large part, is how much attention the company has already paid to the prospect. It is not a reading of the market so much as a reflection of the go-to-market, and a team that ranks its pipeline by it is, to a degree nobody likes to quantify, ranking its pipeline by its own past decisions.

The score is partly a record of your own spending

Trace any behavioural signal back to its origin and the dependency becomes obvious. An email open requires that an email was sent, and the send list came out of a segmentation that already encoded a belief about who fits. A pricing-page visit is far more likely from an account that has sat inside a retargeting audience for a month, and membership in that audience was a media buying decision. Even the raw volume of activity per account follows coverage: territories with more reps generate more touches, more touches generate more responses, and more responses generate higher scores. Rank the resulting list and you have largely reproduced a map of where the budget went, dressed as a map of where the interest is.

The training loop compounds the effect rather than correcting it. Predictive fit models learn from closed-won deals, and closed-won deals only come from accounts the company chose to pursue, so the model learns the shape of last year's targeting rather than the shape of the demand. Segments that were never worked contribute no positive examples, so they score low, so they continue not to be worked, and the model's confidence in its own boundaries hardens with every cycle. It is a system with no access to the counterfactual, which makes it very good at describing its own history and very poor at noticing anything outside it.

None of this makes engagement data worthless, but it is worth being precise about what it is good for. Behaviour carries real timing information, and knowing that an account's attention shifted this week is useful once you have independent reason to believe it fits. What the typical composite score erases is provenance — whether a signal was solicited or arrived unbidden. An unprompted visit to a technical documentation page from an account never marketed to is expensive to fake and suggests someone inside is investigating a real problem, while the same page view from an account retargeted for six weeks mostly tells you the retargeting worked. Both land as the same event and contribute the same points, which is how a metric that averages a signal with its own echo ends up sounding confident about very little.

Situational fit is the evidence you did not generate

The alternative is not a better behavioural model but a different category of evidence altogether, and the useful name for it is situational fit. Firmographics — industry, headcount, region, revenue band — describe a category and change slowly, which is why they were never sufficient; they say an account belongs to a population that sometimes buys, not that this organisation, right now, is in a state where the problem is expensive and someone owns it. Situational fit is that second question, and it has answers. Has the company staffed a team whose existence presupposes the initiative, or named a programme, a risk, or a constraint in its own published reporting? Is it bound by a compliance obligation with a fixed horizon, has it announced a merger that will force the consolidation of exactly the systems you replace, or has it opened a procurement process of the kind public bodies routinely must?

What all of these have in common is that the company published them about itself, for its own reasons, entirely independently of your marketing. That independence is the whole point, because evidence you did not generate cannot be inflated by your own spending and does not degrade into a mirror when you look at it closely. It is also, worth saying plainly, the kind of evidence that sits comfortably within data-protection norms, being organisational and openly published rather than personal and quietly harvested. The disciplined version of the practice stays with public corporate disclosure, official registers and notices, an organisation's own published material, and first-party context a prospect knowingly gave you in a form, a call, or a support conversation. Purchased personal records of murky provenance and covert tracking sit outside that line, and they are not merely legally fragile; they are weak evidence for the question at hand, describing an individual's browsing rather than an institution's situation. The defensible sources and the informative ones turn out to be largely the same set.

The reason teams fall back on scores despite understanding all of this is not ignorance but arithmetic. Assembling a situational case takes real labour — reading filings, postings, disclosures, and the account's own published words, then judging whether the fragments add up to a reason someone would buy this quarter. A capable rep can do that for five accounts before a big meeting, and nobody can do it for five hundred, refreshed weekly. Lead scoring survived a decade of criticism because it was the only thing that scaled, and a mediocre signal covering the whole list will always beat an excellent signal covering two percent of it.

What changes is the request you make of the system

Once that assembly work can be done at the scale of a whole territory, the interesting question stops being how to rank the list and becomes what to ask for instead. The request that follows from this reframing is not "score these accounts" but something closer to an investigation: what is this organisation's situation, what published evidence supports that reading, what would contradict it, and what did we look for and fail to find. The output is not a number but a short, sourced argument about an account, with its reasoning exposed and its gaps admitted — the kind of thing a rep can read in a minute, disagree with in the second, and act on in the third.

That is a reasonable description of what an AI Mission is for, and it maps onto the architecture more cleanly than scoring did. Specialist agents can each own a class of evidence and gather it continuously through the Model Context Protocol integrations that reach the systems where the record lives; a Reasoning Core can weigh fragments no rule could reconcile, because the combination that matters — a hiring pattern plus a disclosed constraint plus a leadership change — is never the combination somebody wrote a branch for; Observations accumulate so that a change in an account's situation registers as an event rather than a quietly refreshed field. Enterprise Knowledge keeps it honest, since a company's own record of why deals were won and lost is the only place a claim about fit can be tested against something other than the marketing that produced the engagement. And Human-in-the-Loop is where the value compounds, because a rep who overrules the assessment produces exactly the labelled counterfactual scoring models never see.

Suspicion is warranted here, because this is exactly the territory where relabelling is most tempting. Gartner has predicted that over forty percent of agentic AI projects will be cancelled by the end of 2027, naming unclear business value and "agent washing" — existing tools re-badged as autonomous — among the causes, and lead qualification is a natural home for the practice. A scoring model that asks a language model to write a rationale beneath the same number has changed nothing about where the number came from. The test that separates the two is whether the output can be argued with: a score can only be overridden, whereas a case with its sources attached can be checked and contradicted, and a system whose conclusions can be contradicted is the only kind that learns anything from being wrong. That is the same distinction running through the category publication documenting how autonomous operations are being adopted across enterprise functions — software that produces an output, versus software that takes responsibility for the reasoning behind it.

What falls out of this for a sales team is a small change in vocabulary with a large change in behaviour. Stop asking the system who is most interested, because interest is the variable you have been paying to manufacture, and start asking what is true about each account that would remain true if you had never contacted them. Keep the engagement data, but demote it to what it is genuinely good at — telling you when to move on a situation you have independently established, not whether the situation exists. The mental model worth carrying is two ledgers rather than one blended score: an attention ledger recording everything you have spent on an account, and an evidence ledger recording what the world has published about it regardless of you. Conventional scoring adds the two together and reports the sum as insight, when the qualification that predicts anything is the second ledger read alone — and the honest diagnostic for any pipeline is how much of the score at the top of the queue would survive if the first ledger were set to zero.