Lead Scoring: Fit, Intent, and Explainable AI That Reps Trust

Lead scoring is a prioritization mechanism: given more leads than attention, rank them so the next hour of sales effort lands where it converts best. That framing matters because it defines success — a scoring model succeeds when reps work the ranked queue and win more, and fails when reps glance at the number, distrust it, and go back to gut feel.

It also defines the honest scope. A score does not need to predict revenue to the dollar or impress anyone with machine-learning vocabulary. It needs to reliably put the pricing-page-revisiting, ICP-fit account above the student downloading your ebook — and to be trusted enough that the ranking changes behavior.

Every workable model separates two questions. Fit: is this the kind of company and person we win with? Industry, company size, geography, role seniority, technology context — attributes that change slowly. Intent: are they showing buying behavior right now? Visits, page depth, recency, frequency, trial activity — evidence that changes daily.

Keep the axes separate, then combine multiplicatively, not additively. Addition lets one axis compensate for the other: a perfect-fit account with zero activity, or a hot-behaving account that is a terrible fit (a student, a competitor, a job-seeker) can both accumulate a high summed score. Multiplication encodes the truth that you need both — great fit with no motion is a future nurture target; great motion with no fit is noise.

The practical output is a two-by-two your team already intuits: high-fit high-intent gets immediate outreach, high-fit low-intent gets nurture and monitoring, low-fit high-intent gets a polite low-touch path, low-fit low-intent gets ignored without guilt.

The traditional approach: a workshop produces a point sheet — plus ten for a whitepaper download, plus five per email open, plus twenty for a webinar, minus five for a free email domain — wired into the marketing automation platform and reviewed never. It works for a quarter, then rots.

It rots for structural reasons:

• The weights are guesses frozen at workshop time, never validated against what actually converted.
• Points accumulate without decay, so a contact who opened newsletters for two years outscores this morning's pricing-page visitor.
• Activity volume masquerades as intent — rewarding your most-emailed contacts, not your most interested ones.
• Nobody can explain any specific score, so the first time a rep sees a "hot" lead that is obviously junk, trust dies and the model becomes decoration.

The lesson is not that manual models are worthless — it is that unvalidated weights, missing decay, and unexplainable outputs are fatal in any model, manual or ML.

The strongest foundation for the intent axis is deterministic behavior — observed actions with known meaning, weighted by what they actually indicate:

• Recency: a visit yesterday outweighs ten visits last quarter, so signals must decay.
• Frequency: three sessions in a week signals active evaluation.
• Depth: pricing and comparison pages outrank blog posts by a wide margin, because of where they sit in a buying journey.

Product signals extend the same logic past the website: a trial crossing a product-qualified-lead milestone — teammate invited, integration connected, real usage threshold — is among the strongest intent evidence that exists. Our signal based selling guide catalogs the full signal hierarchy and what each one is worth.

"Deterministic" is the load-bearing word: every input is a real event that happened, not an inferred probability. That is what makes the score auditable — you can trace any number back to the behaviors that produced it, which is precisely what the rotting point-sheet could not offer.

The pitch for ML-based scoring is real: learned weights beat workshop guesses, and a model can find non-obvious conversion predictors. The classic failure is also real: a black-box percentile that no rep can interrogate. When "82" meets a rep's skepticism and has no reasons to offer, the model loses the argument — and a scoring model that loses arguments with reps has failed at its only job.

Our position, and how we built it into BusinessMCP: the score must ship with its reasons. Every visitor and account carries a deterministic fit-times-intent score with human-readable explanations attached, persisted per visitor and updated as behavior changes.

Explainability also fixes the feedback loop: when sales can see the reasons, they can tell you which reasons are misleading, and the model improves. A black box collects resentment instead of corrections.

One: instrument behavior first — website events with company identification (so the ~98% of visitors who never fill a form still generate intent data), plus product milestones if you have a trial. Without behavioral input, any score is a fit filter wearing a costume. Two: define fit crisply from your last twenty wins and losses — industry, size band, geography, role — and write down the disqualifiers as firmly as the qualifiers.

Three: start with simple, defensible weights — heavy on pricing-and-comparison behavior, meaningful decay (a half-life of days to a couple of weeks for intent), multiplicative combination with fit. Four: wire scores to actions, not dashboards: thresholds that trigger alerts, queue ordering, or automated warm outreach. A score nobody acts on is a report.

Five: validate quarterly against outcomes — did high scores actually convert better? — and adjust weights with sales in the room. This is also the honest moment to admit a build-versus-buy choice: the loop above is buildable on a CRM plus analytics stack with real engineering effort, or it comes assembled in platforms like ours where scoring feeds outreach natively.

• Vanity points: rewarding email opens, newsletter tenure, and event attendance until your most-marketed-to contacts float to the top. If a behavior does not appear disproportionately in the history of closed-won deals, it earns no points.
• Score inflation without decay: the same disease in the time dimension — old signals must sink.
• No sales feedback loop: the model launches, sales quietly notices its misses, and instead of correcting weights the team abandons the queue. Schedule the feedback explicitly; disagreement between the score and rep judgment is training data, not insubordination.