How to build a B2B lead scoring model your reps trust and an agent can act on

Most revenue teams already have a lead score. It lives in a CRM field, it was set up by someone who has since left, and the reps sort by it about as often as they sort by fax number. When they are asked why, the answers are consistent: the high scores were not ready to buy, the score never explains itself, and a student who downloaded four ebooks outranks a director who visited pricing once.

None of that is a reason to stop scoring. It is a reason to build the model differently. This guide covers the six steps, a worked example you can copy, a decision tree for choosing between rules, predictive, and a hybrid, and the one requirement that changes when an AI agent, not a rep, is the one acting on the score.

Why most lead scores get ignored

Lead scores fail in four predictable ways, and each one is a design choice rather than a data problem.

• One number for two questions. “Is this the right account?” and “Is it buying now?” have different answers and different plays. A single total averages them away.
• No reason attached. A rep who cannot see which signals produced a score cannot use it to open a conversation, so it gets ignored.
• No decay. Points accumulate forever, so old activity keeps accounts at the top of the list long after the interest has passed.
• No owner per threshold. Crossing a line triggers nothing in particular, so nothing in particular happens.

The handoff is where this shows up. If marketing and sales do not share a definition of a qualified account, the score becomes the argument instead of the answer; the revenue leakage playbook covers the shared dictionary that has to come first.

Rules, predictive, or a deterministic hybrid

There are three ways to produce the number. The choice depends on how much clean deal history you have and on who, or what, will act on the result.

A score is a description of an account. What to do about it is a separate decision, and most stacks stop at the description. The difference is laid out in buyer intelligence vs. decision intelligence.

Build the model in six steps

A worked example (illustrative)

The points, decay windows, and thresholds below are examples to show the mechanics, not recommended values. Your own weights come from step 3. Each axis is capped at 50.

Decay rule: an intent signal earns full points for 7 days, half points from day 8 to day 30, and nothing after day 30. Thresholds: fit of 30 or more is high fit; intent of 20 or more is high intent.

Look at the single-total column. Accounts A and C land within three points of each other, 57.5 and 55, and a one-number model would route both to a rep today. Only A is buying. C is a strong account that went quiet three weeks ago; calling it now spends a rep’s time and the account’s patience. Two axes make that difference visible. One total hides it.

Fit and intent: four quadrants, four plays

The high-fit, low-intent quadrant is where most pipeline is lost quietly. Those accounts are on the target list, nobody is watching them weekly, and the month their intent rises is the month nobody looked. Coverage of that quadrant, not volume against the top one, is what the sales process coverage metric measures.

Which scoring approach fits your team?

Answer for the team you have today, not the one in the plan.

When an agent acts on the score, reproducibility stops being optional

A rep can ignore a bad score. An agent acts on it, for every account, every time it runs. That raises the bar in three ways. The score has to be reproducible: two identical accounts scored a week apart get the same number. It has to be explainable: the signals behind it travel with it. And it has to be recorded, so that when an action looks wrong, someone can see what the score was and why.

This is why asking a language model to produce a priority score is the wrong design. Language models are good at reading and drafting; they are not built to return the same arithmetic twice. Scout, the Pipeline Research Analyst, scores accounts against your ICP from intent signals, CRM records, and web telemetry. Account scoring runs on sandboxed deterministic code. No language model performs arithmetic that has a numeric consequence, so a score can be reproduced and checked. Qualified accounts go to Herald, the Outbound Engagement Rep, whose drafts carry the signal that produced them.

What happens next is your team’s call. Your team chooses the mode for each action class, based on its risk tolerance, and can change it at any time. Human-in-the-loop (HITL): The action is drafted and held. It does not execute until a named person on your team approves it. Human-on-the-loop (HOTL): The action executes under a standing policy your team sets. A named person supervises and keeps intervention, override, and revocation authority. Every action, in either mode, is recorded to the audit ledger with its rationale, before-and-after state, and the approver or policy behind it. The trade-offs of each mode for revenue actions are covered in human-in-the-loop vs. human-on-the-loop.

Scoring is the first stage of the loop, not the whole of it. The four agents and their coordinator are specified on the Revenue Accelerator Stack page, and why most AI agents stop at top of funnel covers what breaks when a score hands off to a tool that cannot see what happens next.

How to tell whether your model is working

Three checks, run monthly, tell you more than any dashboard color.

• Conversion by band. Opportunity creation should rise from each band to the next. A flat line means the weights are not separating anything.
• Rep acceptance. Track how often reps accept or reject routed accounts, and why. Rejections with a reason are the cheapest calibration data you will get.
• Time to first touch in the act-now quadrant. A correct score acted on next week is a missed one. If this number is long, the problem is capacity, not the model.

The ratio and latency metrics behind those checks are covered in which SDR metrics actually predict pipeline.