Scoring splits in two, because the data splits in two.
- Pre-contact — what you can read: profile fit, services fit. That is
prospect-scoring. - Post-contact — what only a conversation establishes: whether they engaged, what the deal is actually worth, who really decides, when they would move.
This skill is the second half, and it starts from the first. A lead score that re-derives fit from scratch throws away the work already done and produces two different fit numbers for the same account.
The structure
lead score = prospect profile (carried over, capped) + post-contact points (0 until contact)
A workable default — tune the weights, keep the shape:
| Block | Max | Available |
|---|---|---|
Prospect profile (from prospect-scoring, normalised) |
30 | before contact |
| Engagement — replied, opened a thread, asked something real, attended | 15 | after contact |
| Confirmed deal potential — a number they said, not one you inferred | 20 | after contact |
| Decision path — you know who signs and how they buy | 15 | after contact |
| Timeline — a date or a window they named | 10 | after contact |
| Access — you are talking to the person who decides, or to a champion who will carry it | 10 | after contact |
Hard ceiling rule. A lead with zero post-contact points cannot outrank one that has them, however good its profile. Cap uncontacted rows below the top band. Otherwise the queue fills with beautiful strangers and the people who actually replied sit behind them.
Bands come from your pipeline, not from the rubric
A rubric's thresholds are a guess until a real pipeline calibrates them. If the team's live sheet treats 60+ as worth a call and the rubric says 75, the sheet wins — it is the one that has been compared against outcomes. Write both numbers down and reconcile them deliberately; do not let a theoretical threshold quietly override observed behaviour.
Recalibrate after every meaningful cohort, and recalibrate on accounts, not rows — a company with three contacts otherwise counts three times and drags the whole model.
Only calibrate on a clean cohort. Accounts with mostly empty fields will show absurd conversion rates, because they were usually created by hand after they converted. Backfilled records predicting the outcome that created them is the most common way a scoring model gets a flattering, useless number. In one calibration a base looked like it converted at 9.6%; on the clean cohort it was 3.2%.
When the base has no deal values, the model optimises conversations, not revenue — and will systematically favour small accounts, because small firms answer more. Say so out loud rather than presenting the ranking as a revenue ranking. Recalibrate only once deal values exist.
Hard gates beat points
Some facts should never be outweighed by a good total:
- Confirmed non-target → clamp the score below the disqualify line, no matter what else scored.
- Data doubt on a load-bearing field → clamp below the top band. The lead stays in the queue and gets human eyes, rather than being either promoted on bad data or silently dropped.
- Explicit opt-out → out of scoring entirely, out of all campaigns.
A gate that only subtracts points is not a gate. It has to clamp.
Keep the components visible
Store every block as its own column, never just the total. When a lead scores 78, you have to be able to see whether that came from a confirmed budget or from three soft engagement points — those two 78s deserve opposite actions. A single total column makes the model unauditable and, in practice, untrusted by the people meant to work the queue.
Same rule for the inputs behind a block: where a number could come from several sources, keep the sources in separate columns and note which one was used. Merging them loses the distinction between a verified figure and an inferred one exactly when you need it.
Output
One row per lead: Name · Company · Prospect profile · Engagement · Deal · Decision path · Timeline · Access · Total · Band · Action · Data confidence · Gate fired (if any) · What is missing.
Sorted by band, then by what is missing — because the top of a "needs more information" list is usually a better use of an hour than the bottom of the qualified one.
Handoff
Band → action: a call, one qualifying reply, a nurture track, or disqualify. The reply itself is
written by reply-objection-handler, which also triages it for intent and speed. Patterns across
many scored leads — a whole band that never converts — belong in reply-audit and then back into
hypothesis-builder.