AI lead scoring: how to make a score you can actually defend

1. What lead scoring is for

Lead scoring exists to answer one question: of the names in front of me, which ones do I call first? That is it. A score is not a prediction of revenue and not a measure of company quality. It is a ranking device for your own limited hours.

That framing matters because it sets the bar for a good score. A good score is one that changes the order you work the list. If every lead scores between 70 and 80, the model has told you nothing, no matter how confident the number looks.

2. Why most scores are useless

Three failure modes show up over and over.

• The black box. A number appears with no reasoning. You cannot tell whether it is reading something real or pattern-matching on company size, so you either trust it blindly or ignore it.
• Missing data treated as bad data. A company with no website scores low. But no website is sometimes the strongest buying signal in the set, depending on what you sell. Absence is not a negative unless you decided it is.
• Firmographics dressed up as intent. Industry, headcount, and revenue band describe who a company is, not whether anything has changed for them lately. Fit is not intent.

3. Signals, tiers, and weights

A signal is a checkable statement about a business. Not "is a good fit" but "the About page names the owner" or "a review in the last 60 days mentions an unanswered call." Checkable means a person could go verify it in ten seconds.

Signals need tiers, because they are not equal. A workable split is two levels:

• A-tier signals define the fit. If none of them hit, this is not your customer. Weight them 3.
• B-tier signals are supporting evidence. They break ties between leads that already clear the A-tier bar. Weight them 1.

Then every signal gets three possible verdicts, not two: hit, miss, and unknown. Unknown is the one most systems skip, and skipping it is what produces confidently wrong scores. If the source needed to check a signal was not available, the honest answer is unknown, and unknown must contribute zero in either direction.

4. Anti-signals

Positive signals rank the list. Anti-signals remove names from it. They are the disqualifiers you would state out loud: already using a competitor, too large to care, wrong geography, franchise with no local decision-maker.

Anti-signals should not shade a score down by a few points. One clear hit should zero the lead, because a disqualified prospect with six positive signals is still disqualified. Shading produces the worst possible outcome: a well-evidenced lead near the top of the list that cannot buy.

5. Worked example

Real numbers from a pull of HVAC contractors in Nashville, scoring for call-handling software. Five A-tier signals, five B-tier. A hits count 3, B hits count 1.

Note what the unknown row did: nothing. It did not penalise the lead for a gap in Google's response, and it did not get quietly counted as a miss to make the math look complete.

Then set a bar. Two thresholds are enough: a minimum number of A-tier hits, and a minimum funnel score. A lead under either one is dropped, and the drop reason is recorded so you can tell whether your bar is too high.

Funnel score	Tier	What to do
9 or more	HOT	Call this week
6 to 8	WARM	Worth a sequence
3 to 5	COLD	Only if your A-tier bar is 1
Below bar	DROP	Keep the reason, do not delete the row

6. Writing the ICP

The ICP is the input that decides everything downstream. Write it as three blocks.

1. Who, in one sentence. "Owner-operator HVAC companies in Middle Tennessee doing their own phone handling." Specific enough that a stranger could sort a list with it.
2. Positive signals, as checkable statements. Not "established business" but "100 to 2,000 Google reviews." Not "growing" but "a review in the last 30 days."
3. Disqualifiers, stated flatly. "Already running CallRail or WhatConverts." "Three or more locations." These become your anti-signals.

Then tune on one query before you run a hundred. Pick a market you know by eye, run twenty leads, and read the evidence quotes on the ones you disagree with. Nine times out of ten a bad score traces to a vague signal, not a bad model.

7. What public data cannot tell you

Public sources are enough to rank a list. They are not enough to know a business. Here is the honest boundary.

Readable today	Not readable from public sources
Review text, count, rating, recency	Contract renewal dates
Listing category, address, phone, website	Budget or revenue
Website copy, vendor scripts, hours, team size claims	Who actually signs
Whether a competitor's tooling is installed	Internal dissatisfaction that was never written down

Hiring posts, ownership-change records, and trade-directory membership are all readable in principle, but each needs its own source wired up. Until that work is done, the correct behaviour is to report unknown, not to infer.

8. Five common mistakes

1. Scoring before filtering. Run cheap mechanical filters first: no website, no phone, PO box addresses, thin review counts. Never spend a model call on a lead a regex could have rejected.
2. Too many A-tier signals. More than about five and everything scores HOT. Tiering only works if A-tier means something.
3. Weights with decimals. If you are arguing about 2.5 versus 2.75, the signal set is the problem, not the weight.
4. Deleting dropped leads. Keep them with their drop reason. That list is the only evidence of whether your bar is calibrated.
5. Never rereading the evidence. Scores drift as your market changes. Read ten evidence quotes a month and you will catch it early.

9. Questions