If you cannot see the points, the caps, the half-lives and the band thresholds, then no. And a score your reps cannot interrogate is a score they will quietly stop using.
Lead scoring is drifting towards the same place as everything else: a model that produces a confident number and cannot explain it. The pitch is that it is more accurate than hand-set points, which is often true.
Accuracy is not the binding constraint. Adoption is.
The failure is social, not statistical
Here is what happens with an unexplainable score, reliably, in month three.
A rep calls the 92. It is a student doing research. They call the 88. It is a competitor. They call an 11 because they recognised the company name, and it closes.
Now ask them to work the list top-down. They will say yes and they will not do it, because they have three counterexamples and no way to find out what went wrong. There is no conversation to have with a model that cannot be inspected. You cannot say "it is over-weighting opens" because you cannot see that it is.
The score does not get fixed. It gets ignored, and everybody carries on using the account names they recognise, which is the thing scoring was introduced to replace.
A model that is 70% accurate and arguable beats one that is 85% accurate and opaque, because the first one gets used and improved and the second one gets abandoned without anybody announcing it.
The four things you have to be able to see
| Element | What it is | What goes wrong without it |
|---|---|---|
| Points | What each behaviour earns | Nobody can tell what the score rewards, so nobody can tell when it rewards the wrong thing |
| Caps | The maximum any single behaviour can contribute | One contact opening the same email forty times outranks somebody who replied |
| Half-life | How fast engagement decays | The score becomes a lifetime total and your hottest leads are your oldest contacts |
| Band thresholds | Where a lead becomes hot, warm or cold | The bands drift, everyone ends up hot, and the score stops sorting anything |
Three of those four are invisible in most scoring implementations, including plenty that are not AI-based. The half-life is the most commonly missing and the most damaging: a model with no decay is measuring accumulated history rather than current interest, and after eighteen months it is ranking your database by age.
What the score should be built from
Fit, which is stable: industry, size, region, whether they look like your won deals. This changes rarely and should be most of the ceiling.
Engagement, which is volatile and decays: what they did, how recently. This should move the score within a band rather than across all of them.
And it should exclude anything a machine can do on the recipient's behalf. Apple Mail prefetches tracking pixels, so somewhere between 58% and 65% of recorded opens in 2026 are not people, which is covered in Apple fakes most of your email opens. A model that counts opens is partly scoring mail servers, and if you cannot read the model you cannot confirm whether yours does.
This is why Signal Score is built as a readable model: a score out of 100 from fit and recent engagement, with the points, caps, half-lives and band thresholds all visible and editable, and machine opens never counting. It is a deliberately unfashionable design. The argument for it is that the score exists to be acted on by a person, and a person acts on a number they can argue with.
The test to run on yours
Take your ten highest-scoring leads. For each, write down in one sentence why it scored that high.
If you can do it for all ten, your model is readable and the rest is tuning. If you cannot do it for any, your reps already know, and the list you think they are working is not the list they are working.
When an opaque model is the right choice
To be fair to the other side: where the volume is high enough that no human reviews individual leads, and the output drives an automated action rather than a person's next call, explainability matters far less. Scoring millions of records for a routing decision is a different problem.
That is not most B2B teams. Most B2B teams have a few hundred leads a month and a handful of people deciding who to ring, and in that setting the score's job is to win an argument with a sceptical rep. It cannot win an argument it cannot participate in.
Related reading: Apple fakes most of your email opens, and marketing says 400 leads, sales says 12 are real.
See how Signal Score exposes its model, and what else sits in Signal's marketing side.
