Lead Scoring Without a Data Team

Somebody once built a scoring model. Opened an email, two points. Visited pricing, ten. Job title contains director, fifteen. Nobody has touched the weights in two years, nobody can explain why a lead scored 82, and the reps sort by date because at least a date means something. The model is not wrong exactly. It is unexamined, which turns out to be worse.

Not because reps are stubborn. Because of three properties most models share.

Nobody can explain the number. A score of 82 is not an argument. If the person acting on it cannot say what produced it, they cannot decide whether to believe it, so they fall back on their own read. That is a rational response to an opaque instrument.

It was never validated. The weights came from a workshop where people guessed. Nobody has ever checked whether high scores convert better than low ones, which is the only question that matters.

It rewards curiosity, not intent. Points for opens and page views mean a competitor researching you and a student writing a dissertation both score highly, and a serious buyer who read one page and picked up the phone scores nothing.

This is the design error underneath most bad models, and fixing it fixes most of the problem.

Fit: are they the kind of buyer we serve? Size, market, situation. Stable, knowable early, and it does not change.

Intent: are they acting now? Behaviour, timing, specificity. Volatile, and it changes weekly.

Collapse those into one number and you cannot distinguish a perfect-fit buyer with no urgency from a poor-fit buyer who is very active. They score identically and need opposite treatment: the first goes into a long, patient rhythm, and the second should probably be declined.

Across most businesses these outperform aggregate engagement, and all of them are observable without any modelling.

The build takes an afternoon and needs no tooling beyond what you already have.

One, list your last twenty customers. Actual customers, not opportunities.

Two, list twenty conversations that wasted your time. Equally important and usually skipped, because nobody enjoys writing that list.

Three, find what separates them. Usually two or three things, and usually more boring than expected: company size, whether they had a named problem, whether they had bought something like it before.

Four, write it as conditions, not weights. In our size range, in one of our markets, stated a problem we solve, replied within a week. Count how many are true. That is the score, and anyone can explain it.

Five, write down what you excluded and why. Job title, email opens, industry. Recording the rejects stops the model quietly re-accumulating them next year.

One test, run quarterly, that almost nobody runs.

Compare conversion rates by score band. Take a quarter of leads, split into high and low, and compare how many became customers. If high converts meaningfully better, the model works. If the rates are similar, it is not predicting anything and you are adding a step for nothing.

A model that fails this test is worse than no model, because it launders a guess as a number and people defer to numbers. Deleting it is a legitimate outcome and more useful than adjusting weights on something that never predicted anything.

Also check the false negatives: of the customers you won, how many scored low? A model that misses good customers is dangerous in a way that is invisible, because nobody audits the leads that were deprioritised.

The honest section, because for a lot of readers this whole post is a distraction from something more valuable.

If you can contact every lead, do not score. Scoring is prioritisation, and prioritisation only matters when there is more work than capacity. Below that line, effort spent on a model is effort not spent on answering everybody faster, which is the higher-return activity by a wide margin.

If you have fewer than a few hundred leads a year, the sample is too small. You cannot validate anything, and your own judgement on twenty conversations is better than a model built on the same twenty.

If the real problem is definition rather than ranking, fix that instead. When sales and marketing disagree about what a good lead is, a score is a way of avoiding the conversation rather than having it. That one is in the marketing and sales handoff, and knowing which leads to decline outright is in the leads you should turn down.