Lead scoring without a data team

Two dimensions, not one

The mistake in most first attempts is scoring only engagement. Someone who opened four emails scores highly — even if they are a student researching a project and will never buy.

Score two things separately:

• Fit — how closely they match your ideal customer. Industry, company size, role, geography, tech stack. This is knowable before they do anything.
• Engagement — what they have actually done. Visited pricing, downloaded something, replied, booked a call.

Keeping them separate is what makes the score useful. High fit with low engagement means market to them. Low fit with high engagement means be polite and move on. High on both is where sales should spend the day.

A model you can build today

Points are arbitrary until calibrated against real outcomes. Start simple and adjust.

Negative scores matter as much as positive ones. A model that only adds points ranks everyone who has ever existed above everyone new. Decay and disqualification are what keep the list honest.

Calibrate against closed deals, not opinion

Here is the step almost everyone skips. Take your last 20 closed-won deals and your last 20 closed-lost, and run them through the model retrospectively.

If your won deals do not score meaningfully higher than your lost ones, the model is measuring something other than buying intent. Adjust the weights until it separates them. Twenty of each is enough to start — you are looking for a clear gap, not statistical significance.

Where the data quality problem bites

Scoring is arithmetic on your CRM fields. If those fields are wrong, the score is wrong with total confidence.

Three failure modes we see constantly:

• Missing firmographics. You cannot score company size if the field is empty on 60% of records. The model silently treats unknown as zero, and good leads sink.
• Duplicates. The same person split across three records has their engagement divided by three. All three score too low to action.
• Stale job titles. Someone scored as a decision-maker two years ago may have left. Contact data decays at roughly 22–25% a year.

This is why enrichment and deduplication come before scoring, not after. A scoring model on unclean data produces confident nonsense.

Does AI lead scoring actually help?

Honestly: sometimes, and less often than the marketing suggests.

Where it genuinely helps. With a few thousand closed deals, a model can find patterns a human would not guess — that a particular combination of company size, page sequence and timing predicts conversion. That is real, and it beats hand-set weights.

Where it does not. Below roughly 500 closed deals there is not enough signal, and the model will confidently fit noise. Many products marketed as AI scoring are a weighted average with a nicer interface. And every one of them inherits your data quality problems — a model trained on duplicated, half-empty records learns the duplication, not the buying behaviour.

Our own position: get the data clean, build the simple two-dimensional model, calibrate it against real outcomes. If you are still scoring thousands of leads a month after that, then look at automating it.

What good looks like

A working model means sales opens the CRM and knows who to call first, without a meeting to decide. If your team still sorts by date created, the score is not being trusted — and usually that is a data problem, not a model problem.

Last reviewed 2026-07-28.