AI Lead Scoring: Prioritize Your Best Prospects

The problem scoring is trying to solve

If forty enquiries land in a week and you have time to properly chase twelve, the order you work them in decides your month. Most owners sort by whichever came in most recently, or by whoever shouted loudest. Both are poor proxies for who is actually going to buy.

Lead scoring is simply a way of putting a number on each enquiry so the list sorts itself. The AI part means the number is learned from what happened to your previous leads, rather than guessed by whoever set up the form.

Rules-based scoring versus learned scoring

These get lumped together and they are not the same thing.

Rules-based scoring is a points system you write yourself. Visited the pricing page, add ten. Company email address rather than a free one, add five. Outside your service area, subtract twenty. It is transparent, you can debug it over a coffee, and it works from day one because it needs no history.

Learned scoring takes your closed and lost records and works out which patterns preceded a sale. It finds relationships you would not have written down, including uncomfortable ones. It also needs a meaningful volume of past outcomes before it says anything trustworthy, and it will not explain itself as cleanly.

For most local businesses the right sequence is rules first. Get a simple points system running, use it for six months, and only move to a learned model when you have enough labelled outcomes to justify it.

The signals that carry real information

Scoring systems tend to collapse into two families of signal, and both matter.

Who they are

Fit signals: location relative to your service area, business type, apparent size, the service they asked about, how they found you. A referral from an existing customer and a cold click on a display ad are not the same lead and should never score the same.

What they did

Behaviour signals: pages viewed and in what order, time on the pricing page, whether they returned, whether they opened the quote, whether they replied, how quickly. Behaviour is usually the stronger predictor, because intent shows in actions before it shows in words.

Two signals are underrated by almost everyone. The first is speed of response to your first message, which predicts closing better than almost anything else. The second is negative signals, which most scoring models omit entirely: someone who has enquired four times over two years and never bought is not a hot lead, they are a habit.

What you need before any of this works

Lead scoring fails more often from missing groundwork than from a bad model. Before you start, you need:

• Outcomes recorded. Every lead marked as won, lost, or unqualified, with a reason. Without labelled outcomes there is nothing to learn from. This is the step almost everyone skips.
• One place where leads live. If enquiries are split across a form inbox, a phone pad, Instagram messages and a notebook, no scoring system can see the picture.
• Enough volume. A learned model needs hundreds of past leads with known results to find anything real. Below that, you are fitting noise, and a simple rules system will do better.
• A different action for each band. A score with no consequence is decoration. High scores get a call within the hour. Medium scores get a sequence. Low scores get an email and no chasing.

Three ways scoring goes wrong

It learns your habits, not your market. If you have always called the enquiries with a company email first, those will show the best conversion, and the model will faithfully recommend more of the same. It is measuring your behaviour, not customer intent. The fix is to occasionally work a random sample of low scores so the data contains a counterfactual.

Everything scores high. If eighty per cent of leads land in the top band, the system has told you nothing. A useful score spreads leads out.

It quietly writes off good customers. Score bands become self-fulfilling. Leads you never contact never convert, which confirms they were low quality. Keep a small floor of contact for low scores, if only to keep the model honest.

Starting small this month

You do not need a platform. Take your last hundred enquiries, mark each won or lost, and look for the three attributes that separate them. Turn those into a five-line points rule in whatever system already holds your leads. Sort tomorrow's enquiries by it, and record whether the top of the list closed better than the bottom. That single experiment tells you more than any vendor demonstration, and if it works you have a case for building something more sophisticated on top.

For a model that learns from your history, several hundred past leads with recorded outcomes is a sensible minimum, and more is better. Below that the patterns it finds are usually coincidence. A hand-written rules-based score works from day one and often performs comparably for a small business, so start there.

Qualification is a yes or no judgement about whether someone could ever buy: right area, right service, budget in range. Scoring ranks the people who pass that filter by how likely they are to buy soon. Scoring an unqualified lead is wasted effort, so run the filter first and rank afterwards.

Yes, and the common failure is self-fulfilling. If the system ranks a type of lead low and you therefore never call them, they never convert, which appears to confirm the score. Keep contacting a small random sample of low-scoring leads so you can tell a genuine pattern from a habit the model has copied from you.

You need one place where every enquiry and its outcome is recorded, which in practice usually means a CRM, though a disciplined spreadsheet works at low volume. The tool matters far less than the discipline of marking every lead won, lost or unqualified with a reason. That record is the raw material scoring depends on.