What is lead scoring, and can AI do it accurately?

How does lead scoring actually work?

A model learns which attributes and behaviors preceded conversions in your own records, then applies that to open leads. Company size, sector and source usually matter; so does engagement — opens, replies, pages viewed. The output is a probability, and the useful part is the ordering it produces.

The ordering matters more than the number. A founder with four hours and four hundred leads does not need to know that a particular lead converts 62% of the time. They need to know which twelve to call today. Scoring is a sorting problem wearing a probability costume.

This is also why lead scoring is a genuinely good fit for a small business. Large sales teams have capacity to work a whole list. A solo founder does not, so the cost of working the list in the wrong order is the entire value of the model.

How accurate is it?

Accurate enough to change what you do first, which is a lower bar than it sounds and the only one that matters. A well-calibrated model is right about as often as it claims to be. Treat a high score as a reason to call someone first, never as evidence they will buy.

Calibration is the property to ask about. A model that scores a hundred leads at 80% and sees eighty of them convert is calibrated. One that scores confidently and is wrong half the time is worse than an unsorted list, because it directs effort with false authority.

Any honest answer here includes the cold-start problem. A brand-new company has no outcomes to learn from, so early scores lean on general patterns and are weak. They improve as the business accumulates its own history. A vendor implying otherwise is describing something that cannot exist.

What makes lead scoring go wrong?

Three things. Too little history, so the model learns noise. Leaky inputs, where a field that only gets filled in after a deal closes makes the model look brilliant and useless. And drift, where your market changes and yesterday’s pattern quietly stops holding.

Leakage is the one that catches people. A model that includes "contract sent" as an input will predict closed deals almost perfectly and tell you nothing, because by the time that field is set you already knew. Anything only known after the outcome has to be excluded, and the check is boring and necessary.

Drift is the quiet one. Nothing breaks. The scores keep arriving, they keep looking reasonable, and they slowly stop being informative because the business started selling to a different kind of customer six months ago.

How should you actually use the scores?

As a work order, not a verdict. Call the top of the list first, and treat the bottom as deprioritized rather than disqualified. Check periodically whether the leads you were told to ignore ever convert — if they do, the model is mis-ranking and you would otherwise never find out.

That last habit matters more than it sounds. A scoring model creates a self-fulfilling loop: leads it ranks highly get attention, attention converts them, and the model looks accurate. Leads it ranks poorly are never called and never convert, which confirms the ranking without testing it.

The cheap correction is to work a small random sample from the bottom of the list every month. If those never convert, the model is doing its job. If they convert at a similar rate to the top, the scores are not informative and something is wrong with the inputs.

Keep the confidence figure visible in whatever you use the scores in. A ranked list with no uncertainty attached invites a founder to treat position three as meaningfully different from position four, when the model may consider them indistinguishable.

Where newc0 uses it

lead-scoring is one of two confirmed model domains newc0 deploys into the application it builds, alongside sales-forecasting. Both are described on the model library page, including what each will not tell you. In the morning report, the model’s output appears labeled as model output with a confidence figure attached, not as a fact.