Lead Scoring With Machine Learning: Beyond Points-Based Rules

Why the points system stopped working

Most CRMs ship with a scoring screen. Somebody in marketing sat down two years ago and decided a pricing-page visit was worth ten points, a job title containing 'director' was worth fifteen, and an email open was worth two. It felt sensible at the time.

Then the website changed, the sales team started selling to a new segment, and email opens became meaningless once mail clients began pre-loading images. The points kept adding up. Sales quietly stopped trusting the number and went back to reading every lead by hand.

That is the normal lifecycle of rules-based scoring. The rules are not stupid. They are frozen opinions about a market that keeps moving, and nobody owns the job of updating them.

What a machine learning lead scoring model actually does

A predictive lead scoring model looks at every lead you have closed, won or lost, and learns which combinations of attributes and behaviours preceded a win. The output is a probability: this lead has roughly a 22% chance of becoming a customer, that one 3%.

The difference from points is that nobody assigns the weights. The model finds them, including the uncomfortable ones. We regularly see that a signal the team loved, such as downloading a whitepaper, has almost no relationship with buying, while a dull one, such as the company's industry code or the number of people from the same domain who visited, carries most of the weight.

The data you need before it is worth building

The model can only learn from leads with a known outcome, so the first question is how many closed leads you have and whether the CRM records them honestly. A B2B software firm generating 150 inbound leads a month, with a 6% close rate, produces about 100 wins a year. That is enough to start, provided losses are marked as losses rather than left to rot as 'open'.

• Outcomes. Won, lost, and ideally a disqualification reason. Leads stuck in 'open' forever poison the training set.
• Firmographics at the time of the lead. Company size, industry, country. Enriched data captured later leaks future knowledge into the model.
• Behaviour before the first sales touch. Pages visited, forms submitted, event attendance. Anything recorded after a salesperson called is off limits for scoring new leads.
• Source and campaign. Often one of the strongest signals, and usually the messiest field in the CRM.

That third point catches almost every first attempt. If the model is allowed to see 'number of meetings booked', it will score brilliantly in testing and be useless on a lead that arrived ten minutes ago. We cover the general version of this trap in our guide to predictive analytics with machine learning.

How we build one, step by step

1. Export two to three years of leads with outcomes and freeze the data as it looked on the day each lead arrived
2. Build a baseline from your existing points score so there is something honest to beat
3. Train a gradient-boosted model, which handles messy tabular CRM data well, and hold back the most recent months for testing
4. Compare both on the held-back months: of the top 20% of leads each method ranks, how many became customers
5. Write the score and its top three reasons back into the CRM record, not into a separate dashboard
6. Retrain monthly or quarterly and alert when the score's accuracy on recent leads slips

The comparison in step four is the whole business case. If the model's top fifth of leads contains 60% of the eventual wins and the points score's top fifth contains 35%, sales can spend their mornings on a much shorter list. If the gap is small, we tell you, and the project stops there.

When rules are still the right answer

Machine learning is the wrong tool for lead scoring more often than vendors admit.

• You have fewer than roughly 200 closed outcomes, or you changed your product so recently that old outcomes describe a different business
• Your sales team is small enough to read every lead anyway, so ranking saves nobody time
• The real problem is routing (which rep, which territory), which is a rules job
• CRM hygiene is poor and nobody has the appetite to fix it first

In those cases we would rather tidy the rules, remove the points that no longer predict anything and look again in a year. It is also worth reading how AI-assisted lead qualification differs: a language model reading an enquiry can help a small team long before a scoring model has enough data.

Getting sales to trust the score

A score that sales ignores is worth nothing, and trust is mostly a design problem. Show the reasons alongside the number, in plain words: 'industry matches past wins, three visitors from the same company, arrived via partner referral'. Keep the score as bands (hot, warm, cold) rather than a two-decimal probability that invites argument.

Then run it quietly for a month before announcing it. Let the team see that the leads it flagged as hot really did close more often. At SpiderHunts, our machine learning team treats that shadow month as part of the build, because it is the step that decides whether the score gets used.

What it costs and what can go wrong

For a business with a reasonably clean CRM, a first lead scoring model including the CRM write-back usually takes four to eight weeks. Most of that is data preparation and integration rather than modelling. Running costs are small: a scheduled job and a few minutes of compute a day.

The failure modes are predictable. The model learns the habits of your best salesperson rather than the quality of the lead, because the leads they touched closed more often. It reinforces past targeting, so a new segment you want to break into scores low by definition. And it degrades silently when marketing launches a campaign unlike anything in the training data. None of these are reasons not to build one, but each needs a named owner and a monitoring check.

Frequently asked questions