AI lead scoring for sales, and when the number is just a guess

Lead scoring is an appealing idea because it replaces an argument with a number. Sales thinks marketing sends rubbish, marketing thinks sales does not follow up, and a score promises to settle it. That is a real benefit, and it is not the benefit most tools are sold on.

The question worth asking before you buy or build anything is narrower: do you have enough closed history for a model to learn from, and if not, whose history is the number actually based on?

The published threshold, which is higher than people expect

Salesforce documents the data requirements for Einstein Lead Scoring plainly, and the numbers are a useful reality check whichever vendor you are talking to. Its setup considerations state that at least 1,000 leads must be created in the last 200 days, and that at least 120 of those must have been converted to an account and contact.

Sit with those two numbers. A thousand leads inside roughly seven months, and a hundred and twenty of them converted. A business doing twenty leads a month does not reach the first threshold in a year, and a business with a long sales cycle may never reach the second on a rolling basis.

This is not a criticism of the product. It is arithmetic that applies to any model: a classifier needs enough examples of both outcomes to find a pattern, and converted leads are the scarce class.

What happens below the threshold is the part to understand

Salesforce is explicit about the fallback. When you do not have enough conversion data to build your own predictive model, Einstein uses a global model drawn from anonymous data across many customers, and switches to a model built on your data once you have accumulated enough, using whichever performs better.

That is a sensible engineering decision and it has a consequence nobody mentions in the sales meeting: below the threshold, the score is telling you what tends to convert for companies in general, not what converts for you. If your business is unusual in any way that matters, and most good businesses are, the number is confidently describing somebody else's customers.

This is the precise mechanism by which scoring launders a guess. The guess is not wrong because it is a guess. It becomes dangerous when it arrives as 87 out of 100 and a rep stops calling the 40s.

Rules first, model later, for a reason

A rules-based score that you wrote is worse at prediction than a trained model and better at almost everything else. You can read it, argue with it, and correct it on Monday when a rep tells you the rule is wrong.

The practical read across that table: until you are above the threshold, the honest tool is a short list of rules. Four or five signals, written down, agreed with the people who make the calls. It is not AI and it will outperform a borrowed model on the only metric that counts, which is whether reps trust it enough to change their order of work.

Where scoring genuinely earns its place

• The queue is longer than the team can work. Scoring is a triage tool. With forty leads a month and two reps, there is no queue to triage and the score changes nothing.
• Outcomes are recorded honestly. If closed-lost is used as a bin for everything that went quiet, there is no signal to learn.
• The definition of a lead has been stable. Change the form or the offer and you have changed what the history means.
• Somebody owns the threshold. A score with no agreed action is a decoration. The useful question is what the team does differently at 80 that it does not do at 50.

The second point is where most implementations quietly fail, and it is a data discipline problem rather than an AI problem. The same issue sits under every reporting project we have written about, including a real-time dashboard for a sales team: the number is only as honest as the field somebody updates at five o'clock.

The cheaper thing to build first

Before any scoring, build the thing that makes scoring possible later: a record of every lead with its source, its first response time, and its eventual outcome, in one place you can query.

That is unglamorous and it pays immediately, because the first query you run will probably show that response time predicts conversion better than any attribute of the lead, and that is actionable today without a model. If your leads currently live across a form inbox, a spreadsheet and somebody's phone, the first step is migrating from spreadsheets to a database, and for property specifically we have scoped this shape of system in a real estate lead management system.

It also means that when you do cross the data threshold, you cross it with clean history rather than two years of unusable records. That is the single highest-return decision in this whole subject.

The honest summary

If you have a queue longer than your team can work, clean outcome data, and volume in the published range, predictive scoring is a good buy and you should use what your CRM already offers rather than building it.

If you do not, scoring will give you a number that feels like knowledge and is not, and the cost is not the subscription. It is the leads nobody called because a model trained on other people's customers gave them a 40. Our general position on where AI helps in operations and where it adds a layer to debug is in how to use AI in operations without breaking the business.

And there is a prior question that scoring cannot answer: whether a lead is qualified at all is a definition problem, not a probability problem. The team at BDG Labs set out a definition that holds up under scrutiny in a B2B lead qualification framework that survives an audit, and it is worth settling before any model is asked to rank anything.