A lead score sales will use: six decisions before the model.

Why the decisions come first

A scoring model is easy to build and hard to make matter. Most CRMs will produce a number for every lead. The question is whether a rep opens the queue tomorrow morning and works it in a different order because of that number. If not, the project produced a column.

When we are asked to fix a score nobody uses, the model is almost never the problem. The problem is that the team skipped the conversation about what the score is for, and went straight to which tool to buy. The six decisions below are that conversation. Each one takes a meeting, and each one changes what gets built.

1. What, exactly, does the score predict?

"Likelihood to buy" is not an answer. A model learns from examples of an outcome, so the outcome has to be something the CRM records reliably: a meeting held, an opportunity created, a deal won. Each choice produces a different score.

Predicting meetings is tempting because there are many examples, but it teaches the model to find people who take meetings, which is a different group from people who buy. Predicting won deals is what everyone wants and what most teams have too few examples of. Opportunity created, with a strict definition of opportunity, is often the honest middle. Whichever you choose, write it at the top of the page, because every later argument about whether the score is "right" comes back to it.

2. What will anyone do differently?

A score earns its place by changing an action. So name the action before the model: the top band is called within the hour, the middle band goes into a sequence, the bottom band gets nurture and no rep time. Or: the score sets the order of the inbound queue and nothing else.

This decision sets how precise the model needs to be. If the only use is ordering a queue, a rough ranking is enough. If the score decides who never gets a call, the bar is much higher and somebody should review the bottom band by hand for the first months. If no action changes at any score, stop here. You do not need a model.

3. One number, or fit and intent kept apart?

Two different questions hide inside most scores. Fit asks whether this company could ever be a good customer: its size, industry, tools and situation. Intent asks whether someone there is doing something that suggests a need now: visiting pricing, replying, hiring for the role you sell to.

Blend them into one number and a rep cannot tell a perfect-fit account that has gone quiet from a poor-fit student who downloaded everything. Kept apart, the two answers tell the rep what to do: strong fit with strong intent gets a call today, strong fit with no intent goes to outbound, weak fit with strong intent gets a polite reply and little else. Two simple scores are more useful than one sophisticated one.

4. Which data is the model allowed to use?

Three rules. First, only fields that are reliably filled in. A field that is blank on half the records does not teach the model the attribute; it teaches it which channel skipped the form.

Second, nothing that is recorded after the outcome. If "demo completed" is an input to a score that predicts opportunities, the model will look brilliant in testing and useless in practice, because it is reading the answer. It is a common reason a score tests well and then disappoints.

Third, nothing you would be uncomfortable explaining to the prospect. If the model leans on a signal that would sound arbitrary or unfair said out loud, take it out. A slightly less accurate score that everyone can defend will outlive a clever one that nobody can.

5. How does the score explain itself?

A bare number asks the rep for trust and gives nothing back. Every score should arrive with its reasons, in a sentence a salesperson would say: matches the profile, three people from the account read pricing this week, a new head of revenue operations started last month.

Reasons do three jobs. They give the rep an opening line. They let a rep disagree usefully, which is how you find out a rule is wrong. And they make the model's mistakes visible the first time someone reads a reason that makes no sense. Write the reasons to the record, next to the score, where the rep already works. A score that needs a second tab is a score that is not read.

6. How will you know it works, and who owns it?

Decide the test before the build. The simplest honest one: score last quarter's leads as they looked on the day they arrived, then compare what happened to the top band against the bottom band. If the top band did not convert clearly better, the score is not ready, however good it looks on a chart.

Then decide who looks at that comparison every month, because scores decay. Products change, markets shift, a new channel floods the top of the funnel with a different kind of lead. A named owner and a monthly check keep the score true. Without them the score drifts, and nobody notices until sales stops reading it.

What gets built once the six are settled

With the decisions written down, the build is short. The first two weeks go to the data checks, because decision four cannot be answered without them. The first version of the score, with its reasons, goes live inside the CRM in the weeks after, in front of a small group of reps who are asked to argue with it. Their objections are the cheapest testing you will get.

By the end of the engagement the score orders the queue, the reasons sit on the record, the monthly check has an owner, and the comparison from decision six has been run on live leads against the baseline taken at the start. The system is handed to the team that will run it, with the decisions page as its first document.

Try the decisions on the score you have

If you already have a score, you can test it against these six in an hour. Ask three reps what the number predicts and see whether the answers match. Ask what they do differently at 80 than at 40. Open five high-scoring leads and try to say why each scored high. If the answers are vague, the model is not the thing to fix first.