One score, two questions

Churn scoring and expansion scoring get built by different teams, at different times, on different data. Customer success owns one, sales owns the other, and the two disagree about the same account often enough that people stop trusting either.

They disagree because they are reading the same inputs and drawing opposite conclusions from the middle of the range. Heavy usage plus rising support tickets is either a customer in trouble or a customer outgrowing their plan. Nothing in a one-directional model can tell you which.

One score, read at both ends, forces the argument into the open. An account cannot be a red churn risk and a green expansion candidate at the same time, and when the model says it is, that is a definition problem you want surfaced rather than averaged away.

Score against the account’s own baseline, not an absolute. A customer who logs in twice a week and always has is healthy. A customer who logged in daily and now logs in twice a week is the one you want flagged, and an absolute threshold will rank them identically. Almost every score that gets ignored fails here first.

Weight recency hard. Behaviour from nine months ago tells you about a relationship that no longer exists. A 30-day window with a 90-day comparison catches direction; a twelve-month average catches nothing in time to act on it.

Agree what an account is before you score one. If the CRM has three records for the same customer — and after two migrations it usually does — you are scoring fragments. This is unglamorous and it is the step that most often has to happen first. See CRM data quality for what that involves.

The expansion signals are the churn signals inverted, plus two the churn model does not care about: seat growth and use-case spread. A team adding users, or a second department appearing in the usage data, has made the buying decision already — the only question left is whether anyone notices in time to have the conversation.

Keep upsell and cross-sell separate in the model, because they need different evidence. Upsell — more of what they have — is justified by hitting the ceiling of the current plan. Cross-sell — a different product — is justified by a use case appearing that the current product does not serve. Scoring them as one number produces confident recommendations that the account manager cannot defend in the meeting.

Both are cheaper than acquisition, which is the usual argument for doing them. The better argument is that the evidence is already sitting in your systems and costs nothing to read.

A subscription business with five years of history had leads never contacted, opportunities abandoned mid-pipeline, and customers who had lapsed quietly. None of it was visible, because answering the question meant joining four systems by hand.

Every dormant record was scored against last contact, acquisition cost, previous value and the stage it stalled at, then surfaced as a working queue with the reason attached — refreshed each morning rather than exported once. 4,800 records re-scored, £240k of pipeline surfaced from data the business had already paid for.

The mechanism is the one described above: a score built from signals that already existed, attached to an action, and refreshed on a clock. Full case study.

Nobody owns the alert. The score fires into a channel and everyone assumes someone else is on it. If a red account does not create a named task with a date, the model is decoration.

It is built on what is easy to measure. Logins are easy and weak. Whether the customer achieved the thing they bought the product for is hard and strong. Weight toward the second even when it takes work to capture.

It is never recalibrated. A score that has not been checked against actual renewal and churn outcomes is an opinion with a number attached. Write the outcome back and review the weights quarterly.

It measures the account and ignores the person. Accounts do not churn, buyers do. A green account whose champion just left is a red account that has not been re-scored yet.