Building a churn early-warning score that people actually use

Churn rarely arrives as a surprise. In the accounts we analyse, usage starts to fall a median of 74 days before a cancellation notice. The signal is there. The problem is that most health scores are either too vague to trust or too complicated to act on, so they sit in a dashboard while customer success managers rely on gut feel.

Start with three inputs, not thirty

The temptation is to throw every field into a model. Resist it. A score built from a handful of understandable inputs gets used, while a black box with forty features gets debated. These three predict most voluntary churn in B2B SaaS:

• Active users as a share of seats purchased, compared with the account's own 90-day average.
• Frequency of the one or two actions that define value in your product, such as reports exported or invoices sent.
• Support tickets tagged as bugs or billing issues in the last 30 days.
• Optional fourth input: whether the executive sponsor has logged in this quarter.

Weight each input, sum to a score out of 100, and set clear thresholds. Green above 70, amber from 40 to 70, red below 40 is a sensible start. Check it against the last year of churned accounts. If most of them were red at least 60 days before leaving, the score is good enough to ship.

Explain every flag

A score of 38 means nothing on its own. A message that says active users fell from 42 to 19 in five weeks and two billing tickets were opened gives a CSM something to say on the phone. Always show the reason alongside the number, and link to the evidence.

A churn score is only as good as the conversation it starts on Monday morning.

Route alerts to a person

Alerts sent to a shared Slack channel are read by everyone and owned by nobody. Send each flag to the account owner, include a suggested next step, and review the red list in the weekly team meeting. Track what happened after each alert so you can tune thresholds every quarter.

Teams that follow this pattern typically see logo churn fall by 15 to 20 percent within two quarters. Not because the model is clever, but because it is simple enough that people trust it and specific enough that they know what to do next.