Building a Customer Health Score That Actually Predicts Churn

The most common health score mistake is building the formula around whatever data happens to be easy to pull: logins per month, ticket count, maybe an NPS response if you have one. The better starting point is a short list of churned accounts from the last year, examined one by one. What changed in the ninety days before each one left? For most B2B products it is a mix of usage depth (not just logins, but whether the account touched the features tied to their actual use case), support friction (repeated tickets on the same issue, or a ticket that sat open too long), and relationship strength (whether your original champion is still there). Build your signal list from that review, not from a template.

A common design flaw is weighting raw usage too heavily, because it is the easiest signal to instrument. Usage volume is a lagging indicator: it often stays flat right up until cancellation, because the people still logging in are not the ones deciding whether to renew. Support friction and stalled onboarding tend to move earlier and predict better. A practical starting split, before you tune it against your own churn history, is roughly a third usage depth, a third support and product friction, and a third relationship and sponsorship signals like champion turnover or a stalled expansion conversation. Adjust the split only after you have tested it against at least one full quarter of real outcomes.

A single blended score hides which lever to pull. Instead of one number, score each underlying dimension on its own scale first, then combine them into an overall score with documented weights. That way, when an account manager sees a red score, they can see immediately whether the problem is a stalled onboarding, a support backlog, or a champion who has gone quiet, and act on the actual cause instead of guessing. Keep the sub scores visible in whatever tool your team already lives in, whether that is a CRM view or a shared dashboard, so nobody has to open a second system to see why an account went red.

Before you trust a health score in production, run it backward against last year's renewals and churns. Pull every account that renewed and every account that churned or downgraded, then check what the formula would have scored them at three months and one month before the outcome. If churned accounts were not flagged red until the week they canceled, the formula is too slow to be useful, and you need earlier leading indicators, not just a lower threshold. If healthy accounts get flagged constantly, account owners will start ignoring the score within a quarter, which defeats the entire point of building one.

Backtest the formula in this order:

1. Pull every account that renewed and every account that churned or downgraded over the past year.
2. Score each one with the formula as it would have stood three months and one month before the outcome.
3. Check whether churned accounts were flagged red early enough for anyone to act.
4. Adjust the weights and thresholds if churned accounts were flagged too late, then run the check again.

A score with no attached action is just a dashboard decoration. For each band, red, yellow, and green, write down exactly what happens next and who does it: a red score might trigger an automatic executive check in within five business days, a yellow score might route to the account's CSM for a proactive call, and a green score might simply get logged as a candidate for an expansion conversation instead of a rescue one. Review the bands themselves every couple of quarters, because a threshold that made sense at fifty accounts often needs recalibrating once you have five hundred.