Your health score was green. The customer churned anyway.
That's not a data problem. That's a measurement problem.
Most health scores are built on lagging indicators — metrics that tell you what already happened. By the time the score turns red, the customer has already decided to leave. The fix isn't a better formula. It's tracking the right signals in the first place.
Why Your Health Score Is Lying to You
Here's how it plays out. You built a health score. NPS survey results. Login frequency. Ticket count. It looked rigorous. Three months later, a green account churns. You investigate. Usage had been declining for six weeks before the score moved. The NPS from last quarter was fine — because it was from last quarter. The login count was steady — because they were logging in to export their data before leaving.
Every metric you tracked confirmed what had already happened. None of them predicted what was about to happen.
The score was accurate. It was measuring the wrong things.
NPS tells you how a customer felt when you sent the survey. Login frequency tells you they accessed the product — not whether they got value from it. Ticket count is the most misleading of all. Zero tickets can mean a delighted, self-sufficient customer. It can also mean someone who gave up asking for help and is now quietly preparing to leave.
The score is green. They're already gone mentally.
Bain and Company's 2025 retention research maps exactly how this plays out. Customers move through three stages before cancellation: early warning, where behaviour changes sixty to ninety days before; active disengagement, twenty to forty days before; decision-made, the final zero to twenty days. Most CS teams catch accounts at the decision-made stage. Save rates there are below ten percent. The accounts worth saving were visible — by the right signals — sixty days earlier.
What Leading Indicators Actually Look Like
The leading indicator that predicts churn is not a generic metric. It's the one action most correlated with renewal in your specific product.
Not logins. The specific thing a customer must do to extract real, repeated value from what they bought.
For a project management tool it might be new projects created in the last thirty days. For a CS platform it might be workflow completion rate. For a billing tool, the number of automated reconciliations run. Every product has a different answer — but every product has an answer. And that action's fourteen-day trend is more predictive than any composite score built on engagement proxies.
If users don't engage with core features in their first thirty days, they're sixty percent more likely to churn. Not because they're unhappy. Because the product never became part of how they work. That signal is visible in week three. Your health score probably isn't watching it.
Silence is a signal too. A customer filing frequent tickets is frustrated but engaged. A customer who stopped contacting you after month two has often already started looking for an exit. Zero proactive contact initiated by the customer for twenty-one days is one of the strongest leading indicators of passive disengagement — and it won't show up in your NPS.
The System That Actually Predicts Renewal
Five steps. Each builds on the last.
Reverse-engineer your churned accounts. Pick your last five churned accounts. Pull their usage data at thirty, sixty, and ninety days before cancellation. Find the pattern that was there before the score turned red. That pattern — not login frequency, not NPS — is your leading indicator. This takes an afternoon and requires no new tools.
Identify your value action. Ask: what is the one thing a customer must do to get real value from your product? That action's trend over fourteen days is your strongest predictor of renewal. Add it to your customer records this week. Everything else in your health score is secondary.
Replace lagging metrics with leading ones. Declining value-action trend over fourteen days. Onboarding milestone stalled for twenty-one or more days. Zero customer-initiated contact after month two. These three signals — not NPS, not login count — are what your health score should be built on. A customer can look perfectly healthy on the standard metrics while all three of these are moving in the wrong direction.
Set thresholds and flag early. When the leading indicator drops below the threshold, flag the account — before the overall score turns red. The intervention at week six is a conversation. The intervention at week fourteen is a save call. Completely different outcomes. Save rates at week six are above sixty percent. Save rates at week fourteen are below ten.
Recalibrate quarterly. Health scores that aren't recalibrated against actual outcomes lose up to thirty-five percent of their accuracy within six months. Compare your predicted health against what actually happened each quarter. If you're seeing surprises — green accounts churning, red accounts renewing — the model needs adjustment. The model that's right today is built on the customer behaviour of six months ago. That customer base no longer exists.
We built this on MatrixFlows — structured customer records with leading indicator fields connected to product data, with an AI agent that flags accounts at the early warning stage, not the decision-made stage. The flag fires at week six. The CS team makes a conversation, not a save call.
What to Do This Week
Pick your last five churned accounts. Pull their usage data sixty days before they cancelled. Write down when the behaviour that predicts value actually declined — not when your health score turned red, when the signal changed. Compare those two dates. The gap between them is the intervention window you're currently missing.
Then write down the single action in your product most strongly correlated with customers who renew. Add that action's fourteen-day trend to your customer records this week. For every account currently green in your health score, check the trend on that one indicator. If it's been declining for twenty-one days, the score is lying to you.
Your health score isn't wrong. It's measuring exactly what you told it to measure.
The question is whether what you told it to measure is what actually predicts renewal.
For most companies, it isn't.
Once you've rebuilt the signal, the next step is using it for expansion — the same leading indicator system that catches at-risk accounts surfaces the ones ready to grow. And if you want to understand why self-sufficient customers renew at ninety-five percent while high-touch customers renew at seventy-five, the math is here. MatrixFlows is free to start.
.svg)
.svg)
.svg)
