Why customers churn before your data shows it

The industry has the causality backwards

Open any customer success platform and you will find the same taxonomy: usage decline, feature narrowing and login gaps are the leading indicators; cancellation and revenue churn are the lagging ones. Track the leading indicators, the reasoning goes, and you get ahead of the problem.

Leading indicator of what, though. They are leading indicators of the cancellation event. They are lagging indicators of the decision — and the decision is the only thing you could have influenced.

In most B2B environments, customers do not gradually reduce usage and then conclude, at the bottom of the slope, that they should leave. More commonly the underlying decision forms first — explicitly or implicitly — and observable disengagement follows. This is not worth the effort. This stopped doing the job. The person who championed it left. The usage curve bends afterwards.

What your dashboard charts, in other words, is rarely deliberation. It is disengagement following a verdict. And in many organisations that verdict lands well before an account ever turns red.

Decisions sit in a queue before they can be executed

Organisations rarely cancel software the day they decide it is no longer worth keeping. The decision spends weeks or months in an administrative queue — budget cycles, renewal dates, procurement windows, internal alignment, the contract that does not come up for another two quarters. Nothing can happen until the calendar allows it.

That lag is what creates the illusion that usage predicts churn. During the queue period the customer keeps logging in less, keeps narrowing their feature use, keeps going quiet. It looks like a decision forming in slow motion. It is a decision already made, waiting for the paperwork.

Which also explains why the interval is so inconsistent across companies. It is not a behavioural constant. It is the length of your customers' renewal cycle.

Telemetry cannot see the threshold

Ask how much warning a health score really buys you and the published answers scatter. Some analyses put the earliest detectable signals at sixty to ninety days before cancellation. Others describe customers disengaging weeks before any score moves. Practitioners writing candidly about their own dashboards describe the same pattern from the other side: the score was green, the renewal probability was high, the customer left anyway.

The estimates disagree because almost all of them infer the decision date from behavioural data. That is circular. You cannot use the usage curve to establish when the decision happened if the question is whether the usage curve arrives after the decision.

The distinction is simple and it is the hinge of this whole argument. Behavioural telemetry can identify observable disengagement. It cannot establish when the account mentally crossed the threshold. Those are different events, and only one of them is instrumented.

There is a related pattern worth sitting with. Numerous customer success benchmarks report that a large share of churning customers never open a support case or voice dissatisfaction before they leave. The exact percentages vary by source and segment, which is reason enough not to hang an argument on any single one — but the direction is consistent across all of them. Silent churn is not an edge case. Most customers who leave never file the complaint that would have told you something was wrong.

Why save plays keep missing

This is not an academic distinction. It determines whether your retention motion can work at all.

Every intervention in that sequence is aimed at month nine. The decision happened in month four. And the reason code that ends up in your churn analysis describes the final conversation rather than the cause, which is how a retention problem gets misdiagnosed as a pricing problem and answered with a discount.

Health scores are not the problem

None of this means the score is worthless. It means the score is answering a narrower question than the one it gets asked.

A health score does not answer will this account churn. It answers has disengagement become observable yet. That is a genuinely useful thing to know — it tells you where to send a CSM, which accounts need coverage, where onboarding failed at scale. It just cannot tell you whether you are early or late, because it has no view of the decision that preceded the behaviour it measures.

Treat it as an observation instrument rather than a prediction instrument and it stops disappointing you. The prediction has to come from somewhere else.

Renewal is another hiring decision

The clearest way to see all of this is to stop treating renewal as a continuation and start treating it as a purchase. Your customer re-hires you every renewal, against the same forces that governed the original decision — what pushes them, what pulls them, what they are anxious about, what habit now favours.

Read that way, churn is not a retention failure. It is demand collapse inside an existing account, and it obeys the same timeline as any other purchase decision: something changes, a verdict forms, action follows when circumstances permit. Bob Moesta puts the general case plainly — people do not switch because of an event, they switch because progress stopped. The usage decline you are watching is not the progress stopping. It is what happens after.

What to do: find the date, then calibrate to it

The fix is not a better model. Models trained on behavioural data will keep finding the behavioural signal, which keeps arriving after the fact. The fix is to establish, for your product, how long the gap actually is — and that requires asking people.

A firing interview reconstructs a cancellation backwards, from the moment they left to the point the product stopped doing the job. Done well it produces a date: the month it was really decided.

The questions that produce a date are the ones anchored to moments rather than opinions. When did you first start wondering whether renewing still made sense. What had changed in the business by then. Who raised it first, and what did they say. When was the last time you genuinely expected this to solve the problem. People cannot reliably tell you why they left, but they are surprisingly good at telling you when something shifted, and the when is what you are after.

Then go back to your own telemetry for that account and look at what was happening at that moment, before the obvious decline started. Whatever you find there is your genuine leading indicator, and it will be specific to your product rather than borrowed from a vendor's benchmark.

A dozen of these also surface something no dashboard can: what the customer was trying to get done, whether that job changed, and what they hired instead. An account that left because the job disappeared is a different problem from one that left because a competitor did the job better — and both look identical in a usage chart.

Figures on signal lead times and silent churn are drawn from customer success platform research and practitioner analyses published between 2025 and 2026; they vary widely, which is part of the argument above. The framing of switching as progress stopping is Bob Moesta’s. The claim that usage decline is a lagging indicator of the decision, the administrative-queue explanation for the gap, and the case for calibrating early warning against an interview-established date are ours.