By the time a renewal is on the calendar, the decision is usually made. The signals that predicted it were in support, usage, org changes and sentiment — scattered across systems nobody joined together.
Renewal forecasts tend to be optimistic until roughly six weeks out, when they collapse toward reality. That is not a modelling failure. It is what happens when the first real information arrives six weeks before the date — because the earlier information existed in systems nobody was joining together.
The signals below are ordinary. None require new instrumentation. What they require is being read together, on a timeline, against a definition of normal for that account.
Structural signals: the account changed shape
The champion changed jobs. This is the highest-value single event in retention and the one least likely to be tracked. The person who bought, defended the budget and absorbed the internal cost of adoption has left, and their replacement inherited a tool they did not choose, at a price they did not negotiate.
Sponsorship moved down. A quarterly review that used to include a VP now includes a manager. Nothing was said, but the account's internal priority has been restated.
The buying committee shrank. Fewer people from the account appear in threads and meetings over successive months. Usage may be stable while the relationship narrows to a single point of contact — which is a single point of failure.
Usage tells you whether the product is being used. The relationship tells you whether it will be renewed.
Why usage-only health scores mislead
Behavioural signals: usage changed in the wrong way
Total usage is a weak signal and a noisy one. What predicts is the shape of the change:
- Breadth collapsing — the account still logs in but has stopped using features it previously relied on.
- Concentration — activity that was spread across a team is now one person, often the same one in every thread.
- Workflow abandonment — a configured, integrated workflow silently stops running.
- Seasonality break — usage that has a reliable rhythm misses its cycle, which is a stronger signal than a gradual decline.
Sentiment signals: the tone changed
Support tickets are the most under-read retention dataset in most companies. What matters is not volume but trajectory: the same issue reported three times, escalation language appearing where it did not before, response-time complaints, and — the clearest of them — a ticket that mentions an evaluation or a competitor.
Meeting transcripts carry the same information earlier. A customer who has begun saying 'we need to justify this internally' has told you the renewal is contested, months before it appears anywhere else.
Commercial signals: the paperwork is talking
Invoice disputes, a shift from annual to monthly, a request for a shorter term, procurement appearing earlier than usual, or an unusual interest in the termination clause. Each is individually explicable. Together, and in the ninety-day window, they are a pattern.
Score the account, act on the signal
Health scores fail for a predictable reason: a number tells nobody what to do. An account at 42 produces a meeting; the specific fact that the champion left six weeks ago and nobody has met the replacement produces an action.
So keep the score for prioritisation and surface the contributing signals for action. The owner should receive the account, the top signals with dates, the history that makes them meaningful, and a suggested next step — not a colour.
The same signals predict expansion
It is worth noting that this machinery is symmetric. Breadth increasing, a new senior sponsor appearing, a team growing, a workflow being extended — these are the leading indicators of expansion, and they are the same data read in the other direction.
Teams that build churn detection and stop there have paid for the expensive half of the work and left the profitable half on the table.