Nobody cancels out of nowhere. They stop logging in as often. They open the pricing page twice. Their card fails and the retry fails. They write a second ticket about the same problem, shorter and angrier than the first. They ignore three emails in a row. Then, weeks later, they cancel — and the cancellation survey asks them why, as if it were a mystery.
Churn is a signal problem before it's a customer problem. The signals exist; they're just spread across a product database, a helpdesk, a billing system and an email tool that don't talk to each other. Retention work is mostly the discipline of bringing them together and acting on them while there's still something to save.
The signals that matter
- Usage decay. Frequency dropping against the customer's own baseline (not the average), and depth narrowing to one or two features. The best single predictor in almost every product we've seen.
- Support signals. A second contact about the same issue. Negative sentiment. A ticket that closes without a resolution. Support teams sit on the richest churn data in the company and are almost never asked for it.
- Billing. Failed payments, downgrade-page visits, invoices opened but unpaid. Involuntary churn from card failures alone is routinely 20–40% of total churn in subscription businesses, and it's the easiest kind to recover.
- Engagement. Email and push interaction falling off a cliff rather than tapering.
- Lifecycle stage. A new customer who hasn't reached the activation milestone in the first weeks isn't at risk of churning — they're already gone and haven't noticed.
Scoring without over-engineering it
You don't need a data-science team to start. Weight the signals above, add them up, and pick a threshold. A customer with falling usage, an unresolved ticket and a failed card is at risk; you didn't need a model to tell you that, you needed the three systems in one view. The advantage of starting with rules is that everyone can explain a score, which means the people acting on it trust it.
Once you have a few months of outcomes — who actually left, who stayed — train a model on them and let it re-weight the signals. Calibrate it: a score of 0.8 should mean roughly 80% of customers at that score churn without intervention. Then leave it alone and review it monthly. The model is never the hard part.
The hard part is the same as it always was: doing something useful, quickly, for the right customer, before the score becomes a cancellation.
Match the intervention to the cause
This is where most retention programmes fail: one save flow for everything, usually a discount. But the signals tell you why someone is leaving, and the why decides the fix.
- Activation gap (never got to value) → a guided nudge to the milestone, ideally a human offering to set it up with them. Discounts are irrelevant; they don't want more of a thing they haven't used.
- Value gap (used it, stopped) → education, a check-in call, a relevant feature they missed. Ask what changed.
- Service failure (bad experience) → a make-good and a person, fast. The apology is the intervention.
- Billing failure → a dunning sequence: retry schedule, a plain email, an SMS, a one-tap card update. Recovery rates of 50–70% are normal when this is done properly.
- Price → an offer, and only here. It's the least common cause and the most common response.
And for the accounts that matter — the top few percent by value — none of the above. A named human, with the context already summarised, picks up the phone.
The discount reflex
Discounting at the cancellation screen trains customers to threaten cancellation. It also hides the real cause: a customer who was leaving over a service failure and stayed for 30% off will leave next month, and now you've paid for it. Use offers last, sparingly, and only when the cause is genuinely price.
Measuring it honestly
Two rules. First, cohort retention curves, not a blended churn rate — you want to see whether this month's new customers are staying longer than last month's. Second, holdouts: a slice of at-risk customers who get no intervention. Without a holdout you'll credit your save flow with every customer who was going to stay anyway, and you'll believe your own campaigns. Report saves, recovered revenue and cohort curves weekly, in the same dashboard your support and marketing numbers live in.
Where to start this week
Pick the three signals you can get today — usually usage frequency, open tickets and failed payments — put them in one list, sorted by customer value. Call the top ten. You'll learn more about why customers leave in those ten calls than from a year of exit surveys, and you'll have the beginnings of a playbook by Friday.
