Customers at risk of canceling usually show it first in behavior: they use the product less, use fewer of its features, hit more problems, or have billing trouble. Compare each customer's recent activity with their own earlier activity, score the signals, and review the highest-scoring accounts first.
Start with a plain points-based score you can explain. Move to a statistical model once you have enough past cancellations to test it against.
Warning Signals to Look For
| Signal | What to measure | Why it matters |
|---|---|---|
| Falling usage | Logins or key actions in the last 30 days versus the prior 30 | Declining use usually comes before cancellation |
| Narrow use | Number of distinct features used | Customers who rely on one feature have less to lose by leaving |
| Support friction | Open tickets, repeat contacts, long resolution times | Unresolved problems erode confidence |
| Billing trouble | Failed payments, downgrades, removed seats | Failed cards cause involuntary churn; downgrades often precede full cancellation |
| Engagement drop-off | Email opens, admin logins, champion leaving the account | In business accounts, losing the main user is a strong warning |
| Time to value | Days from signup to first meaningful result | Customers who never reach a first result rarely stay |
Which of these matter in your business is an empirical question: check them against customers who actually left. Is churn usually a symptom of poor adoption? Compare early usage of customers who stayed with those who left; if the gap is large, adoption is the lever.
Build a Simple Risk Score
Before any modeling, a points score gets you a working list. Illustrative weights:
| Condition | Points |
|---|---|
| Usage down more than 40% versus the prior 30 days | 3 |
| Uses only one feature | 2 |
| Two or more open support tickets | 2 |
| Failed payment in the last 30 days | 3 |
| Seats or plan reduced in the last 60 days | 3 |
Add the points for each customer and rank them. Review customers at 5 or more first. The weights above are starting guesses; adjust them against your own history.
Check That the Score Works
- Score the past. Take customers as they looked 60 days before a date, score them, then see who actually left in the following 60 days.
- Compare to the base rate. If 4% of all customers left but 20% of the top-scored group did, the score has lift of 5 in that group. If the top group churns at roughly the base rate, the score is not finding anything.
- Avoid leaking the answer. Do not use signals recorded after the customer started leaving (a cancellation survey, a final invoice) when scoring the past; the score would look accurate and be useless.
When to Move to a Model
Once you have enough past cancellations, a logistic regression or gradient-boosted model can weigh the signals from the data instead of from guesses. Keep the same test: score a past date, check the outcomes afterward. See churn prediction for the modeling side.
What to Do With the List
- Match the action to the signal: a failed card needs a billing fix, not a discount; low adoption needs onboarding help.
- Track whether contacted customers stay at a higher rate than similar customers who were not contacted; otherwise you cannot tell whether the outreach helped.
To measure the result, track churn rate by signup group with a cohort retention table.
Frequently Asked Questions
How can I predict which customers are at risk of canceling?
Compare each customer's recent behavior with their own earlier behavior: falling usage, fewer features used, rising support issues and billing problems are common early signs. Score those signals and check the score against past cancellations before acting on it.
How can I spot which accounts are at risk before they churn?
Rank accounts by a risk score built from usage decline, narrow feature use, open support tickets, failed payments and plan reductions, and review the highest-scoring accounts first.
Is churn usually a symptom of poor adoption? How would I tell?
Often, but check your own data. Compare early usage, such as time to first result and features used in the first weeks, between customers who stayed and customers who left. A large gap means adoption is the main lever.
How do I know whether my churn risk score works?
Score customers as they looked at an earlier date, then compare the share who actually left in the following weeks among the highest-scored group with the share among all customers. A score that finds a group churning at several times the base rate is working.
Do I need machine learning to predict churn?
No. A points-based score built from a handful of signals is a reasonable start and is easy to explain. A statistical model becomes worthwhile when you have enough past cancellations to weigh the signals from data and test the result.