Three approaches
| Approach | How it works | Strong when | Watch out for |
|---|---|---|---|
| Rules and health scores | Points or bands from usage, tickets, NPS and CRM fields | You want a score everyone can understand and audit | Averages that let one weak area hide; stale manual fields |
| Machine learning on outcomes | A model learns which patterns came before past churn | You have many renewals and clean history | Too few churn events to learn from; hard to explain a single score |
| AI that reads conversations | Language models read calls, tickets and email for risk | Most of the warning is in what customers say | Made-up or unsupported claims unless every risk must cite its source |
They aren't exclusive. A practical setup uses AI to find evidence in conversations, and plain rules to turn that evidence into a score people can check.
The data you need
- CRM: accounts, ARR, renewal dates, owners, and closed renewals (won and lost) to test against.
- Conversations: call recordings, support tickets, email.
- Product: seat and feature usage over time.
- Engineering: bugs and blockers linked to accounts.
- Billing: payment history, late invoices and requested downgrades.
How to test any churn prediction
Run it on last year's renewals as they looked before the renewal date, then check:
- Coverage of churned revenue (recall): what share of the ARR that actually churned was flagged 90, 60 and 30 days before the renewal?
- Precision: of the accounts flagged Red, how many actually churned or shrank?
- Lift: how many times more likely is a Red account to churn than the average account?
- Lead time: how early did the first warning appear?
Judge precision against your base churn rate. If 10% of your renewals churn and 30% of the accounts flagged Red churned, the score finds churn 3 times better than picking accounts at random.
Questions to ask a vendor
- Can a CSM see why an account is at risk, down to the call, ticket or email?
- How do you treat accounts with little data?
- Can we backtest on our own history before we switch?
- How will we know the actions we take actually changed the outcome?
Horizon combines AI that reads conversations with a rules-based health score, lets admins backtest it against past churn, and measures actions against a holdout group.
More on this topic: churn, definitions, guides and tools.
Updated November 3, 2026