Build a Customer Health Score That Predicts Churn

    You are a customer analytics lead. Build a health score that predicts renewal, not one that describes how the account feels.

What I sell: {{product}}
Customer count and segments: {{base}}
Signals available to me: {{usage_support_billing_engagement_data}}
Churn I've had: {{who_left_and_when}}
Current health scoring, if any: {{existing_approach}}

Start with the failure mode. Most health scores are a weighted average of whatever data was easy to get, tuned until the colours look plausible. They then fail in the only way that matters: an account goes from green to churned with no amber in between. Build backwards from churn instead.

Part 1 — Learn from the accounts that left.
For each churned customer, ask: what was observable 90 days before they cancelled, and how did it differ from accounts that renewed? Look specifically at the leading signals, not the lagging ones:
- Usage trend, not usage level. A big account declining is at more risk than a small account that was always small.
- Breadth: how many people, teams, and features. Single-user, single-feature accounts churn easily whatever the volume.
- The champion: still there, still engaged, still responding?
- Support pattern: not ticket volume, but unresolved tickets, repeat tickets on the same issue, and — the strongest of all — a sudden drop in contact from an account that used to talk to us
- Meeting acceptance and email response rates
- Invoice behaviour: late payment, questions about the contract, requests for a shorter term
- Value proof: has anyone on their side ever articulated the outcome they got?

Part 2 — Build the score.
- Pick the five to seven signals with actual separation between churned and retained accounts. Say which of mine you cannot evaluate because I have not supplied the data.
- Weight by predictive power, not by importance-in-principle
- Set thresholds from the observed distribution, not from round numbers
- Add the overrides that beat any score: champion departed, budget owner changed, acquisition, a competitor evaluation. One of these should turn an account red regardless of usage.
- Segment separately. A single score across enterprise and SMB will be wrong for both.

Part 3 — Make it honest.
- Say what the score cannot see. If half your value is delivered in meetings, product usage tells you little.
- Define the action tied to each band. A score with no playbook attached is a dashboard, not a system.
- Set the review: check quarterly whether the accounts that churned were actually flagged, and retune. A score nobody validates drifts into decoration.