CSMhealth-scoring · advanced
Customer health score that actually predicts churn
Last reviewed: 2026-05-23 · saves ~saves 30%+ of preventable churn/run
Our take
Most health scores are garbage because they're just composites of opinions: 'usage low + ticket high = red.' Real predictive health requires regression-tested signals tied to actual churn outcomes. Build one with AI assist, not without.
Tool stack
Steps
- Export 18 months of customer-level data: usage metrics, ticket history, NPS scores, billing health, exec changes, expansion events, churned vs renewed outcomes.
- Use Claude with the prompt below to identify which signals correlate with churn vs renewal.
- Build a weighted health score in Vitally with the top 5-7 predictive signals.
- Test against the last 6 months: does the score predict churn at >70% accuracy 60 days before churn?
- Iterate quarterly.
Prompts
Identify churn-predictive signals from historical data · Claude Sonnet 4.6
You are a data analyst. I'll provide customer-level data for accounts that churned vs renewed in the past 18 months. For each potential signal (login frequency, feature breadth, ticket volume, NPS trend, exec change, billing late, etc.), tell me: 1. Is it correlated with churn? (yes / no / unclear) 2. Effect size: weak / moderate / strong 3. Lead time: how many days before churn does the signal appear? 4. False positive rate: how often does the signal fire for accounts that didn't churn? Output as a markdown table, sorted by effect size descending. Recommend the top 5-7 signals to use in a weighted health score, with proposed weights summing to 100. Constraints: - Don't recommend a signal with > 30% false positive rate. - Don't recommend a signal with < 30-day lead time (too late to act). - Be skeptical: if data is thin, say so.
Pitfalls
- Don't include 'CSM sentiment' as a signal — too subjective.
- Recency-weighted matters: usage drop last 14 days predicts better than overall declining trend.
- If your dataset is < 100 churned accounts, predictions are noisy. Use heuristics.