Churn-risk watch

When to run

• Recurring: weekly digest across the book (recommended)
• On demand: "which accounts are at risk" / "how's the health of my book"

Computer scopes

• CRM: accounts, renewal dates, contact changes
• Support: ticket volume, severity, and sentiment trend
• Product usage: activity trends by account and by seat
• Email + meetings: engagement cadence, champion activity

Workflow

1. Compute per-account signals: usage trend (30/90d), support sentiment delta, champion or power-user inactivity, meeting cadence break, contact departures (bounces, title changes), renewal proximity.
2. Score into bands (stable / watch / at-risk) with the signal breakdown attached; renewal proximity multiplies severity, it doesn't create it.
3. For each at-risk account, attach the receipts: the sentiment-shifted threads, the usage curve, the departed champion's record.
4. Diff against last week's digest: new entrants, escalations, recoveries — the delta is the story, the full list is the appendix.
5. Recommend one intervention per at-risk account (exec outreach, success plan reset, training push) matched to the dominant signal.

Guardrails

• Autonomy: L0. Observes and reports; never contacts customers or edits records.
• Sentiment is a model judgment: show the underlying threads, invite disagreement.
• Digest visibility follows account permissions.

Output

Eval fixtures

• Set: 24 months of sandbox history for 80 accounts, 19 with known churn outcomes.
• Rubric: at-risk band recall on churned accounts at T-60 days; false-alarm rate on stable accounts; every flag evidenced.
• Ship gate: T-60 recall ≥ 0.7 with false-alarm rate ≤ 0.15.