sboghossian Ops — Churn Risk Detector
Purpose
Early identification of at-risk accounts gives the customer success team time to intervene before a cancellation decision is made. Churn in B2B legal AI is expensive to reverse — once a firm has moved their matters to a competitor, switching costs are high. This skill defines the predictive model inputs, the risk scoring logic, and the graduated response playbook.
Input signals
The model draws on the following signals, evaluated per user (for consumer/SMB) or per tenant/account (for firm/enterprise):
| Signal | Weight | Churn indicator |
|---|---|---|
| Login frequency trend (last 30d vs prior 30d) | High | Declining login frequency |
| Messages-per-session trend (last 30d vs prior 30d) | High | Declining depth per session |
| Feature breadth (number of distinct features used) | Medium | Drift to single feature (over-reliance = fragility) |
| Support ticket volume (last 30d) | Medium | Rising ticket volume without resolution |
| NPS score (most recent) | High | Score ≤ 6 (detractor) |
| Plan downgrade history | High | Any downgrade in last 90 days |
| Time since last success event | High | >14 days since a completed draft, research, or matter milestone |
| Explicit churn-signal events | Critical | See table below |
Explicit churn-signal events (highest weight)
These events are strong leading indicators and should immediately spike the risk score:
| Event | Risk bump |
|---|---|
| User exports all their data ("Export all") | +25 points |
| User submits a delete-account inquiry | +30 points |
| User raises a billing dispute or chargeback | +20 points |
| User explicitly mentions a competitor by name | +15 points |
| User cancels a renewal reminder | +10 points |
Risk score output
The model outputs:
{
"userId": "<string>",
"tenantId": "<string>",
"riskScore": <0–100>,
"topSignals": ["<signal description>", "..."],
"suggestedAction": "<string>",
"evaluatedAt": "<ISO timestamp>"
}
riskScore is a 0–100 integer where 0 = no risk and 100 = cancellation imminent.
topSignals lists the top 3 contributing signals in descending weight order.
suggestedAction is derived from the tiered playbook below.
Response playbook
| Risk score | Action |
|---|---|
| 0–50 | No action. Monitor passively. |
| 51–69 | Flag in CSM dashboard. CSM may reach out at discretion. |
| 70–89 | CSM outreach required. Assign to account CSM; outreach within 48 hours. Goal: understand friction and remove blockers. |
| 90–94 | Discount or incentive offer. CSM calls with a concrete retention offer (discount, feature unlock, dedicated onboarding session). |
| 95–100 | Cancel-prevention flow. Escalate to senior CSM or account manager. Offer exit interview. Last-resort save: pause subscription instead of cancel. |
Integration
- Scores are recalculated daily per user/tenant.
- Scores are surfaced in the ops dashboard under "At-Risk Accounts."
- Score changes of ±10 or more in a 24-hour window trigger a Slack notification to
#csm-alerts. - Pairs with [[ops-nps-collector-in-chat]]: a detractor NPS score (≤6) triggers an immediate score recomputation.
- Pairs with [[ops-credit-burn-rate-watcher]]: an account approaching credit exhaustion without upgrading is a moderate churn signal.
Model calibration
The risk model should be recalibrated quarterly against actual churn outcomes:
- Compare predicted scores (30 days prior to churn) against accounts that actually cancelled.
- Adjust signal weights based on which signals were most predictive.
- Track false positive rate (high-risk accounts that did not churn) to avoid over-triggering CSM workload.
Related skills
- [[ops-nps-collector-in-chat]] — provides the NPS detractor signal
- [[ops-credit-burn-rate-watcher]] — provides credit exhaustion signals
- [[onboarding-upgrade-prompt-when-credits-low]] — the in-product retention touchpoint for credit-low users