churn-prevention
When to use
- Net retention dropped and the team cannot name which of the four churn causes is dominant — defence-spending is uniform when it should be cause-specific.
- A health score exists but does not predict — it tracks usage but misses relationship and fit signals — and CS plays are running on bad triggers.
- A board ask names "are we losing customers we should have kept, or customers who never fit?" — the answer requires the four-way classification, not a single number.
Do NOT use to fix days 0–30 onboarding (route to
onboarding-design), drive upsell or expansion (route to
expansion-playbook), or build product-led retention loops (route
to retention-loops).
Cognition cluster
- Mental model 30 — Inversion. Do not ask "how do we keep this
account?" — ask "name the reason this account will leave." The
inversion forces a cause; the cause picks the move. See
docs/contracts/mental-models.md § 30.
- Mental model 16 — Leading vs. lagging indicators. Cancellation
is lagging; usage-decay, relationship-decay, and fit-mismatch
signals are leading. A health score built on lagging signals can
only confirm churn after the cancel request lands. See
mental-models.md § 16.
- Mental model 3 — Pareto (80/20). ~20 % of accounts carry ~80 %
of revenue risk. Uniform health-monitoring across the book is
theatre; weighted monitoring is reasoning. See
mental-models.md § 3.
- Context-spine — product + customer-segment. Read the
product slot for which capabilities the segment was sold
(value-churn lives here when capability and pitch diverged), and
the customer-segment slot for the segment's switch-event
patterns — fit-churn shows up early in segments whose switch
event differs from the ICP. See
context-spine.
Procedure
Step 0: Inspect — classify the last 20 churn events
Inspect the most recent 20 cancellation events. Tag each as one of:
- Involuntary — payment failure, dunning, card expiry. Not a value problem; an ops problem.
- Value — capability shipped does not match what was sold or what the buyer needs now.
- Relationship — champion left, sponsor change, exec turnover; product still fits, relationship does not.
- Fit — buyer was never the ICP; usage and pain never matched.
A book with > 30 % involuntary is an ops fix, not a CS fix. A book
with > 30 % fit is a marketing / qualification fix upstream, not a
CS fix.
Step 1: Define health-score signals per cause
One leading signal per cause, falsifiable, computable from
existing telemetry:
- Involuntary — payment-method age, dunning-retry depth.
- Value — feature-usage decay vs paid-tier ceiling (used / available), session length trend.
- Relationship — primary-contact response-latency, executive-meeting cadence vs contract baseline.
- Fit — segment classification at signup vs ICP; in-product behaviour mismatch (using read-only when sold workflow).
Health score = weighted aggregate per segment; weights derived from
Step 0's cause distribution. Do not average across causes — average
hides the binding signal.
Step 2: Set early-warning thresholds with confidence bands
For each signal, compute the historical threshold where the signal
flipped to a churn event within 60 days. Attach a confidence band.
A threshold without a band over-triggers in low-volume cohorts and
trains CS to ignore the alert.
Step 3: Map cause → play
Each cause gets one default play and one disqualifier:
- Involuntary — payment-retry + alternate-method outreach. Disqualifier: account in voluntary cancellation queue.
- Value — capability-gap interview; if real gap, route to product; if perception gap, route to enablement.
- Relationship — multi-thread outreach to second sponsor + exec sponsor injection.
- Fit — controlled wind-down; do not invest CS hours in saving a fit-mismatch account.
Step 4: Run the early-warning loop weekly
Weekly: pull accounts crossing threshold per signal; tag with
cause; assign default play; record outcome at +30 days. Outcomes
that do not match the play's expected lift become Step 1 signal
revisions next quarter — not next week.
Step 5: Hand back
Hand the cause-classification of the last 20 events, the
per-cause signal definitions, and the cause → play map to CS
operations and to
expansion-playbook for the
healthy-account expansion-trigger logic. Net retention work
without the cause split is spending money in random directions.
Related Skills
WHEN to use this
- Designing a churn-cause classification and weighted health score.
- Running the weekly early-warning loop and tuning thresholds.
WHEN NOT to use this
When the agent should load this
- "Why are accounts churning?"
- "Design a health score that actually predicts."
- "Classify last quarter's churn — value or fit?"
- "Welche Plays für Relationship-Churn?"
Output
churn-classification.md — last 20 events tagged with cause; cause-distribution percentages with bands.
health-signals.md — per-cause leading signal · threshold · confidence band · weight in aggregate health score.
cause-play-map.md — per-cause default play · disqualifier · expected lift at +30 days.
Gotcha
- A health score that aggregates without segmenting by cause hides the binding signal; CS plays based on the aggregate burn hours on the wrong account.
- "Engagement dropped" is not a cause; it is an observation. Engagement drops because of value, relationship, or fit — diagnose the cause before triggering a play.
- Fit-mismatch accounts surface as save targets when they should be wind-down targets. Saving a fit-mismatch account costs CS hours and produces a louder churn one cycle later.
Do NOT
- Do NOT run uniform CS plays across the book; weight by Pareto-risk-tier.
- Do NOT change health-score thresholds inside a quarter without an A/B holdout — concurrent changes destroy the signal.
- Do NOT invest save-cycles into accounts whose churn cause is fit; route to a controlled wind-down and tighten qualification upstream.
Runnable example
Mid-market SaaS, gross retention slipped from 92 % to 86 % over two quarters.
- Cause classification — of last 20 churns: involuntary 15 %, value 30 %, relationship 25 %, fit 30 %.
- Health signals — involuntary: payment-retry-depth ≥ 2 → 38 % cancel-in-60d (band 22–54). Value: feature-usage decay > 30 % MoM → 51 % cancel (band 38–64). Relationship: primary-contact silent 21+ days → 44 % cancel (band 30–58). Fit: ICP-classification ≠ ICP-purchased → 71 % cancel (band 58–82).
- Cause-play map — relationship-churn plays generate +18 % save-rate at +30 days; value plays generate +9 % only when paired with a product commit on the capability gap.
- Hand-off — classification + signals + plays → CS ops weekly; tightened qualification rule fed back to marketing (fit-mismatch upstream); healthy-account triggers handed to
expansion-playbook.