Predictive Churn System Design

Churn is not random. Users who leave show behavioural patterns before they go — reduced session frequency, declining feature engagement, increasing support tickets, shrinking usage depth. AI can identify these patterns at scale before any human would notice. The PM's job is to define what signals predict churn, design the interventions, and measure whether prediction and intervention together actually reduce churn.

---

Context

The three phases of a predictive churn system:

Phase	What it does	PM involvement
Predict	AI model identifies users likely to churn within [N days]	Define the prediction target and feature signals
Intervene	Product or team takes action on high-risk users	Design the intervention playbook
Measure	Did the intervention reduce churn for treated users?	Define the measurement methodology

The churn prediction trap: Most implementations get stuck at Phase 1 — they identify at-risk users but have no defined intervention. This skill designs all three phases.

---

Step 1 — Define the churn event

Define: what counts as churn, prediction window, current churn rate, and different churn types to track separately.

Step 2 — Define the churn signal features

Three signal categories:

Engagement signals: login frequency, session duration trend, core feature usage, feature breadth, days since last core action

Lifecycle signals: account age, onboarding completion, plan tier, billing event proximity

Sentiment signals: support ticket volume, CSAT/NPS, AI output negative feedback, cancellation page visits

Rank top 5 signals by expected predictive power.

Step 3 — Define the model approach

Start with rule-based scoring (weighted signals → risk score). Upgrade to ML (logistic regression → gradient boosting → survival analysis) when rules plateau. Minimum AUC-ROC > 0.70 before production use.

Step 4 — Design the intervention playbook

By risk level: HIGH (human outreach within N hours), MEDIUM (automated in-app nudge), LOW (standard engagement). Include intervention content principles and suppression rules.

Step 5 — Define the measurement methodology

Holdout control group (80% treatment / 20% control). Measure lift = (Control churn rate – Treatment churn rate) / Control churn rate. Track weekly and monthly metrics.

Quality check before delivering

Churn event is specific — not "user stops engaging"

Prediction window is defined

Signal features include lifecycle and sentiment — not just login frequency

Intervention playbook has specific message copy

Holdout control group is designed

Minimum detectable lift is defined before measurement begins

Suggested next step: Before building the model, audit your event logging. Every signal needs to be reliably logged with a timestamp and user ID. Spend one week confirming that all your top 5 signals are being logged correctly.