P4-34 AI: Churn Prediction

Predict which users are likely to churn before they cancel — and trigger retention flows.

---

Scoring Card

Dimension	Score	Rationale
Pain	3/5	Churn is the silent killer — most teams react after cancellation
Revenue	3/5	Retained revenue is the highest-ROI dollar; justifies enterprise tier
Build	2/5	ML pipeline, multi-signal feature engineering, threshold tuning
Moat	2/5	Model improves with data but churn prediction is a known ML pattern
Total	10/20

---

Classification

Vitamin AI Layer

Vitamin

Churn prediction transforms reactive customer success into proactive retention. Teams survive without it, but those who activate it catch at-risk users weeks before cancellation — turning a loss into a save. It is the AI feature most likely to justify enterprise pricing on its own.

---

The Pain It Kills

By the time a user hits the cancel button, it is too late. The decision was made days or weeks ago, and no last-minute discount can reliably reverse it.

• Churn costs 5–25x more to replace than to retain. Acquiring a new customer at $50–$200 CAC vs. sending a retention email at $0 cost.
• Most indie SaaS teams discover churn from their Stripe dashboard — after the fact.
• Teams that do track engagement signals do it manually: “check who hasn’t logged in this week” — unscalable and unreliable.
• No affordable tool connects engagement data, billing data, support signals, and NPS scores into a single churn probability.

---

What It Does

1. Ingest multi-signal data — engagement patterns (logins, feature usage, session depth), billing data (plan changes, failed payments), support tickets, NPS scores.
2. Train a churn prediction model — ML model (gradient-boosted trees or logistic regression) trained per-tenant on historical churn events.
3. Score every contact — output a churn probability (0–100%) updated daily.
4. Auto-trigger retention flows — when churn probability exceeds a configurable threshold (default: 70%), automatically enroll the contact in a retention sequence.
5. Surface feature importance — explain why this contact is at risk (e.g., “login frequency dropped 60% in last 14 days”, “NPS score declined from 8 to 4”).

---

Competition & What We Replace

Tool	Pricing	Limitation
ProfitWell Retain	Acquired by Paddle	Billing-focused, no engagement data, no multi-channel retention
ChurnZero	Custom pricing (expensive)	Enterprise-only, heavy implementation, 6+ week onboarding
Custom ML	$0 + engineering time	4–8 weeks to build, requires dedicated ML engineer to maintain
Baremetrics	$50–$500/mo	Analytics only — no prediction, no automation

GrowthOS churn prediction is integrated, automated, and affordable — it combines engagement data, billing signals, and NPS scores that no standalone tool can access without complex integrations.

---

Moat & Defensibility

Data breadth is the moat (2/5).

• Churn prediction ML is a known pattern — the algorithm is not the moat. The moat is having all the signals in one place: engagement events, billing data, NPS scores, support interactions, and contact metadata.
• Standalone churn tools must integrate with 3–5 systems to get the same data that GrowthOS already has natively.
• The model improves per-tenant over time — 3 months of data produces a useful model, 12 months produces a strong one.

---

Interoperability Advantage

The power of churn prediction in GrowthOS is that every input signal and every output action is already connected. No Zapier. No CSV exports. No manual workflows.

---

What Ships

• Churn probability score per contact — 0–100%, updated daily
• At-risk-of-churn segment — auto-populated, usable across all modules
• Auto-trigger retention flows — configurable threshold triggers email sequences, nudges, or Slack alerts
• Churn risk dashboard — tenant-level view of at-risk contacts, trends, and saves
• Feature importance — per-contact explanation of why the model flagged them (top 3 contributing factors)
• Historical accuracy tracking — model precision/recall over time, so tenants can see the model improving

---

What Does NOT Ship

• Custom ML model training — no tenant-facing training UI; model trains automatically on tenant data
• Win-back campaigns — use Email Sequences for post-churn win-back; churn prediction focuses on prevention
• Revenue impact forecasting — no “this churn will cost you $X” projections; we show probability, not financial modeling
• Real-time scoring — scores update daily, not in real-time; sufficient for retention use cases

---

Build vs Buy

BUILD.

No open-source churn prediction system exists with multi-tenant support and native integration to a contact graph. The ML component is standard (scikit-learn / XGBoost), but the feature engineering pipeline that pulls from events, billing, surveys, and contact data must be built natively.

Estimated effort: 5–6 weeks.

---

Dependencies

Dependency	Why
Contact Graph (P1-01)	Central data record for all contact signals
Lifecycle Emails (P1-03)	Retention sequences triggered by churn prediction
Stripe Integration (P2-31)	Billing data — plan changes, failed payments, downgrades
Surveys & NPS (P1-05)	Satisfaction signal — NPS decline is a strong churn predictor
Event data (3+ months)	Model requires historical engagement and churn events to train