Part I + Part II
AI-Powered Churn Prediction and Retention Activation
Predict churn risk, identify key churn drivers, and activate segment-specific, brand-aligned retention messaging.
So what: We can identify high-risk customers with solid discrimination and turn predictions into actionable retention communication.
1. Can We Classify Churn Well?
Results from tuned Logistic Regression and Random Forest, evaluated with imbalance-aware metrics.
Model Comparison
| Model | ROC-AUC | PR-AUC | Churn Precision | Churn Recall |
|---|
Recommendation by objective: RF for ranking quality, LR when catching churners is prioritized.
Performance Snapshot
So what: In this PoC both models are viable; at larger production scale we expect stronger ranking and calibration, with final model/threshold selection aligned to campaign cost vs missed-churn trade-off.
2. What Factors Influence Churn, and in What Direction?
Top Churn-Increasing Signals (LR Direction)
Top Churn-Reducing Signals (LR Direction)
Most Influential Features (RF Importance)
Caution: charge-related variables and tenure are correlated, so business actions should be validated with controlled tests.
So what: Contract type, tenure, and support/security consistently explain churn risk and directly inform retention strategy.
3. LLM Activation: From Risk Score to Retention Email
Score customers with churn model
Assign segment by profile and risk pattern
Use segment-specific prompt template
Generate Vodafone-style retention email
Segment Prompt Strategy (3 segments)
- Price-sensitive monthly: savings, plan flexibility, billing transparency.
- Service-support gap: reliability reassurance, support/security improvements.
- Loyal renewal: recognition, perks, premium continuation offers.
Sample Email (Support Gap Segment)
So what: This closes the loop from analytics to customer communication with consistent tone and higher relevance.
4. Brand Control and Safe Operations
Generation Controls
- Fixed structure (subject, intro, bullets, CTA, close).
- Tone constraints (friendly, clear, positive, professional).
- Forbidden content rules (no mention of model/churn prediction).
Validation Controls
- Automatic checklist pass/fail after generation.
- Auto-regenerate with targeted feedback if failed.
- Sample manual QA with brand/marketing approval gates.
So what: The solution is not only effective, but also governable and brand-safe for scaled deployment.
5. Recommendation and Rollout Plan
Recommended Operating Choice
- Use Random Forest for ranking and prioritization quality.
- Keep Logistic Regression as a challenger when recall-heavy strategy is needed.
- Set campaign thresholds based on outreach capacity and expected retention ROI.
30-60-90 Plan
- 30: pilot with one segment and controlled audience.
- 60: threshold optimization and A/B tests on offer framing.
- 90: multi-segment rollout with quality and uplift dashboards.
Scale-up Positioning
- This presentation demonstrates feasibility on a small PoC dataset.
- Next phase trains the same architecture on larger, broader historical data.
- Expected gains: stronger churn recall/precision balance, richer segmentation, and higher retention uplift per campaign.
So what: We can move from model insight to measurable retention impact with a low-friction phased rollout.