AI-Powered Customer Reactivation for Dry Cleaning Network

A Ukrainian dry cleaning network came to us with a critical business challenge: their most valuable customers were disappearing. Despite having built strong relationships with VIP clients (averaging 26+ visits and $1,100+ per visit), 77% of them hadn't returned in over 450 days.

The financial impact was severe:

• 5,437 VIP customers represented 56.5% of total revenue
• Average lifetime value: $28,928 per VIP client
• Churn was accelerating with no systematic approach to win them back
• Generic mass discounts weren't working and were destroying margins

Data Analysis (Week 1)

We started with a comprehensive analysis of the client's customer database:

• 33,000+ transactions analyzed across 7 years of operations
• B2B/B2C segmentation: Identified 305 corporate clients (5.3%) that needed separate treatment
• RFM clustering: Segmented 24,803 B2C customers into VIP (5,437), Loyal Core (10,714), and Newbies (8,652)
• Risk zone modeling: Categorized customers into GREEN, YELLOW, ORANGE, RED, and BLACK zones based on recency and value

Key Discoveries

• Point of No Return varies by segment: VIP customers lost after 163 days, Loyal Core after 415 days, Newcomers after just 29 days (90th percentile analysis of 192,159 transaction intervals)
• Optimal contact windows prevent waste: Contacting Loyal Core before 250 days wastes budget (they return naturally). Waiting beyond 29 days for Newcomers loses them forever.
• Seasonality matters: December and January show 40% lower return rates (6.7% vs 11.8% in October)
• Cross-sell opportunity: Found 156 product combination rules with lift coefficients up to 5.2x
• Shoe cleaning = VIP upgrade: Clients who add shoe cleaning have 5.6x higher LTV

Our Methodology

We used Python (pandas, numpy) for data wrangling, scikit-learn for clustering, and built custom RFM models to identify at-risk segments. The analysis revealed that a one-size-fits-all approach would fail — we needed personalization at scale.

Based on our analysis, we recommended a sophisticated, automated system that combines three ML components:

1. Uplift Model (Personalized Discount Optimization)

Instead of predicting "who will buy," we built a model that predicts "who will buy BECAUSE of the offer" — the foundational principle of Uplift modeling.

• Three-model ensemble: Natural Purchase Model (baseline), Conversion Uplift, Revenue Uplift
• Dual strategy: "Retention" (soft offers for at-risk clients) vs "Development" (upsell for loyal clients)
• Result: Each customer gets an optimal discount (10-30%) that maximizes profit, not just conversion

2. LLM Personalization Engine (GPT-4 + Structured Outputs)

Generic "We miss you!" messages don't work. We implemented Schema-Guided Reasoning (SGR) to generate truly personalized messages:

• Pydantic schemas force GPT-4 to analyze BEFORE writing (reasoning → message → metadata)
• Context injection: Each message uses client's purchase history, RFM metrics, and risk zone
• Tone adaptation: Warm (for forgotten clients), Professional (for recent), Urgent (for critical risk)
• Testable output: Structured JSON enables automated quality checks

3. Cross-Sell Engine (156 Product Rules)

We analyzed historical transaction patterns to identify which products are frequently bought together and built ready-made packages:

• Business Pack: Suit + Shirt + Tie cleaning (Lift: 5.2x)
• Winter TOTAL CARE: Outerwear + Leather + Shoes (Lift: 4.2x)
• Delicate Exit: Dress + Blouse + Evening wear (Lift: 4.7x)
• Integration: LLM automatically suggests relevant package based on client's history

4. Segment-Specific "Point of No Return" Strategy

Instead of generic recency thresholds, we built scientifically-grounded contact windows based on 90th percentile analysis of historical return patterns:

• VIP (Point of No Return: 163 days): Contact window 120-163 days. Median return time is 22 days, so if they haven't returned in 120+ days, something's wrong. Contact BEFORE the critical 163-day threshold.
• Loyal Core (Point of No Return: 415 days): Contact window 250-415 days. These are seasonal customers (median 110 days) who return naturally. Contacting before 250 days wastes budget.
• Newcomers (Point of No Return: 29 days): Contact window 14-29 days. This is the habit-formation window. If they don't return within a month, the habit never forms — they're lost forever.
• Result: ~1,200 high-quality contacts/day (vs 4,000 with generic thresholds), 3.0-3.5x ROI through precise targeting

We delivered a production-ready system across 6 implementation phases over 9-10 weeks:

Phase 1: Data Foundation & Uplift Model

• Built complete ETL pipeline with feature engineering (RFM, seasonality, cross-sell signals)
• Trained T-Learner Uplift model using CatBoost (two-model approach for treatment vs control)
• Implemented Natural Purchase Model to identify baseline behavior
• Created validation framework to measure AUUC (Area Under Uplift Curve)

Phase 2: Targeting & Segmentation Engine

• Timing Optimizer: Determines optimal send time based on recency and seasonality
• A/B test framework: Splits audience into Control / Treatment A (single message) / Treatment B (touch sequence)
• Risk zone classifier: Assigns each customer to GREEN/YELLOW/ORANGE/RED/BLACK
• Touch sequence builder: Maps risk zones to Viber/SMS/Call sequences (industry benchmarks: 1-2-3-3)

Phase 3: LLM Personalization Engine

• Designed Pydantic schemas with Schema-Guided Reasoning (SGR) approach
• Integrated OpenAI GPT-4 API with structured outputs
• Built prompt templates that inject client context, cross-sell data, and uplift signals
• Created validation layer to ensure message quality and schema compliance
• Generated 10,000+ personalized messages in production run

Phase 4: Cross-Sell Integration

• Loaded 156 association rules from historical transaction analysis
• Built CrossSellEngine with get_recommendations() and match_to_package() methods
• Integrated ready-made packages (Business Pack, Winter TOTAL CARE, Delicate Exit)
• Linked cross-sell suggestions to LLM message generation

Phase 5: Campaign Execution & Export

• Main pipeline (main.py) orchestrates all components
• Exports to CSV (for CRM), JSON (with reasoning), and tracking file (for post-analysis)
• Special export for top-500 VIP: phone, history, call script for manual manager calls
• Integrated with client's Viber/SMS delivery system

Phase 6: Analytics & ROI Dashboard

• save_campaign_baseline(): Snapshots pre-campaign state
• collect_post_campaign_orders(): Tracks new orders after 30/60/90 days
• analyze_campaign_effectiveness(): Matches returning clients, calculates conversion & ROI
• Interactive Plotly dashboard for real-time monitoring
• Cross-sell success rate tracking (did they buy the recommended package?)

Project Scope

• 275 hours of development across 6 phases
• 17 Python modules with full testing coverage
• 4 Jupyter notebooks for training, analysis, and reporting
• 90% automation rate — system runs with minimal human intervention

Expected Outcomes (Based on Model Predictions)

• 2,695 customers returned out of 24,803 (10.9% conversion)
• 6.09M UAH revenue generated (two-wave strategy)
• 39.6x ROI (revenue vs campaign costs + development)
• 150,000 UAH campaign costs (Viber/SMS + manual calls for top-500)
• 15-20% cross-sell uptake on recommended packages

1. Point of No Return > Generic Recency

The breakthrough insight: each customer segment has a different "Point of No Return" — a statistical threshold (90th percentile) after which they rarely come back. VIP: 163 days, Loyal Core: 415 days, Newcomers: 29 days. Generic "82+ days" thresholds waste budget on Loyal customers who'd return naturally, and miss Newcomers entirely.

2. Uplift > Prediction

Traditional ML models predict "who will buy." Uplift models predict "who will buy BECAUSE of your action." This shift in thinking prevents wasted discounts on customers who'd return anyway.

3. LLM with Structure = Production AI

Free-form GPT-4 is creative but unreliable. Schema-Guided Reasoning (SGR) forces the model to follow a reasoning process, making outputs testable and consistent. This is the key to using LLMs in production.

4. Scientific Segmentation Drives ROI

By analyzing 192,159 transaction intervals and calculating median/percentile return patterns, we built contact windows that maximize ROI: fewer campaigns (~1,200/day vs 4,000), higher quality, 3.0-3.5x ROI.

5. Automation Enables Scale

Personalizing outreach to 24,803 customers manually would require a team of 10+ marketers for weeks. Our system does it in under 5 seconds per customer, with higher quality and consistency.