AI personalization in hospitality: 3 use-cases that work today
Half the vendors promise an "AI-first programme", but 90% are just an OpenAI wrapper around an empty prompt. What actually works today (2026):
Use-case 1: Churn prediction
Gradient boosting on 8-12 features: days since last stay, stays per 12mo, avg spend per stay, engagement score, last-3-email-opened, dashboard-login. Output — "will churn in 90 days" probability. Typical accuracy — AUC 0.78-0.85 after 6 months of data.
What to do with it? Trigger an automated re-engagement campaign 30 days before predicted churn. Per industry data (McKinsey 2024) — re-engagement at the right moment adds 18-23% to retention rate.
Use-case 2: Next-best-offer
Collaborative filtering on the reward catalog + feature-based ranking. "Guests like you used SPA-day and loved it" — a simple recommender, but works better in hospitality than in retail, because there are fewer rewards (10-30 vs thousands of SKUs).
Effect: conversion on push with a recommended reward — 12-18% vs 3-5% generic push. That's 3-5× density per touchpoint.
Use-case 3: Review sentiment AI
Best use-case: classify reviews (TripAdvisor, Google) on 6 dimensions (room / F&B / service / cleanliness / value / amenities) with extraction ("complaining about smell in the 4th-floor hallway"). It gives the GM actionable insight without reading every review.
BERT-based models are enough (no supplementary deep learning). Accuracy in English — 0.85+ on labeled 10K reviews. Cheaper than any ML startup will "sell" you.
What NOT to do
- Don't replace human curators with AI agents. Luxury and upscale guests require human touch. AI is co-pilot, not replacement.
- Don't deploy "general AI" (Gemini / GPT-4 over PMS). Cost higher, accuracy lower than specific models, latency high.
- Don't start with computer vision (face recognition at check-in) — privacy nightmare, ROI unclear.
Architecture advice
AI features should be feature-flagged (toggle per tenant), explainable (admin sees "why the model decided that") and trainable (retrain at least monthly). In TTE we do exactly that — AI layer isn't "magic", it's a tool with a transparent API.
