The adoption gap
Travel and tourism supports 357 million jobs and moves $10.9 trillion a year, yet AI adoption beyond basic chatbots sits at just 15–25% — well below the 88% cross-industry average. The lag isn't reluctance. It's that generic LLMs weren't designed for an industry where an airline manages thousands of fare classes, a hotel juggles 20–30 interconnected systems, and a cruise line coordinates multi-week itineraries across jurisdictions.
A domain-fitness framework
Patil proposes five dimensions to evaluate whether a business area needs a specialized model:
- Proprietary terminology density — revenue metrics, rate codes, cabin designations
- Brand voice criticality — where personalization drives differentiation
- Data fragmentation — value locked in disconnected legacy systems
- Seasonal and temporal complexity — demand cycles requiring specialized reasoning
- Labor crisis and knowledge retention — capturing institutional expertise before it walks out the door
Hospitality scores high on all five.
Why fine-tuning beats retrieval
The core argument: embedding industry expertise inside model weights through fine-tuning outperforms general-purpose models augmented with retrieval pipelines. Self-hosted deployments also address the data privacy constraints that matter in regulated markets.
The proposed architecture uses specialized agents — revenue management, guest experience, operations — coordinating through an orchestrator that mirrors how hotels actually run. Revenue management is the clearest quick win: without sophisticated systems, hotels leave 2–5% of potential revenue on the table, and every one-point gain in guest satisfaction correlates with a $1.42 RevPAR lift.
Takeaway for operators
Generic LLMs plus retrieval will get you a chatbot. Domain-specific, fine-tuned models get you a system that actually understands the business. If AI is going to move hospitality from 15% adoption toward the 88% cross-industry norm, the models will need to know what a rate code is before they see one in a prompt.