Enterprise
Travel companies build domain-specific AI to cut costs, boost accuracy
From Trip.com's small language model to Booking.com's fine-tuned GPT integration, travel firms are moving beyond generic LLMs to solve industry-specific challenges.
Key takeaways
- Trip.com's Wendao small language model, trained on 20 billion travel-specific data points, saves over 10,000 customer service hours daily in the Chinese market.
- Generic LLMs struggle with travel's complexity—airline fare rules, real-time pricing, and regulatory requirements often trigger hallucinations in foundation models.
- Companies like Booking.com and Turneo are building sector-specific "harnesses" that combine major LLMs with proprietary travel data rather than training models from scratch.
- Japan Airlines cut incident report generation time from 60 to 20 minutes using a small language model fine-tuned on historical airline data.
- Domain-specific models reduce liability risk in a heavily regulated industry where AI errors can result in fines rather than just poor customer experience.
Travel companies are increasingly building AI models trained specifically on travel industry data rather than relying solely on general-purpose large language models like ChatGPT, a shift driven by accuracy concerns and the unique complexity of travel commerce.
Trip.com was an early mover in this space. In July 2023, co-founder James Jianzhang Liang announced Wendao, a small language model trained on 20 billion items of travel-specific data—far fewer parameters than ChatGPT-3.5's 175 billion, but focused exclusively on user searches, destination guides, pricing, and schedule information. The model now powers customer service across chat, email, and phone channels in the Chinese market, saving more than 10,000 staff hours daily, according to the company, as first reported by PhocusWire.
The hallucination problem
Generic foundation models struggle with travel's intricate requirements. Airline fare rules alone can involve carrier restrictions, minimum stay requirements, ticketing deadlines, and cancellation penalties—all subject to rapid change. When a general-purpose LLM lacks exposure to this specialized data, it often generates plausible-sounding but incorrect responses, a phenomenon known as hallucination.
Abhijeet Patil, co-author of "Practical Agentic Coding," has written that foundation models "will answer anyway: confidently, fluently and incorrectly." A well-trained domain model, by contrast, knows its boundaries and can acknowledge when it lacks information.
Building sector-specific harnesses
Not every company is building models from scratch. Turneo, a hospitality e-commerce platform, uses major LLMs from OpenAI and Anthropic but wraps them in what CEO Matija Marijan calls a "sector-specific harness." The system proposes pre-arrival upsell emails to hotel guests, learns from human edits, and adjusts its approach based on guest engagement patterns.
"After a couple hundred emails, we'll bring down the frequency of rerunning its learning loop," Marijan said. The system tracks whether guests browse specific offerings and tailors follow-up messages accordingly.
Booking.com took a similar hybrid approach with its AI trip planner, integrating OpenAI's GPT models with proprietary data on properties, pricing, availability, and unstructured content like user reviews. At launch, senior director Adrienne Enggist noted the company had "spent years fine-tuning our structured data" before layering in natural language capabilities.
Narrow use cases show early wins
Sector-specific models prove particularly effective for focused applications. In March 2025, Japan Airlines deployed JAL-AI Report, built on Microsoft's Phi-4 small language model and fine-tuned with the airline's historical incident reports. The offline-capable app helps cabin crew document medical emergencies and delays, cutting report generation time from an hour to 20 minutes.
HomeToGo's group CTO Si Matthews said the vacation rental metasearch engine has moved into small language models because of "the need for strong customer data and transaction control." He anticipates broader industry adoption as the AI market matures and companies recognize common needs.
Why it matters
The shift toward domain-specific AI reflects a fundamental tension in travel technology: generic models offer power and convenience, but travel's regulatory complexity, real-time pricing volatility, and high liability exposure demand precision. As Acai Travel CEO Riky Vittoria noted, errors in an AI agent handling fare rules or consumer protection regulations "isn't a bad answer, it's a fine." Companies that build or fine-tune models on their own data gain accuracy advantages while maintaining control over sensitive customer information—a critical consideration as AI regulation tightens globally.
These details were first reported by Mark Frary for PhocusWire.