Three years is a long time in travel tech and even longer in the world of artificial intelligence (AI). Today, if a travel company is not working with large language models (LLMs) for some aspects of their business, people raise an eyebrow.
This wasn’t so much the case in July 2023. ChatGPT had only been announced the previous November.
People sat up and listened when Trip.com's co-founder and executive chairman James Jianzhang Liang stood up at a corporate event and said, “The opportunity for vertical LLMs is that artificial intelligence-generated content has met a tremendous challenge—their responses are not reliable.”
Liang was outlining the rationale for the company’s launch of a small language model (SML), Wendao, which roughly translates as “ask the ways.” Typically trained on far fewer data points than LLMs, they benefit from a tighter focus on a specific sector.
Wendao was trained on just 20 billion items of unstructured travel data, such as user searches and interactions and destination guides, as well as real-time pricing and schedule data. ChatGPT-3.5, by comparison, was trained on 175 billion.
The company said that Wendao, used in the Chinese market to power the company’s online chat, email and phone-based customer service, was saving the customer service team over 10,000 working hours daily, freeing up time to focus on presale activities.
Abhijeet Patil, co-author of the book “Practical Agentic Coding,”haswritten about the difference between frontier and domain-specific models: “A foundation model thinks it knows your industry. It does not. But it will answer anyway: confidently, fluently and incorrectly. That is hallucination ... A well-trained domain-specific model is the opposite. It knows its domain deeply, and just as importantly, it knows where its domain ends. Ask it something outside its expertise, and it can be trained to tell you honestly that it does not have the answer. That is not a limitation. That is wisdom.”
The biggest challenge for frontier models is that they have not been exposed to enough travel-specific data. Take the example of airline fare rules. These can require travelers to use a specific carrier and not codeshare partners, minimum stay requirements, ticketing requirements and cancellation penalties, among many other things. When this is coupled with the rapid pricing changes in travel, this is a headache for the mega LLMs.
The ‘secret sauce’
According to Matija Marijan, CEO of hospitality e-commerce platform Turneo, the major players are powerful but lacking in certain areas.
“The LLMs built by OpenAI and Anthropic are the most powerful out there. They're not the cheapest, but they're most powerful. They don't have access to all the data into the world. Our data and our workflows are the secret sauce.”
What Turneo is doing is not building a sector-specific LLM or SLM but rather a sector-specific harness that helps hotels upsell through pre-arrival emails to guests.
Turneo’s AI model proposes the email to send, but the hotel team can edit that. Every five emails, the AI re-runs its learning loop to learn from this human training to optimize the email.
“After a couple hundred emails, we'll bring down the frequency of rerunning its learning loop. It goes, ‘What can I learn from all the edits that our people are doing?’ You're implicitly training the harness,” said Marijan.
“If the guest engages by browsing a particular thing in email and then you send them a WhatsApp on the day of arrival, that data is inside the engine,” he said. “It's like having a person whose only job is to figure out how do I make the most money from the guest?”
A growing trend
Many travel companies are increasingly looking at such sector-specific models.
Booking.com’s AI trip planner was built by integrating OpenAI’s GPT models with the online travel agency’s proprietary data on properties, pricing and availability.
At its launch, Adrienne Enggist, the OTA’s senior director of product marketplace, said, “We’d spent years fine-tuning our structured data, like pricing, availability, cancellation policies. But now we could layer in unstructured data, like user reviews, natural language descriptions, and generate curated suggestions based on both.”
Narrow user cases can be particularly well suited to sector-specific models.
In March 2025, Microsoft CEO Satya Nadella announced the company had developed an app in conjunction with Fujitsu for Japan Airlines to help the cabin crew report onboard incidents, such as medical emergencies or delays.
Built on Microsoft’s Phi-4 small language model, optimized for use in offline settings such as onboard a plane, JAL-AI Report was fine-tuned using JAL’s past incident reports. Cabin crew enter keywords and phrases to quickly generate a report that is then handed over to ground staff. The airline said it cut the time taken to generate a report from an hour to 20 minutes.
Vacation rental metasearch engine HomeToGo has also embraced sector-specific models.
Group CTO Si Matthews said that the company has moved into SLMs because of the need for strong customer data and transaction control.
“We believe there's industry demand for sectorized LLMs, and that will come with the maturity of the AI market. When everyone stops poking fun at each other and gets together and understands the commonality of purpose, we believe that that will happen for travel.”
He said that the main driver for this is improved yield through things such as accelerating service response times and “all of those things that can just be done in a deterministic, non-human way so much faster to give a better experience.”
“I think we will see sectorized LLM and the value chain of that will be very deterministic based on the individual company. For us, we see value in having our own as much as being a part of the ecosystem as it matures.”
Ron Glickman, head of sales and marketing for Acai Travel, a PhocusWire Hot 25 Travel Startup for 2024, said that the sums might not add up for sector-specific models
“Every OTA has solved the customer conversation in some way,” he said. “Nobody has fully solved all types of servicing at scale, and that's where margins erode."
With all the complexity of travel, there could be a concern with adopting sector-specific models over liability, particularly in light of the EU’s AI Act.
Acai Travel CEO Riky Vittoria also pointed out a potential blocker for travel companies using their own specific models.
“The AI agent doing the work behind the scenes is the harder part. It has to speak fare rules and know DOT regulations, EU 261 and consumer law in every market. Get that wrong and the error isn't a bad answer, it's a fine.”