Travel & AI — hasn't this been solved? AMA w/ KAYAK's Chief Product Officer
AMA HOST WILL GO LIVE ON MAY 7TH @ 11AM EST
Hello everyone! Matthias here. As Chief Product Officer at KAYAK, I lead our AI initiatives and development of intelligent travel interfaces.
First thing I'll say is that AI in travel is deceptively complex. Many companies claim to have "solved" it, but most solutions fall short in critical ways - either they don't access real-time pricing, they hallucinate travel information, or they create frustrating user experiences.
At KAYAK, we've been working with AI since our founding. Our commitment to innovative technology has driven features like KAYAK Price Check and Ask KAYAK, plus collaborations with ChatGPT and Microsoft Copilot.
And today I'm excited to share that we've just launched KAYAK.ai - combining conversational AI with our comprehensive database of real-time prices from 400+ providers. What makes it different is that it's built specifically for travel, with accuracy at its core from a brand you can trust.
What I can discuss today:
The technical challenges we overcame to make KAYAK.ai reliable and accurate
How we designed a conversational interface that actually works for complex travel planning
The surprising insights we've gained from early user interactions
Why we believe this approach solves problems that traditional travel search doesn't
How AI is changing the way travelers search for and book trips
The challenges and opportunities that come with using AI for travel planning
Common misconceptions about AI in travel:
That it's just a voice interface slapped onto existing search
That large language models alone can solve travel planning
That accuracy isn't as important as convenience
That users don't want conversational interfaces for complex purchases
I'm here to have an honest conversation about where AI in travel currently stands and where it's headed. Ask me anything about our approach, challenges we've faced, or how we see the future of travel search evolving.
Looking forward to your questions and feedback!
Replies
1. Is it really possible for any travel company to offer a truly agentic flight search that finds the best prices for users, when the business is incentivized to maximize margins without losing customers?
2. With AI here , can we expect anything m in the visa/eVisa space - especially in simplifying the process for travelers?
KAYAK
@esafwan On 1.: The landscape of selling airline tickets has evolved beyond just the ticket itself. It now encompasses a range of ancillary services such as baggage options, insurance, upgrades to premium cabins, loyalty programs, airline credit cards, and an array of services designed to enhance the overall travel experience. It’s too early to tell but it does not appear that agentic search would negatively impact this.
Imagine a world where an agentic booking would seamlessly integrate your frequent flyer number, understand your seat preferences, and streamline the booking process, allowing you to save time while still enjoying a personalized experience.
On 2.: Travel visas are not our area of expertise however we all know that visa requirements can be complicated, and AI search machines are really useful for recommendations tailored to personal circumstances. Having said that, when it comes to matters like visas, I make it a point to personally verify information by cross-referencing with the source links.
I played with Kayak AI a few weeks ago and was very impressed. I'm curious to learn more about what's happening under the hood. I've seen other people claim that interfaces like this just reduce the natural language query to API params but I'd be very curious to understand the massive technical challenges that I'm sure had to be overcome to make this. Well done!
KAYAK
@steveb
At a high level, our system interprets a user's query and generates a structured response. This response is then analyzed and refined to deliver the requested search results, real-time information, and even create a personalized itinerary for the user.
Some of the technical challenges we’ve tackled include:
Prompt engineering: Ensuring we use the right terms when writing the prompts for the LLM to ensure effectiveness. A prompt that works well on GPT-3.5 may behave differently on GPT-4 or other models.
LLM Roundtrip Orchestration: A single request to a LLM is often not enough to answer a user’s query. Orchestrating multiple requests with intermediate responses is quite challenging.
Structured data: Crafting structured and semi-structured response expectations is key to effectively process LLM-replies (i.e. locations)
Intent disambiguation: Understanding and refining what users are trying to say when they use different terms and ensuring we address what they mean.
Error handling and moderation: We need (and use) safeguards to ensure our responses comply with policy and platform guidelines, while still directly addressing the user's query in a helpful way.
Third party model (and API) changes: since we are working with OpenAI, we have to make sure we are up to date with all the new model changes. Every model version upgrade has to go through a thorough testing.
Robust testing: We developed a range of testing strategies to ensure the quality and reliability of the user experience, covering everything from edge case handling to regression testing across prompt versions.
@elica_farjadian Very cool. Did you create internal tools for the testing and eval process or is there a third party product that does a good job of this?
KAYAK
@steveb internal for now.
Raycast
I love the idea of being able to riff with a travel AI on several prospective travel itineraries... but that would require the AI making suggestions and then having the ability to earmark or set aside ideas that I like but am not firm about yet. This could be solved with something like Claude's Artifacts — any plans to allow for this kind of "itinerary window shopping"?
KAYAK
@chrismessina Absolutely, we want you to be able to come back to your plans and easily share them with others. Right now, you can save individual results using the heart icon, and those become shareable within kayak.ai. Expanding what can be saved and shared from the chat itself is definitely on our list!
Product Hunt
I'm actually pretty curious why travel AI's don't feel more personal and remember previous interactions. Like if I'm flying from Japan to my family in the states, and the AI knew the different states my family lived in, and my trip budget. It could recommend me a cheaper trip by stopping in LA, waiting a day or two, then flying to a different state on budget airline with the added benefit of seeing more family members!
Is there a particular reason trip planners and booking AIs don't take this approach?
KAYAK
@gabe Great use case! Our first version focuses on keeping context within a conversation, like tracking the dates, destinations, and preferences you mention. We’re also exploring how additional layers of memory and personalization, such as recognizing travel habits or preferred airports and airlines, could help support tailored travel planning.
Product Hunt
Wonder if @tavishi_gupta1 from @Tern has some questions here?
Hey, Matthias, incredibly insightful. Your emphasis on reliability and deep travel-specific tuning in KAYAK.ai hits a crucial point. One thought.. have you explored adaptive preference modeling that evolves with a user's implicit signals (not just search patterns, but context switching, decision fatigue cues, etc.)? This could unlock a new class of personalization, especially for high-intent users planning multi-leg journeys.
KAYAK
@jason_linus Currently, KAYAK.ai leverages the same personalization signals as KAYAK.com, such as search history and bookings. Looking ahead, we're exploring ways to incorporate richer intent signals to deepen our understanding of user needs and continue to enhance the overall experience.
@elica_farjadian sounds good! Curious — when thinking about deeper intent signals, are there specific behaviors that stand out to you? Like time spent on certain results, filter tweaks, or redoing searches? I imagine it’s tricky to tell what’s real intent vs. noise sometimes.
Product Hunt
Why is this a misconception? What is different compared to existing search?
When I first tried using kayak.ai, "a voice/text interface slapped onto existing search" is what it felt like to me. But it's possible I was looking for a too specific itinerary -- a 2 day stay on a nearby island with a fair number of air BNBs but not many hotels.
So I tried using it to find a comfortable itinerary for a long-haul flight with my toddler this summer. It felt exactly like a voice interface on existing search. And, follow-up questions/prompts didn't seem to change its strategy, it just answered my questions without trying to improve its search results.
Here is the chat, if you can/want to see it: https://kayak.ai/chat/GSO81_MARoGPxD_zlYxc2Q
Sounds great, Matthias. Most travel AI tools I’ve tried feel either too basic or give outdated info. How does KAYAK.ai make sure the prices and suggestions stay accurate in real time?
I launched iPlan.ai a few years ago, an itinerary planner for every vacation, and sold my share at the beginning of the AI hype.