Key takeaways
Booking isn't the bottleneck
At Phocuswright's Travel Marketing AI Summit in London, Skyscanner's chief AI officer Piero Sierra drew a line between two kinds of travel request. As PhocusWire reported, an AI booking a low-cost flight is starting to look realistic. The classic demo request, a five-star family holiday somewhere exotic, still falls apart.
It doesn't fall apart at the payment step. It falls apart because there are too many decisions to make before anyone gets near a booking button.
Travellers seem ready anyway. The same session cited Phocuswright research showing that 28% of UK travellers would already let an AI assistant book their flights and hotels. People are willing. The technology underneath isn't quite there yet.
How one simple request turns into thousands of searches
Take a request that sounds easy: a warm, seven-night family holiday in the next two months, somewhere with a good beach, for around £2,000.
Any AI chatbot can understand that sentence. Searching the market behind it is another matter. Before you have a single price, the request has already split into a pile of separate questions:
- Which destinations fit, and which nearby places are worth a look too?
- Which airports could you fly from and into?
- Which departure dates, return dates and trip lengths work?
- Which flights and fares go with each of those?
- Which hotels, room types and board options are available?
- Which suppliers sell each piece, and at what price?
Multiply those together and one sentence can mean thousands of combinations. Each one needs live prices and availability, and every live check costs time and money.
What is query fan-out, and why does it matter for travel?
The summit discussion also touched on query fan-out. That's when an AI takes one request and deliberately splits it into several searches, so it doesn't have to bet everything on one reading of what you meant.
For looking up information, that makes sense. For travel, it gets expensive fast, because every branch can set off another round of live supplier searches.
That leaves two bad options. Search everything and it's too slow and costs too much. Let the AI pick a few places and it may quietly throw away the best trip before you ever see it. The real work sits in between: deciding which parts of the search are worth running without losing the good options.
The missing layer: deciding what to search
AI travel products are often drawn as a straight line. You talk to the AI, the AI calls a supplier, the supplier takes the booking. For open-ended trips, that skips the hardest part. Something has to sit between understanding the request and paying for live searches.
Here's how we think that should work:
- Understand what the traveller is asking for.
- Separate the must-haves, like budget or school holiday dates, from the nice-to-haves.
- Come up with a wide enough set of destinations and dates that good options aren't missed.
- Narrow that set down using clear rules, search and ranking.
- Check live prices and availability for what's left.
- Rank the results and explain the trade-offs in plain language.
- Only then hand the chosen trip over to be booked.
Put simply: work out what to search, search it, check it, rank it, then book it.
Narrow first, then search
This is the problem we've been working on at MyEscapePlan. We don't ask a model to pick your holiday for you. We treat an open-ended trip as a search problem: expand the request into every reasonable destination, date and trip length, then drop the combinations that are unlikely to be worth pricing before the expensive live searches start.
In our September 2026 benchmark, across 12 flight and accommodation scenarios, that cut the destination and date combinations sent to live search by 67.7%. Of the options a full, price-led search ranked in its top ten, 67.5% were still in ours.
That doesn't mean AI travel booking is solved, and it isn't a booking conversion rate. It shows something narrower but useful: you can measure how much searching you cut and how much quality you keep.
That gives anyone building an AI travel agent a clear target. How much live searching can you skip while still keeping the trips most likely to be the best ones?
Live prices have to come before confidence
A model can give you a convincing reason why Mallorca suits a family. What it can't know is whether the flights are £90 or £490 on your dates, whether a suitable hotel still has rooms, or whether moving the trip by two days changes the price completely. Only live supply can answer that.
The AI should help decide where to look. Travel systems should confirm what's actually available. Only once prices and availability have been checked should the AI sound sure of itself.
The same request should get the same answer
Chat-based AI is unpredictable by design. Ask the same thing twice and you may get two different answers. That's fine for brainstorming. It isn't fine when money is involved.
If a traveller, a travel agent or another AI sends the same request twice, the system should be able to tell you why a destination made the list, why another didn't and which rule made the difference. Budgets, exact dates, airport choices and availability checks shouldn't depend on a model happening to say the same thing twice.
What works is a mix. Let AI handle the interpretation, and let clear, predictable rules handle the constraints and the transaction.
Booking still matters. It just comes last.
AI agents still need secure payments, sign-in, supplier connections, changes, cancellations and reliable follow-through. None of that answers the question that comes first. Before an agent says "I can book this for you", it needs a good answer to "why this trip?"
That matters most when the traveller hasn't chosen a flight, a hotel or even a destination yet. Plugging an AI straight into supplier systems makes a great demo. On its own, it doesn't solve the search problem.
We've written before about why AI travel tools still leave you doing the comparing yourself.
What a useful AI travel agent needs
The next wave of travel AI will probably look less magical than the demos and do far more work behind the scenes. It needs to:
- Understand what the traveller wants.
- Keep enough options open to leave real choice.
- Control how far the search spreads across destinations, dates and suppliers.
- Run live searches only on the combinations that matter.
- Check prices and availability against live supply.
- Rank the results and explain the trade-offs.
- Then book.
Skyscanner is right that agentic travel is on its way. For complex trips, the breakthrough won't be an AI that finally learns to press the Book button. It will be the search and decision-making underneath that makes pressing it a safe bet.