Key takeaways
AI is becoming the front door to travel
The shift is already visible in traveller attitudes. Phocuswright research presented at its Travel Marketing AI Summit in London found that 28% of UK travellers would already let an AI assistant book their flights and hotels.
That does not mean traditional search disappears tomorrow. It does mean the first step of a trip is no longer guaranteed to be a search form. Instead of opening a flight or hotel form with a fixed destination and dates, a traveller who is willing to use AI can describe the outcome they want and ask it to narrow the possibilities.
Travel discovery is becoming a visibility problem
At the same summit, travel executives discussed the move from SEO toward generative-engine visibility. The important difference is that an AI answer does not behave like a directory of blue links. It can interpret the request, fan it out into several sub-queries and then present only a small shortlist.
For a travel company, that changes the commercial problem. A product can be technically available and still never reach the traveller if the AI never includes it in the consideration set.
Visibility matters because you first have to make the shortlist. But being visible is only the first layer of the problem.
Visibility is necessary. It is not control.
If an external AI decides which five hotels, flights, cruises, tours or destinations are worth considering, a travel company can end up pricing and fulfilling somebody else's shortlist.
The company may still own inventory, supplier relationships and the transaction. What it may no longer own is the decision immediately before those systems are called.
That matters because travel companies know things a general-purpose model does not inherently know: which products they actually sell, which contracts matter, which suppliers they prefer, which combinations work operationally, which customer segments convert and which rules must never be relaxed.
AI visibility asks: can the model see my products? Decision ownership asks: can my products, rules and customer context shape what the model decides is worth shopping?
A simple agent loop is not a travel decision system
The simplest agentic architecture is attractive: give a model a prompt, connect it to search tools and let it keep calling them until it has an answer. That is useful for prototypes and some bounded tasks. It is not, on its own, a production strategy for open-ended travel discovery.
Travel creates long-horizon problems. TRIP-Bench, a 2026 research benchmark for realistic travel-planning agents, uses 18 tools and more than 40 travel requirements. Its harder dialogues can involve more than 150 tool calls and more than 200,000 tokens of context. The paper reports that even advanced models fall below 10% success on some hard subsets.
That does not prove that every production travel agent will fail. It does show why 'give the model more tools' is not the same thing as controlling a complex decision process.
The problem is not just capability. It is reliability and control.
Travel has hard constraints. Dates can be fixed. Budgets can be binding. A family room cannot be replaced by a cheaper room that does not fit the party. A preferred supplier can matter commercially even when another result looks superficially similar.
A general-purpose agent may understand those instructions, but a production travel system also needs to prove that it preserved them, explain when it relaxed a preference and produce consistent behaviour when the same request is evaluated against the same data.
Travelport CTO Andrew Jordan recently described the same trust gap from another angle. Travelport's U.S. research found that among consumers already using AI for travel, 32% use it to research prices and options while only 8% use it to complete the booking, and around 72% say AI-powered booking causes stress or anxiety.
The implication is bigger than checkout. Fluent answers are not enough. The infrastructure behind the answer has to be accurate, bookable, serviceable and accountable.
Read PhocusWire's interview with Travelport CTO Andrew Jordan
Your products have to shape discovery
A travel company's product set should not be an afterthought that gets queried once an AI has already decided where the customer should go. Products should help shape the decision itself.
That means discovery needs access to structured information about inventory, rates, contracts, product attributes and commercial policy. It also needs traveller intent and, where appropriate, first-party context such as preferences and previous behaviour.
Spotnana's Steve Singh has described curation as a strategic problem as conversational interfaces show travellers only a handful of offers. Spotnana's stated target is to present an option the traveller would have chosen from the full list at least 95% of the time. He also argues that richer, accurate product information becomes more important as natural-language requests become more specific.
That is an important shift. The quality of an AI travel experience is no longer just the quality of the language model. It is also the quality, structure and governance of the product information feeding the decision.
Decision ownership means your rules survive the AI layer
For travel companies, a useful decisioning layer should combine several things before live shopping begins:
- Customer intent: what the traveller is actually trying to achieve.
- Your products: the destinations, routes, stays, experiences or other products the business can actually sell.
- Business rules: supplier priorities, commercial policy, eligibility and operational constraints.
- Traveller context: preferences, profile and relevant first-party signals where they can be used lawfully.
- World and market context: seasonality, reachability, affordability and other travel-specific signals.
- Outcome history: what verified successfully, what travellers pursued and where previous shortlists failed.
The model can still play an important role in interpreting open-ended language. But the decision should not disappear into a prompt. Hard constraints, product eligibility, ranking signals and relaxations need explicit ownership.
Search everything is not a strategy
Agentic AI can make generating possibilities extremely cheap. Live travel shopping is different. Every extra possibility can become another flight, hotel, package or supplier request.
Amadeus CEO Luis Maroto warned in July that AI could push look-to-book ratios higher, with more searches behind every booking. As Skift reported in September, Amadeus also says the costs of AI, including compute, search load, security and data governance, are arriving now, while the gains in conversion and personalisation may come more slowly.
A naive agent can therefore create the wrong economic outcome: generate more possibilities, fan out into more supplier calls and produce the same booking at a higher infrastructure cost.
The goal should not be to search everything faster. It should be to decide what is worth searching before expensive live shopping begins.
A stronger travel-AI architecture
For open-ended discovery, we think the architecture needs to separate understanding from decisioning and decisioning from verification:
- Interpret the traveller's natural-language request.
- Compile dates, budget, origin, party and other hard constraints into a structured contract.
- Bring the company's products, rules and relevant traveller context into the decision.
- Generate a broad enough candidate set to preserve real choice.
- Rank and prune candidates using reproducible travel-specific logic.
- Send only the candidates worth confirming into live pricing and availability systems.
- Use verified outcomes to rank the final options and improve future decisions.
This keeps AI where it is strongest - interpreting messy intent and communicating trade-offs - while keeping constraints, product eligibility and shopping policy under explicit control.
How MyEscapePlan approaches the decisioning layer
MyEscapePlan Discovery handles one part of that layer: deciding what is worth pricing. It sits upstream of existing search, pricing and booking systems and turns an open-ended travel request into a ranked set of destination, date and trip-length candidates before live inventory is requested. It does not replace inventory providers, and it makes no live supplier calls itself.
A model reads the traveller's sentence; a deterministic engine decides the answer. The engine uses typed constraints, versioned travel data, reproducible ranking signals and explainable scoring. Hard constraints stay hard, and soft preferences are relaxed only under explicit policy, with any relaxation disclosed. Same request, same data version, same ranking, which means a result can be inspected and tested.
In the canonical September 2026 Discovery benchmark, across four scenario families and the Flights and Accommodation tracks with five repeats per scenario, 40 requests in total, MyEscapePlan reduced the logical shopping space by 67.7% on average while retaining 67.5% Recall@10 against a price-led reference.
Those numbers need the right caveat. Logical shopping-space reduction is not the same as billable supplier-request reduction. Recall@10 measures agreement with a price-led reference rather than traveller fit, and at 67.5% roughly a third of the reference's top options did not survive. The value of the benchmark is that the trade-off is measurable: how much of the space can be removed, and how much relevant supply survives?
What changes over the next few years
Conversational discovery will become normal. Travellers will increasingly describe outcomes rather than fill in rigid forms. General-purpose assistants, specialist travel agents and travel brands themselves will all compete to become the interface where those requests begin.
That makes AI visibility a new distribution requirement. Product information will need to be richer, more structured and easier for machines to interpret.
But the strategic advantage is unlikely to come from visibility alone. If every company exposes the same product data to the same general-purpose models, the differentiator becomes what happens between intent and shopping: which products are eligible, how they are ranked, which constraints are binding and what the system learns from outcomes.
Complex travel will also continue to need accountable transaction and servicing infrastructure. The future is less likely to be one model replacing the travel stack than AI deciding more intelligently what the existing stack should be asked to do.
Questions travel companies should be asking now
- Can AI systems reliably see and understand our products, not just our marketing pages?
- If an AI creates the shortlist, which of our own rules and commercial priorities influence it?
- Can we show why one product was included and another was excluded?
- Are hard traveller constraints enforced by a controlled system or left to prompt interpretation?
- How much live search fan-out does one conversational request create?
- Can we measure shortlist quality before we optimise for fewer supplier calls?
- Do we retain the demand signals created before booking, or does an external AI own that part of the customer journey?
Take back control of travel discovery
Travel companies do not need to choose between ignoring AI and handing the entire decision to a general-purpose agent.
There is a third option: keep the decision about what is worth shopping inside systems you control, where your products, your rules and your customer context can shape it.
Visibility gets you into the conversation. Decision ownership lets your business shape the answer.