The big airline booking sites have spent twenty years optimising for a single user journey: type two airports, pick a flight, pay. As of 2026 that journey is starting to look quaint. Travellers now ask their phone things like “I've got £900 and four days in May, find me somewhere warm with good food and direct flights from Manchester” — and increasingly get a real answer back, followed by a real booking.
We've been testing the new wave of AI travel assistants for the past few months on real trips. The headline: they've crossed the line from gimmick to genuinely useful for several core tasks, while still failing on others in ways that haven't improved at all. Here's the honest map.
Where AI is now genuinely better than searching yourself
1. Open-ended discovery
The classic “where should I go?” problem has always been hard on traditional sites. You can't put “somewhere warm in November under £800 with direct flights from Bristol” into a search box and get sensible results. AI assistants do this well in 2026 — they reason over typical weather, flight availability, and average prices, and they come back with three or four plausible options. The discovery layer is a real time-saver.
What works: combining constraints (budget, weather, flight duration, activity type). What still doesn't: factoring in your existing flight-credit pile or knowing which airline has a sale on right now. Those need real-time data the assistants don't always have.
2. Reading documents
Throw a screenshot of your boarding pass, hotel confirmation, or rental-car receipt at most modern assistants and they'll extract the times, addresses, confirmation numbers and the relevant fine print. This is small magic. The same goes for translating a foreign-language confirmation into English so you can spot the surprise non-refundable clauses. For frequent flyers this single feature has replaced 80% of what we used to do with TripIt.
3. Re-planning after disruption
This is the genuinely new capability. When a flight cancels, the best AI assistants in 2026 can simultaneously: look up the next available flights across multiple carriers, summarise your compensation entitlement under UK261, find hotels near the airport, and draft the email to the airline asking for re-routing. The whole thing takes ninety seconds and beats sitting on the phone for an hour.
We tested this on a real cancellation at Amsterdam Schiphol in February. The assistant correctly identified that the cancellation didn't qualify as an extraordinary circumstance (technical fault), produced the correct compensation amount, named the regulation, listed three alternative flights with actual price quotes, and proposed two hotels within 4km of the airport. Total time from cancellation announcement to having a plan: 4 minutes. Old-school passenger next to me at the desk: still in the queue.
4. Trip-specific research synthesis
For complex destination research — visa rules, vaccination requirements, plug types, tipping norms, current border policies — AI assistants now reliably produce summaries that previously required reading half a dozen government websites and travel forums. They're not perfect, and you should always cross-check anything safety-critical (visa rules, immunisation, money rules in volatile regions), but the time saving on routine research is real.
Where AI still gets things wrong
1. Live availability and prices
The single biggest current limitation. Many assistants will confidently tell you that a flight is £142 when, by the time you click through, it's £230. The booking-engine integrations are getting better — some now talk directly to airline APIs — but most still rely on cached data or scraped meta-search. The rule of thumb: trust the AI for shortlisting, trust the airline website for the final booking.
2. Loyalty awards
AI assistants don't yet understand airline-alliance loyalty programmes well. Asking “what's the best use of 100,000 Avios for a holiday in May?” produces answers that are usually wrong about the actual availability — points award seats are typically not visible through the public-facing data the assistants use. Award redemption remains a domain where a human with a points-blog browser tab open still wins.
3. The local context
Where to eat dinner in a small Italian town isn't something an AI assistant should be your only source on. The data underlying the answer is whatever ended up on the open web — which trends toward over-reviewed, over-touristed venues. The assistant will list five restaurants, all of which were good five years ago but are now the “Instagram-discovered” spots where the locals stopped going. A real local recommendation still beats an AI one for hyper-local choices.
4. Picking the wrong airport
A common failure mode: the assistant cheerfully books you into Beauvais (BVA) when you wanted Paris, or Stansted (STN) when you wanted London-anywhere-but. The cost saving on the flight is wiped out by the £50 ground transfer. The fix is to always specify the airport, not just the city, when prompting. The newer assistants are starting to surface the transfer cost automatically, but many don't.
How to use AI travel assistants without getting burned
- Use them for discovery and shortlisting, not final booking. Confirm prices and availability on the airline or hotel website itself.
- Always cross-check safety-critical info. Visa rules, vaccination requirements, customs and currency restrictions belong on a government website at the end of the chain.
- Be specific about constraints. Airport (not just city), departure-time window, baggage requirements, whether layovers are acceptable, and whether you need refundable.
- Use them as a translator and reader at the airport. Photographing a foreign-language menu, confirmation or signage and asking “what does this say and is there anything I should worry about?” is the killer travel-AI use case.
- Use them during disruption. The combination of speed, regulatory knowledge and parallel search is genuinely more useful than the airline's desk queue.
The Boarding Time view
Two thoughts for the long run. First, the right model is the AI does the routine work and the human keeps the judgement work. The assistants are excellent at squeezing time out of admin (reading confirmations, drafting emails, summarising rules). They're not yet excellent at the parts of travel that depend on taste, mood and surprise — which is most of what makes a trip memorable. Don't outsource the soul of the trip to the model.
Second, the underlying data quality matters more than the assistant's reasoning capability. The reason we've been steadily building structured, agent-accessible travel tools on Boarding Time — and publishing them with well-known agent skills — is that the better assistants in 2026 are the ones that know which authoritative source to ask, rather than the ones that try to know everything themselves. The quiet trend of the year is that good travel sites are starting to publish to AI assistants in addition to humans, which makes the assistants better and the sites more discoverable. Expect more of that in 2027.