The Hardest Part of Travel Planning Isn't the Itinerary — It's the Decisions Underneath It
Generative AI has made itinerary generation nearly instant — but speed was never the hard part of travel planning. The harder part is the decision-making underneath it: trade-offs between cities, hotels, and pace that no traveler can fully specify upfront. This piece argues that the next real opportunity in travel AI isn’t prettier itineraries, but helping travelers understand the choices behind them.
For years, travel technology has treated planning as a generation problem: give a system a destination, dates, budget, and preferences, and it produces an itinerary. Generative AI has made that step dramatically faster — ask for ten days in Japan and get a full day-by-day plan in seconds. That's genuinely impressive, and I think it also solves the wrong part of the problem. The harder part of travel planning was never writing the itinerary. It's making a series of decisions under uncertainty, understanding the trade-offs behind them, and feeling confident enough to actually act on the result.
That distinction changes what a good travel AI should be built to do.
Discovery and Planning Are Not the Same Activity
People often say they enjoy planning trips, but what they usually mean is that they enjoy discovery — watching videos about a town in Italy they'd never heard of, saving hotels they may never book, collecting restaurants and neighborhoods and experiences that expand what a future version of the trip could look like. Discovery answers "where could I go, what looks interesting, what kind of trip do I want" — and it's enjoyable precisely because nothing has to be decided yet.
Share
Start a trip Fortrip can actually defend.
Sketch a brief. We come back with an itinerary that reads like a recommendation, not a checklist — and tell you what we’d change about it.
New research from 3,901 anonymized trip-planning conversations shows how travelers really use AI: opening messages under 30 characters, budgets and mobility needs that surface only after planning starts, and a verification step almost nobody asks for. France leads, Japan ranks sixth, and most planning happens on a Tuesday afternoon.
Three days, two nights, $230 total including flights, hotel, food, and entry fees — and it felt more restorative than a night at a beach resort. A full Sukhothai guide covering ancient ruins with zero crowds, sunrise at a hilltop Buddha, and exactly how to get there.
Planning starts later, once those possibilities have to become a real trip: five cities and eight days, a hotel you love that's far from everything else you want to do, a restaurant with availability on exactly one evening. Changing one city ripples into the hotels, trains, and reservations around it. At that point travel stops being a collection of interesting ideas and becomes a constraint problem — discovery expands the possibilities, planning has to reduce them, and reducing them means making trade-offs.
Travel Planning Is Fundamentally About Trade-Offs
There is rarely one objectively correct itinerary. Three nights in a city or four? A better hotel farther from everything, or a simpler one in a better location? One more destination, or a slower pace through fewer of them? Every choice improves one dimension while making another one worse — which is why I don't think the most useful question for travel AI is "how fast can we generate a complete itinerary." The more useful question is whether the system can help the traveler see the trade-offs behind it: what they gain by choosing one option, what they give up, which constraints are actually driving the recommendation, and how one decision changes everything downstream of it.
There usually isn't a perfect itinerary to produce. There's a better decision to help someone reach, given the real constraints of their trip — which is the premise behind building optimization and validation as their own layer at FortripAI, separate from itinerary generation itself: not a faster way to produce a plan, but a way to test and explain the trade-offs a fast plan tends to hide.
Speed Isn't the Same Thing as Control
Faster is usually better in AI product design, but travel is a high-consideration decision — someone may be spending thousands of dollars, using a meaningful share of their annual vacation time, and going somewhere they've never been. In that context, speed alone doesn't give a traveler what they actually want, which is a sense of control over a decision that size. The system itself can analyze options instantly; what the traveler needs is enough visibility into why a city got three nights instead of two, how a hotel choice affects daily travel time, or which parts of the plan are genuinely flexible versus locked in by something else. A travel AI should expose that decision space rather than collapse it into a single answer — not "here is the best itinerary," but a plan the traveler can see the reasoning behind.
The First Prompt Is Almost Never Enough
Itinerary generation often assumes the system already understands the traveler, and usually it doesn't. "I want to go to Japan for ten days" tells you very little — the preferences that actually matter (hating to rush, caring more about hotel location than luxury, wanting one countryside stop, being willing to wake up early to avoid crowds) tend to surface later in the conversation, sometimes only after the traveler reacts to something concrete. Few people can write a complete specification of themselves in the first message, which means conversational planning isn't just a different interface — it's a process of progressively discovering constraints the traveler couldn't have stated up front. The AI asks, the traveler reacts, a new preference becomes visible, and the plan changes. That loop is part of the planning itself, not overhead on top of it.
The Itinerary Shouldn't Be the End of the Process
This is also why I don't think a first generated itinerary should be treated as the finished product. Before seeing a plan, "I want a balanced trip" is abstract. Seeing a day with two museums, a market, a neighborhood walk, and dinner is concrete — and it's often the moment someone realizes "this is too much, I'd rather do fewer things and spend longer in each place." The itinerary revealed a preference the traveler couldn't have articulated beforehand, which is why the useful loop looks less like generate-and-accept and more like generate, react, adjust, challenge, refine — until the trip actually feels like the traveler's own rather than a plausible-looking draft.
Personalization Is Contextual, Not Just Historical
Most systems assume that knowing someone's past behavior tells you what they'll want next. Travel doesn't really work that way — someone who usually books a central luxury apartment in a major city might want a remote lodge in Costa Rica, nightlife in one destination and total isolation in the next. The traveler hasn't become a different person; the context has. Preference in travel is closer to a function of traveler × destination × trip context than a fixed profile, and historical behavior is only one input into it. A good system has to ask not just "what does this person usually like" but "what is this person likely to value here, on this trip."
AI Shouldn't Promise to Remove Uncertainty
Travel planning has a limit worth being honest about: AI can't make the rain stop, keep traffic predictable, or guarantee a restaurant stays open. Even good information here is often probabilistic, so I don't think a serious travel AI should promise certainty — its job is to help someone make decisions with uncertainty, by surfacing things like which time window usually has less traffic, which route is faster but less reliable, or which attraction is quieter early in the morning. Control, in this context, doesn't mean knowing exactly what will happen. It means understanding why you're making a decision and what might follow from it.
Where AI Is Actually Useful, and Where It Isn't the Differentiator
Social media is extremely good at creating desire — showing people places they didn't know existed, making them curious, helping them imagine a trip before any real constraint has entered the picture. AI becomes genuinely useful once that discovery turns into something constrained: not "where should I go" but "I have eight days in mid-September, I want somewhere in the Balkans or MENA with mild weather, I want to see most of the country because I may not return, and I'd love to catch a festival if the timing works." That's a decision problem, and general-purpose models are already quite capable of reasoning through it at the inspiration stage.
I don't think broad discovery is where a travel company can build something defensible, though — the more interesting ground is closer to where real decisions and transactions happen, where accurate pricing, availability, geography, transport, and supplier data start to matter more than conversational fluency. The closer travel AI gets to the point where someone is actually about to book something, the more specialization starts to matter.
The Harder Problem Is the One Worth Building For
I don't think the future of travel AI is prettier itineraries generated faster — that part is becoming commoditized quickly. The more meaningful opportunity is helping travelers navigate the complexity underneath a plan: understanding what they actually want, surfacing a constraint they forgot to mention, showing a real alternative, explaining the reasoning behind a recommendation, and adapting cleanly when something changes.
Travel planning was never primarily an itinerary-generation problem. It's a decision-making problem, and the best travel AI won't just tell a traveler what to do — it'll help them understand their choices well enough to book the trip with confidence.
