The table has four browser tabs open, a half-folded paper map with a coffee ring on the Balkans, and a route that looked fine an hour ago and now does not. Three countries, eighteen days, and a night train that turns out to run every other day, not every day. This is the moment most people reach for an AI model and ask it to just plan the whole thing. That is the wrong ask, and it is worth understanding why before you type anything.
A language model has never missed a ferry. It has never sat in a bus station at 6 AM discovering the 7 AM departure was cancelled for the season. It generates a plausible-looking itinerary by pattern-matching against travel writing, and plausible is not the same as correct. Ask it to build a trip from scratch, and the result reads well and ignores how transport actually connects on the ground. Ask it to argue with a route already built, and it earns its keep.
Before you go
Start with your own rough shape. Countries, weeks, the two or three things you actually want out of the trip. Not a model’s shape, yours. That part is not negotiable, because a model with no shortlist and no constraints will hand back its own generic version of “three weeks in Southeast Asia,” and that version has never met your actual priorities.
Where a model helps here is narrowing before you commit and then arguing after you sketch. A matcher that filters destinations on climate, budget and travel style does the narrowing job faster than a chat conversation because it is built to compare, not to converse. That kind of tool earns its place before the model does, not after.
Once you have a shortlist and a rough route, hand the model your constraints as hard rules, not preferences. “No more than one travel day in three.” “Two nights minimum anywhere, no one-night stops.” Then ask it directly: where does this route break those rules? Where do the travel days stack up? What’s likely underestimated? A model given explicit rules to check a plan against is doing arithmetic and pattern comparison, both things it is genuinely competent at. A model asked to invent a plan from nothing is doing improvisation dressed up as expertise.
One caution belongs here in bold letters. Models are confident about transport connections, opening hours and border rules, and confidently wrong often enough that anything time-critical needs a second source. A stated train frequency, a visa-on-arrival claim, a museum’s Monday closure: verify each one against an official schedule or a recent traveler report before it becomes part of the route. The failure mode is not that the model refuses to answer. It answers cleanly and sounds certain. That is exactly the problem.
While you are booking
This is where the arithmetic pays off. Two candidate routes through the same region can look similar on a map and differ by nine hours of total transit once the legs get added up properly, and a model will do that addition faster and more patiently than a spreadsheet at midnight. Feed it both routes leg by leg and ask for a total transit-hour comparison, not a vibe check on which route sounds nicer. A side-by-side comparison of two candidate routes on total transit time settles that argument before either number gets trusted for a booking.
The same goes for budget. A vague number like “keep it under two thousand euros” is not a plan; it is a hope. Turn it into a per-day figure across accommodation, transport, food and one buffer category, and suddenly there is something to hold a specific day against and adjust in real time.
Where the model stops being useful is anything that changed this month. Live prices, current availability, a hostel that closed in March, a ferry route suspended for repairs. None of that lives in a model’s training data in a form anyone can trust, and no amount of clever phrasing fixes that. Check those against a live booking site, not a conversation.
While you are travelling
This is the part most people skip, and it is the most useful one. Keep a running note, plain text is fine, that the model can read back and help update: what actually happened each day, what got skipped and why, what a day genuinely cost versus what got budgeted. Two minutes at the end of a day, dictated or typed.
That log does two things. It turns week one into a sharper week two, because the model can compare actual pace against the plan and flag where it is consistently running over or under. And it turns into something real to hand to the next person asking for advice on the same region, built from what happened rather than from what was hoped would happen.
The honest caveat: this only works with connectivity, and a good chunk of interesting travel does not come with reliable connectivity. Write the note anyway, offline, and sync it when a signal shows up. Do not let the tool decide whether the habit is worth keeping.
None of this makes AI a planner. It makes it a stubborn colleague who checks the math and remembers what it was told yesterday, and who still has no idea whether the 7 AM bus is actually running today. Keep that gap in mind every time the schedule looks a little too clean.
—
About the author
Wouter van Kemenade has backpacked through Southeast and South Asia, including Thailand, Vietnam, Laos, Nepal, the Philippines and Indonesia. He runs Pack Lightly, a multilingual travel reference platform.