I Let Three AI Agents Plan the Same 10-Day Trip. One Booked Me Into a Town With No Hotel.

Eighty-one percent of travelers now say they use AI somewhere in the planning process. I believe that number because I'm one of them, and so is everyone I met in a hostel in Kigali last month.

Here's the thing, though. Using AI to plan a trip and using AI to plan a trip you'll actually enjoy are two different skills, and nobody teaches the second one.

So I ran an experiment. Same trip, same dates, same budget. Ten days across Rwanda and Uganda, roughly $1,400 all in, two people. I gave three different AI planning tools the exact same prompt and let them build full itineraries. Then I checked every single claim they made against reality.

Two of the three plans would have cost me money and a night of sleep. One was genuinely good. The differences were not about which model was smarter. They were about what I asked.

What the tools actually got wrong

The first plan was beautiful. Day by day, hour by hour, with restaurant names and travel times down to the minute. It was also fiction.

It routed me from Kigali to Musanze on a road that exists, but scheduled the drive at 2pm on a Friday during a market day, which added ninety minutes the plan didn't account for. It recommended a guesthouse that closed in 2024. And it booked me into a village that, as far as I can tell, has never had a hotel in it.

The second plan was more honest about uncertainty, which I appreciated, but it was so hedged it became useless. "Consider visiting a local market." Which market? What day? Is it the one with the goats or the one with the phone parts, because those are different experiences and only one of them is what I wanted.

The third plan was the one I'd use again. Not because it was more accurate, but because it was structured the way a real traveler thinks: a base, a radius, two anchor activities, and a lot of deliberate empty space.

Look, the pattern wasn't subtle. The tools that failed were the ones given permission to invent. The one that worked was given constraints and told what to do when it didn't know something.

The prompt that changed everything

My first prompt was: "Plan a 10-day trip to Rwanda and Uganda for two people, budget $1,400."

That prompt is a trap. It asks for a deliverable, so the model produces a deliverable, and the easiest deliverable to produce is a confident-sounding fantasy.

Here's what I use now, and I've used it on four trips since:

You are planning a trip. Rule one: never invent a place, price, or schedule. If you don't have reliable information, say "unverified" and tell me what to check. Rule two: give me a base town and a radius, not an hour-by-hour schedule. Rule three: for every recommendation, tell me why it fits a budget traveler who hates guided tours. Rule four: list the three things most likely to go wrong.

That fourth rule is the one that does the work. Asking a model to predict its own failure modes is weirdly effective. It surfaces the road that floods, the border post that closes early, the guesthouse that only takes cash.

Then I verify. This is the part people skip. I take the plan and I check three things myself:

1. Does the accommodation exist, and does it have recent reviews from the last six months?
2. Is the transport route real, and roughly how long does it actually take?
3. Is there anything happening on those dates that changes the plan (holidays, closures, weather)?

That's maybe twenty minutes of work. Twenty minutes is the difference between an itinerary and a wish.

Where AI planning actually shines

I want to be fair to these tools, because they've changed how I travel in ways that are genuinely good.

Language. I don't speak Kinyarwanda or Luganda. A model that can give me ten phrases with pronunciation, plus the cultural context for when to use them, has made more real conversations happen for me than any phrasebook. The trick is asking for the informal version, not the textbook one. "How would a 25-year-old in Kigali actually say this to a friend" gets you somewhere a dictionary won't.

The boring middle. Visa requirements, departure taxes, whether the bus station has luggage storage, what the SIM card situation is at the airport. This is where AI saves me the most time, because it's exactly the stuff I'd otherwise spend two hours in a browser tab graveyard figuring out. It's still worth a second source on anything that could get you turned away at a border, but as a first draft it's excellent.

Replanning. This is the underrated one. My flight got moved six hours in Kigali. Instead of panicking at the airport, I typed the new arrival time and asked what to cut. The model gave me a reshuffled two days in about ten seconds. It wasn't perfect. But it was 80% right in the moment I needed 80% right.

The rule I actually live by now

The agent plans. I decide. The ground decides last.

That's it. That's the whole system. AI is a planning assistant with infinite patience and zero accountability. It will happily send you to a hotel that doesn't exist because the alternative is saying "I'm not sure," and it doesn't feel the cost of being wrong. You do.

So use it for the shape of the trip, the language, the logistics, the replanning. Do not use it for the last mile of truth. Confirmation numbers, recent reviews, a phone call to the guesthouse. That's yours.

One more thing, and I mean this. The best parts of my last four trips were the parts no model suggested. The detour because a motorbike driver said the road ahead was boring. The two hours I spent waiting out rain in a shop that sold only fabric. Nobody's itinerary has room for that, because it hasn't happened yet.

Leave the gaps. Let the gaps fill themselves.

Stay curious, stay paranoid.