I Let AI Agents Plan My Vacation—and It Wasn't Terrible

WIRED

3 days ago

The latest wave of AI tools claim to take the pain out of booking your next trip. From transport and accommodation to restaurants and attractions, we let AI take the reins to put this to the test. Photo-Illustration: Wired Staff/Victoria Turk
The worst part of travel is the planning: the faff of finding and booking transport, accommodation, restaurant reservations—the list can feel endless. To help, the latest wave of AI agents, such as OpenAI's Operator and Anthropic's Computer Use claim they can take these dreary, cumbersome tasks from befuddled travelers and do it all for you. But exactly how good are they are digging out the good stuff?
What better way to find out than deciding on a last-minute weekend away. I tasked Operator, which is available to ChatGPT Pro subscribers, with booking me something budget-friendly, with good food and art, and told it that I'd prefer to travel by train. What's fascinating is that you can actually watch its process in real time—the tool opens a browser window and starts, much as I would, searching for destinations accessible by rail. It scrolls a couple of articles, then offers two suggestions: Paris or Bruges. 'I recently went to Paris,' I type in the chat. 'Let's do Bruges!'
Armed with my decision, Operator goes on to look up train times on the Eurostar website and finds a return ticket that will take me to Brussels and includes onward travel within Belgium. I intervene, however, when I see the timings: It selected an early-morning train out on Saturday, and an equally early train back on Sunday—not exactly making the most of the weekend, I point out. It finds a later return option.
So far impressed, I wait to double-check my calendar before committing. When I return, however, the session has timed out. Unlike ChatGPT, Operator closes conversations between tasks, and I have to start again from scratch. I feel irrationally slighted, as if my trusty travel assistant has palmed me off to a colleague. Alas, the fares have already changed, and I find myself haggling with the AI: can't it find something cheaper? Tickets eventually selected, I take over to enter my personal and payment details. (I may be trusting AI to blindly send me across country borders, but I'm not giving it my passport information.)
Using ChatGPT's Operator to book a train ticket to Bruges. Courtesy of Victoria Turk
Trains booked, Operator thinks its job is done. But I'll need somewhere to stay, I remind it—can it book a hotel? It asks for more details and I'm purposefully vague, specifying that it should be comfy and conveniently located. Comparing hotels is perhaps my least favorite aspect of travel planning, so I'm happy to leave it scrolling through Booking.com. I restrain myself from jumping in when I see it's set the wrong dates, but it corrects this itself. It spends a while surveying an Ibis listing, but ends up choosing a three-star hotel called Martin's Brugge, which I note users have rated as having an excellent location.
Now all that's left is an itinerary. Here, Operator seems to lose steam. It offers a perfunctory one-day schedule that appears to have mainly been cribbed from a vegetarian travel blog. On day 2, it suggests I 'visit any remaining attractions or museums.' Wow, thanks for the tip.
The day of the trip arrives, and, as I drag myself out of bed at 4:30AM, I remember why I usually avoid early departures. Still, I get to Brussels without issue. My ticket allows for onward travel, but I realize I don't know where I'm going. I fire up Operator on my phone and ask which platform the next Bruges-bound train departs from. It searches the Belgian railway timetables. Minutes later, it's still searching. I look up and see the details on a station display. I get to the platform before Operator has figured it out.
Bruges is delightful. Given Operator's lackluster itinerary, I branch out. This kind of research task is perfect for a large language model, I realize—it doesn't require agentic capabilities. ChatGPT, Operator's OpenAI sibling, gives me a much more thorough plan, plotting activities by the hour with suggestions of not just where to eat, but what to order (Flemish stew at De Halve Mann brewery). I also try Google's Gemini and Anthropic's Claude, and their plans are similar: Walk to the market square; see the belfry tower; visit the Basilica of the Holy Blood. Bruges is a small city, and I can't help but wonder if this is simply the standard tourist route, or if the AI models are all getting their information from the same sources.
Various travel-specific AI tools are trying to break through this genericness. I briefly try MindTrip, which provides a map alongside a written itinerary, offers to personalize recommendations based on a quiz, and includes collaborative features for shared trips. CEO Andy Moss says it expands on broad LLM capabilities by leveraging a travel-specific 'knowledge base' containing things like weather data and real-time availability. Courtesy of Victoria Turk
After lunch, I admit defeat. According to ChatGPT's itinerary I should spend the afternoon on a boat tour, taking photos in another square, and visiting a museum. It has vastly overestimated the stamina of a human who's been up since 4:30AM. I go to rest at my hotel, which is basic, but indeed ideally located. I'm coming around to Operator's lazier plans: I'll do the remaining attractions tomorrow.
As a final task, I ask the agent to make a dinner reservation—somewhere authentic but not too expensive. It gets bamboozled by a dropdown menu during the booking process but manages a workaround after a little encouragement. I'm impressed as I walk past the obvious tourist traps to a more out-of-the-way dining room that serves classic local cuisine and is themed around pigeons. It's a good find—and one that doesn't seem to appear on the top 10 lists of obvious guides like TripAdvisor or The Fork.
On the train home, I muse on my experience. The AI agent certainly required supervision. It struggled to string tasks together and lacked an element of common sense, such as when it tried to book the earliest train home. But it was refreshing to outsource decision-making to an assistant that could present a few select options, rather than having to scroll through endless listings. For now, people mainly use AI for inspiration, says Emma Brennan at travel agent trade association ABTA; it doesn't beat the human touch. 'An increasing number of people are booking with the travel agents for the reason that they want someone there if something goes wrong,' she says.
It's easy to imagine AI tools taking over the information gateway role from search and socials, with businesses clamoring to appear in AI-generated suggestions. 'Google isn't going to be the front door for everything in the future,' says Moss. Are we ready to give this power to a machine?
But then, perhaps that ship has sailed. When planning travel myself, I'll reflexively check a restaurant's Google rating, look up a hotel on Instagram, or read TripAdvisor reviews of an attraction, despite desires to stay away from the default tourist beat. Embarking on my AI trip, I worried I'd spend more time staring at my screen. By the end, I realize I've probably spent less.