[ DISCOVERY · THE DISPATCH ]

I Let an AI Agent Book and Pay for My Whole Trip and the Scary Part Wasn't the Mistakes

I gave an AI agent my card and a budget and told it to plan four days in Porto. It booked the whole thing. The mistakes were funny. What it got right is the part that actually kept me up at night.

When an AI agent books travel in 2026, you expect it to mess up the obvious stuff. And it did! That part was almost reassuring. It’s the other part that I want to talk about.

Here’s what I did. There’s a new wave of these tools — autonomous booking agents, the kind that don’t just suggest an itinerary but actually go off and do it, hit the booking sites, enter the payment details, confirm. Different from the planner bots that hand you a nice list and leave the clicking to you. This thing clicks. So I gave one my budget — €600, four days, Porto — connected a payment method, and said go. Then I sat back and, somewhat insanely, let it.

First, it tried to book me into a hotel that was technically in Porto but a 35-minute drive from anything I’d want to see, because it optimised hard on “price + rating” and ignored “but is it near the actual city.” Classic. It also picked a flight with a five-hour layover when a direct was twelve euros more, because twelve euros is twelve euros to a thing that does not have to sit in an airport for five hours.

And — my favourite — it booked a “highly rated local food experience” that turned out to be a tourist-trap port-wine tasting with a 4.6 star average generated by, I’m fairly sure, four hundred bots. It cannot taste wine. It cannot smell a trap. It just sees numbers and trusts them, which, hold that thought, because it’s the whole point.

I caught all of these. They were easy to catch. Wrong date, weird location, dodgy tour — these are visible errors. You read the itinerary, you go “no, that’s daft,” you fix it. Fine.

The scary part was the speed and the confidence.

I watched this thing spend real money — my real money — in about ninety seconds, with a fluency that made me deeply uneasy. It didn’t hesitate. It didn’t go “hmm, are you sure about €340 on flights?” It just executed, smoothly, and presented me with confirmations like a competent assistant who’d already done it. The total was reasonable. The choices looked plausible. And that plausibility is the trap, because the moment something looks like a sensible human decision, you stop checking it as hard.

I almost didn’t catch the food-trap one. It was sitting right there among genuinely good picks, formatted identically, presented with the same calm authority. The good and the garbage looked exactly the same. There’s no tone of voice. No “I’m not sure about this one.” Everything arrives equally confident. That’s the danger. Not that it’s wrong, but that it’s wrong in the same font as when it’s right.

Credit where it’s due. The logistics-stitching was excellent. It found a flight-plus-hotel combination I wouldn’t have bothered to assemble, handled the time-zone math flawlessly, and rebooked a hotel without complaint when I rejected the first pick. The grunt work — the tab-juggling, the cross-referencing, the form-filling I personally despise — it ate that alive. For pure logistics it genuinely saved me an hour of tedium.

It just has no taste and no nose for nonsense. It’s a brilliant operations clerk and a terrible travel companion.

If you want to try this — and it’s worth trying, it’s wild — here’s my setup after one full run:

• Capped virtual card, every time. Already said it. Saying it again because it matters most.
• Demand a confirmation gate. Use a tool that shows you the full plan and waits for your “yes” before any payment. Avoid the fully-hands-off mode for anything that spends money. Review, then approve.
• Sanity-check the three things it’s worst at: hotel location (not just rating), layover length, and any “experience” or tour it picks. Those are its blind spots.
• Keep your own cancellation eye on it. Make sure bookings are refundable or cheaply changeable, because you WILL want to override one of its choices.

I’d do it again. I will do it again. But I’ll never quite shake the feeling of watching a confident machine spend my money in under two minutes and realising that the failure mode I should fear isn’t the dumb mistake — it’s the smart-looking one I don’t bother to question.

Let the agent do the clicking. Keep the budget on a leash and the taste decisions firmly your own. Confidence is not the same as being right — not in people, and definately not in software.

This article contains affiliate links marked rel="sponsored". We may earn a commission at no extra cost to you.