United Airlines Just Banned AI Agents From Booking. Should Your Site Do the Same?

United Airlines has banned AI agents from booking flights on a traveller's behalf. If you run a booking or checkout system, the same question is heading your way: let them in, or lock them out?

United Airlines updated its terms of service recently to ban AI agents from booking flights on a traveller's behalf. Not from searching, not from comparing fares, but from actually completing the purchase. If an agent books it, United says it can cancel the ticket.

That's a big airline with a legal team deciding this is worth writing into a contract. It's worth asking why, and whether your own booking or checkout system needs the same thinking, because the problem isn't specific to flights.

What an AI agent booking a flight actually looks like

Someone asks an AI assistant to "book me the cheapest flight to Manchester on Friday". The assistant doesn't just search. Increasingly, it can fill in the form, enter payment details it's been given, tick the terms and conditions box, and submit. No human looks at the confirmation screen before it's done.

If you've not come across the term before, it's worth reading what an AI agent actually is before you decide how worried to be. The short version: it's software that can take a sequence of actions on its own, not just answer a question.

Why an airline cares enough to ban it

A few things break when an agent books instead of a person:

  • Consent gets fuzzy. Terms and conditions, fare rules, baggage restrictions, all accepted by software reading text, not a person weighing it up. If there's a dispute later, who agreed to what?
  • Mistakes have no second check. A person booking a flight usually glances at the dates, the airport, the name on the ticket. An agent acting on a vague instruction might book the wrong one and not notice.
  • Fraud gets easier to automate. Stolen card details can be fed to an agent and tested against a booking flow far faster than a human could type.
  • Customer service gets harder. When the traveller calls to query a booking they didn't personally make, the airline has to work out what actually happened and who's responsible.

None of that is unique to flights. It applies to anything with a checkout: trade bookings, event tickets, equipment hire, subscription sign-ups, anything where money changes hands through a form.

Where it bites in a smaller booking system

You might not think an AI agent will ever try to book your plumbing callout or your dog groomer slot. But the pattern worth noticing isn't "agents booking flights", it's "software submitting forms meant for people". That already happens with scraper bots, and it's increasingly happening with agents acting on a user's behalf through browser extensions or assistant apps.

A few concrete places this causes trouble:

  • Deposits. If your booking flow takes a card deposit to hold a slot, an agent submitting card details it's been handed, rather than details entered live by the cardholder, is harder to trust. Your payment processor's fraud checks were built around a person typing on their own device.
  • Availability races. An agent that retries on failure, because that's what agents are built to do, can hammer a booking slot faster than a person refreshing a page, and end up holding availability a real customer wanted.
  • Terms acceptance. If your terms exclude liability or set cancellation rules, and those are "accepted" by code rather than read by a person, you want that documented clearly, not assumed.

What I'd actually do about it

I wouldn't start by trying to detect and block every bot, that's a losing arms race and you'll catch real customers in the net. I'd start with the parts of your flow that matter most if something goes wrong.

  • Decide what "acting" versus "drafting" looks like in your own flow. If you're building any automation of your own that touches customer bookings or orders, it's worth being deliberate about where it should stop and ask rather than just complete the action. I've written about when an AI agent should act versus just draft, and the same logic applies to anyone else's agent interacting with your site.
  • Add a confirmation step a bot can't skip. Something that needs a human glance, a one-time code to the booker's phone, a confirmation email they click, puts a person back in the loop at the one point that matters most.
  • Put it in your terms, properly. If you want to restrict agent-made bookings, say so clearly, the way United has, rather than relying on it being implied. It doesn't fix the technical problem but it gives you a position to stand on.
  • Keep a proper audit trail. If a booking ever gets disputed, you want to know exactly what was submitted, when, from where, not just that a booking exists. That's the same discipline behind building an audit trail you can actually trust.
  • If you want to support agents deliberately, build a proper interface for it. The businesses that will do well out of this aren't the ones blocking everything, it's the ones offering a real API for trusted partners to book through, with proper authentication, rather than leaving automated tools to fill in a form meant for humans. That's an integration project, not a form tweak, and it's worth scoping it as one.

United's move is a signpost, not a one-off. More businesses with anything transactional online are going to have to decide, explicitly, whether an AI agent booking on a customer's behalf is something they allow, restrict, or quietly design around. Better to make that decision on purpose than find out the hard way when a booking goes wrong and nobody can say who actually agreed to it.