United Airlines is blocking customers' AI agents. Would you let one log in to your product?

by•

Sooner or later a customer's assistant will show up at your product holding their password. I'd like to hear what other makers plan to do when it does.

Last week United Airlines said it will restrict personal AI agents from booking, managing or checking in for flights on a traveller's behalf. I'm going by CBS News and TechCrunch's reports of United's statement. I haven't found the policy text itself, and neither report gives an effective date or says how United will detect an agent.

United gave three reasons, as quoted in those reports. Sharing account credentials with a third-party assistant "can expose sensitive information." An agent could make purchases or account changes, or spend MileagePlus miles, that the customer didn't intend. And some steps legally need the person: at check-in a traveller has to confirm for themselves that their bags hold nothing banned. The restriction covers United's own site and app and, indirectly, travel agencies and aggregators.

There's a commercial reading as well. Henry Harteveldt, an aviation analyst quoted by CBS, said United doesn't want another powerful intermediary between it and its customers, and that searches cost airlines money. CBS also notes that Southwest went the other way and made its inventory viewable to ChatGPT users. TechCrunch adds that Bank of America recently urged customers not to share banking passwords with AI assistants.

I don't want to argue about whether United is right. What interests me is that every product with a login is about to face a small version of the same decision, and most of us haven't sat down and made it.

A few things I keep turning over.

The shared password is the real problem. When someone gives an assistant their login, the assistant can do everything they can do, and you can't tell the two apart in your logs. Blocking agents and welcoming agents are both better than letting it happen quietly.

Reading and acting are different risks. An agent that looks up my booking is one thing. An agent that spends my miles is another. If I were opening a door, I'd open the read-only one first and watch what people do with it.

Some confirmations only mean something when a person gives them. United's baggage example is a legal one, but most products have a softer equivalent: deleting an account, sending something to another person, agreeing to terms.

Saying no has a cost too. If people want their assistant to handle a chore and you refuse, some will go to a competitor that allows it, and others will hand over the password anyway.

For Murror the question is uncomfortable in a specific way. What people write with us is about as private as data gets. I can imagine someone reasonably wanting their own assistant to read their entries and help them spot patterns, and I can imagine that same access going badly wrong. My instinct is that if we ever allow it, it should be a narrow permission the person grants and can revoke, never a shared password, and that nothing gets written or deleted without them. We haven't built anything like this and I haven't settled it. I'd rather decide it deliberately than discover it in our logs.

So, open questions, and I'd love people to answer each other, not just me:

If a customer's AI agent logged in to your product tomorrow with their password, would you know? And would you mind?

Which actions in your product would you never let an agent complete without the person?

Has anyone already built an official way in for agents, like an API, scoped tokens or an MCP server? What did people actually use it for?

2 views

Add a comment

Replies

Be the first to reply

Have a question or a thought to share? Add a comment above to start the conversation.