United blocks AI agents from booking, as OpenAI brings businesses into ChatGPT
United Airlines will no longer let personal AI agents book for travellers without its permission. A day earlier, businesses could claim their profile in ChatGPT. Your next customer may be an agent.
What happened
United Airlines will no longer let personal AI agents it has not authorised book flights or manage reservations for travellers. The airline announced the move on Wednesday 11 November, a day after OpenAI gave businesses the chance to claim their own profile in ChatGPT.
United is restricting access on its own website and app, and indirectly through travel agencies and booking sites. Its reasons: login details a customer gives to an AI assistant can expose sensitive information, and such an assistant may make unintended purchases, change an account or spend MileagePlus miles. An agent checking a passenger in could also conflict with federal rules, because, according to United, the traveller has to confirm personally at check-in that their bags hold no banned substances. Bookings made through an agent before 11 November will be honoured, United told Travel Weekly.
The analyst Henry Harteveldt told CBS that United does not want yet another, and potentially more powerful, intermediary alongside Google, comparison sites and online travel agencies. Flight searches also cost airlines money, he said. United itself points to banks and retailers that took similar steps. Other airlines are taking a different route: in October Southwest made its flights searchable in ChatGPT, with booking completed on its own site, and Delta’s terms let customers use agents but leave them responsible for what those agents do. On Hacker News the story drew 675 points and nearly 700 comments.
On 10 November OpenAI began a beta in which businesses claim their profile in ChatGPT, much as businesses already do with Google. They prove ownership through a business phone number, email address, domain or Shopify account, and whoever manages the profile passes an identity check; then a blue tick appears. A business can update its details, prices and stock and see, over the past 28 days, how often ChatGPT mentions it, how many people click through and which topics come up. For now only US online retailers and online services may publish a Business Agent of their own. OpenAI says it shares no conversations with businesses and, according to Digiday, that paying does not change how often ChatGPT recommends a business.
“Sharing account credentials with personal, third-party, AI assistants can expose sensitive information.”
United Airlines in its announcement of 11 November 2026
The background
How an agent books for you
An agent can book through the ordinary website: it opens a browser, logs in with your username and password and clicks through the pages, just as you would. To the site it is then hard to tell apart from you. Or it uses an entrance the business offers itself, such as an API: a fixed connection through which software requests data and gives instructions. That is how Southwest’s link with ChatGPT works.
Agents can also identify themselves. Since August 2025 the network company Cloudflare has recognised agents that sign every request digitally, ChatGPT’s agent among them. A site then sees which agent is calling, but not whether the customer asked it to.
This fight has been fought before
Ryanair has taken on booking sites that pull its fares from its website with software, known as screen scraping, and resell the tickets. In July 2024 a jury in Delaware found that Booking.com had broken the US anti-hacking law; Booking.com said it would appeal. Ryanair did sign sales agreements with a number of other sites.
Banks went through the same thing. Financial apps logged in with the customer’s details, posing as that customer. The European Banking Authority banned this on security grounds. Since 14 September 2019, banks in the EU have had to offer such apps an entrance where they identify themselves as what they are.
Where a person has to confirm
The rule United refers to is part of the US regulations on hazardous materials. A passenger is given information about what may not be taken on board when buying a ticket and when checking in. If they can do so without anyone else’s involvement, online for instance, they must acknowledge that they have been shown that information.
Banks moved before United did. Five large US banks block external agents, American Banker reported in late October. An anonymous source told the paper about a bank where an agent reset a password, logged in, changed the accounts approved for transfers and moved $100,000 out; the fraud team then advised blocking all external agents. In November several UK banks followed, according to The Telegraph, and on 20 September Amazon began blocking Meta’s shopping agent Muse.
Our analysis
Agents come in through the front door or through your customer’s password
The pattern is familiar from booking sites and financial apps: a new intermediary logs in as the customer. Businesses first shut the door, and then an entrance emerges where the intermediary identifies itself and the business sets the rules. United’s reasons are real: a shared password, a purchase nobody intended, a statement a person must make. That there is also a commercial interest, as Harteveldt says, does not make them any less true. Our expectation: blocking is the first step, not the end state.
What the coverage tends to skip: a ban is hard to enforce. An agent working in an ordinary browser with the customer’s own password looks like that customer, and a digital signature only helps with agents that want to be recognised. And the check-in rule asks a passenger to acknowledge that they have seen the information. That is one step that can stay with a person, even when an agent does the rest. So the question is not yes or no, but which steps belong to the customer.
Anyone with a webshop or customer portal would do well to decide that now. What may an agent read, what may it propose, what does the customer confirm in person, and who is liable when it goes wrong?
What stands out
- How United will enforce it is unknown United has not said how it will recognise agents. The Register asked and has yet to receive a reply.
- OpenAI turns the worries into conditions A Business Agent must say that it is an AI, and buying or sharing personal data needs the user’s explicit approval.
- Amazon wants agents to ask first According to Amazon, it is normal for a delivery app or online travel agency to ask a business whether it wants to take part. Amazon now asks the same of agents.
- Banks have been here before When financial apps logged in as the customer, Europe ended up with a separate, regulated entrance.
What it means for you
If you run a webshop or customer portal, AI agents will turn up, whether you invite them or not.
- Decide what an agent may see Prices, stock and the status of an order can be open. Customer data and payment details should not be.
- Let the customer confirm what matters A payment, a change to the account or a statement the customer must make personally: always route those through a person.
- Give agents their own entrance Better a connection with limited rights than a customer who shares their password. Then you can see which agent is acting, and switch it off.
Where we come in · Development
We develop an entrance of your own for agents
We develop webshops, apps and integrations, with an API of their own for AI agents. You decide what an agent may read, what it may propose and which step the customer confirms in person. That way nobody has to hand their password to an agent.
Do you know what an AI agent may do on your site?
Tell us what customers arrange with you online. We’ll think it through with you and always reply within one working day.