You ask your AI agent to order you a pair of shoes, book a table for Saturday, find a plane ticket for Christmas. A few minutes later, it comes back with an unexpected answer: "I couldn't."
This is not a breakdown. The site closed the door on it.
What is happening
According to an investigation by the American site The Verge, users of several AI agents, including Meta's Muse, xAI's Grok and OpenAI's autonomous assistants, have been finding for several weeks that they can no longer complete certain tasks on major sites.
Amazon has told Meta that its Muse agent's access violated its rules. Other agents can no longer make purchases there. The airline Delta prevents agents from booking tickets. The reviews site Yelp only accepts agents from companies that pay to access its data. And as we explained on Monday, major banks are blocking agents that want to move money. Users also report blocks on car sales and train booking sites.
Why the sites say no
To protect themselves. An agent that logs in with a customer's credentials, that buys, books or pays on their behalf, is a risk of fraud and error. And if something goes wrong, who is responsible?
To keep their customers. This is the least admitted argument, and probably the most important. A site like Amazon also makes money from the advertising and suggestions its visitors see. An agent that comes to buy a product without looking at anything else sees no advertising. It deprives the site of what keeps it alive.
To get paid. A site's data has value. Yelp is upfront about it: agents are welcome, provided they pay.
This standoff is not new. Last year, Amazon had already banned the shopping agent from the start-up Perplexity, before suing it.
In 1994, a Dutch engineer, Martijn Koster, proposed a very simple convention: a small text file, called robots.txt, that a site places at its root to tell crawlers what they are and are not allowed to visit. It is not a law, only a courtesy rule that search engines chose to respect for thirty years. With AI agents, which no longer merely read but act, this old convention is no longer enough, and sites are moving to barriers.
The other path: paid doors
Not everyone is closing. The online store platform Shopify has chosen the opposite, opening the shops it hosts to AI agents, under rules it sets. Banks, for their part, are preparing read-only access.
A web with two doors is taking shape: one for humans, one for agents, often paid, always controlled.
What it changes
For AI companies, this is a blow. Their promise rests on agents capable of doing everything for us online, like those OpenAI presented in early October. Every closed door reduces their usefulness.
For users, this is likely to make the web a little more painful: more checks to prove you really are a human, and services reserved for those who pay.
What to take away
The web was built for humans. AI agents are entering it as a new kind of visitor, and the owners of the premises are starting to set their terms.
The future of agents will not be decided only in the labs, but in these negotiations, site by site. An agent capable of doing everything is not much use if it is not let in anywhere.