Amazon kicks out AI shoppers while Walmart gives them a $50 limit

“Continued access by an unauthorized AI agent violates Amazon’s Conditions of Use, to which our customers have agreed.” That is the message X user Nic Cruz Patane got overnight into October 15, 2026, when he asked his Grok-based bot to order something from Amazon. He posted a screenshot with a short note: his bots could no longer place orders there. The same day Elon Musk weighed in on complaints from Grok bot owners. First he suggested asking Grok to buy elsewhere, since there are far more shopping sites than Amazon. Then he wrote: “Maybe we should start a competitor to Amazon.”

Six days later Walmart did the opposite of Amazon: instead of shutting bots out, it let them in. A shopping bot will get its own profile inside its owner’s account, with its own order history and a spending cap of, say, $50. Months earlier, Target had rewritten its terms so that a purchase made by a bot counts as a purchase made by the customer. And American Express has promised to cover mistakes made by AI agents, with a catch: if you asked for “the best,” that request is too vague to count as an error.

Three retailers and one card network have given three different answers to a question that sounded theoretical a year ago. Who pays when software does the buying?

Ten million shoppers without hands

The stores have a reason to hurry. On September 8, 2026, Meta launched Muse in the US , a personal AI agent. An agent is a program that carries out a task step by step on its own: it finds a product, compares prices, fills in the order form. From day one Muse could pay through Stripe’s Link at more than a million merchants. Before paying, Muse shows the person a card with the order and the total, and no money moves without a “yes.”

Demand surprised even Meta. On the third-quarter earnings call, Mark Zuckerberg said that in the first week the company had to work with Stripe to raise credit limits. Put simply, Stripe capped how much money Meta could push through it, and purchases hit that ceiling within days because volumes were above what anyone expected. By October more than 10 million people in the US and Canada had set up Muse, and shopping is its most common use. The average purchase, according to the company, is over $200, though that includes travel bookings.

In other words, America’s biggest stores suddenly have a crowd of customers who never look at the shelves, never scroll past an ad and never click a banner. Each store had to decide whether to let them in.

Amazon locks the door

Amazon moved first and hardest. Around September 20, 2026, it blocked Muse , and Muse users got that same “unauthorized AI agent” message. Amazon has three complaints: Meta gave no notice, the agent doesn’t identify itself as an agent, and, in Amazon’s view, it stores customers’ logins. Meta replies that Muse has no visibility into passwords or payment methods and uses logins from secure storage without seeing them. By mid-October owners of Grok-based bots were complaining about the same message, although Amazon has not confirmed blocking Grok.

The legal footing is thinner than it looks. In November 2025 Amazon sued Perplexity over its Comet shopping bot, accusing it of hacking. In March 2026 a lower court barred Comet from entering shoppers’ accounts, but on August 4 an appeals court lifted that ban , reasoning that it is the person who enters the store and the bot is something like their browser.

So Amazon now leans mostly on its own terms of use, which include a section on agents: a bot must identify itself and must not hide that it is a bot. There is some irony here. In April 2025 Amazon itself began testing Buy for Me , an AI feature that buys products on other brands’ websites without the shopper ever leaving the Amazon app. Outside agents are locked out of Amazon, while Amazon’s own agent shops elsewhere.

Walmart gives the bot a wallet on a leash

On October 21 Walmart laid out its approach . The company is frank about the problem: today an agent signs in with the customer’s login, and its orders blend with the person’s so completely that no one can tell them apart. If the bot overspends, it is hard even to work out who pressed the button.

The fix is called Agent Profiles. Each agent gets a separate profile inside the customer’s account, with its own order history. By default the customer approves every purchase, but they can set a limit instead, and Walmart’s examples are $50, $100 and $200. Walmart hasn’t said yet whether a limit applies to a single order or to a week or month. The agent never sees card details and cannot add or change payment methods, and the store may ask for approval even within the limit. A profile can be deleted at any time, and agents that break the rules will be blocked.

Bots will come in through a separate entrance, by invitation only for now, and customer profiles are promised within the next month. Walmart hasn’t said which agents will be supported at launch.

Target: what the bot bought, you bought

Target took a third path back in the spring. On March 22, 2026, it updated its terms for an integration with Google’s Gemini: you can ask Gemini in a chat to buy something from Target, and the bot places the order itself. The agent counts as “authorized” once you have given it permission to act on your behalf. Under the new terms , its purchases are “considered transactions authorized by you,” even if the outcome doesn’t match what you intended. Target also says it cannot guarantee an agent will act exactly as you intend in all circumstances. Reviewing what the agent did, and promptly reporting anything suspicious, is the customer’s job.

Target has said its assistant buys nothing without customer approval and that such orders follow the standard return policy. Still, under the contract, the bot’s mistake is the human’s problem.

Who refunds the word “best”

Banks and card networks are looking for answers too. On April 14, 2026, American Express announced protection for purchases made by agents. The conditions: the agent is registered with Amex, and the cardholder’s authenticated purchase intent is passed along with the order. In that case the company promises to protect customers from charges caused by agent error.

The fine print is where it gets interesting. On its program page , Amex itself says claims may not be available where purchase intent is subjective or non-verifiable, and gives two examples: “best” and “really nice.” The first step, “where possible,” is a return to the merchant, and only then a claim to Amex. Full terms and the list of registered agents are promised later, and the whole program is still described in the future tense. Finextra columnist Maren Solberg noted that vague requests are exactly where agents go wrong, and put it in her headline : Amex will pay for agent errors, except in the cases agents actually get wrong.

If an agent picks the wrong size or orders ten packs instead of one, the error is right there in the order. But people rarely hand a bot an exact product code. Picture it: you ask for “good headphones under $100,” the bot picks the model with the best reviews, and you needed noise cancelling. For Amex that is not an agent error, because “good” can’t be verified. Those are the disputes the protection steps around.

Shoppers seem to sense the catch. Payments company Worldpay’s 2026 report surveyed people in seven countries. Among those open to AI shopping at all, 80% would give an agent some role in payment, yet 59% still want to approve every purchase themselves. It is an odd arrangement: the software does all the work, and the last click still belongs to a person. As long as “best” doesn’t count as a bot’s mistake, people are right to keep the “yes” button for themselves. Would you hand it to the software, knowing you will answer for its taste anyway?