Amazon shuts out Meta's AI agent — in agentic commerce the merchant decides, for now

What happened

Amazon has stopped Meta's AI agent Muse from buying on amazon.com on behalf of customers. Anyone who tried from Sunday evening onwards was met, according to GeekWire, with a notice that continued access by an unauthorised AI agent violates Amazon's conditions of use. GeekWire reported it first on 21 September 2026; Finextra and TechCrunch picked it up the same day. Amazon gives three reasons: Meta did not announce the access, the agent does not identify itself as an agent while browsing, and it apparently captures and stores customers' login credentials. An Amazon spokesperson frames the claim as follows — applications that offer to make purchases on behalf of customers from other businesses should operate openly and respect the service provider's decision on whether to take part. Meta disputes the credentials point: Muse has no visibility into passwords or payment methods, and login data goes into secure storage that the agent can use without seeing it. Amazon has not lifted the block so far. Muse has been available since 8 September 2026; Finextra points to a partnership with Stripe for checkout and payments. Amazon has previously shut out agents from Google and OpenAI and sued Perplexity over its Comet browser.

Who is affected

Every online shop and every marketplace seller in Europe that has to decide over the coming months whether AI agents may order from them — and everyone assuming that question will be answered for them somewhere else.

Assessment

This is not the next agentic commerce report, it is the first visible doorman case. For months the industry has been debating mandates, spending limits and burden of proof as though it were a matter of scheme rules or supervision. Amazon shows that the first decision falls much earlier and is entirely mundane: do my terms of use say that automated third parties may buy here? What is striking is what Amazon objects to — not the paying, but the fact that the agent does not identify itself. An orderer you cannot identify is not a customer in a dispute, it is a risk.

For German merchants this matters for two reasons. First, hardly any shop has a rule on this in its terms and conditions at all, and silence in case of doubt means permission. Second, the liability question hangs on it, and that question is unresolved in the card world anyway: who carries the loss when an agent orders wrongly using stored credentials? As long as no scheme and no supervisor answers that, your own contract text is the only tool a merchant reliably holds. I see no sense in shutting agents out wholesale — the channel is coming. But leaving it open without noticing is something different from opening it deliberately.

What to do now

1. Look at your own terms of use: is automated access by third parties covered? If not, that is the shortest open building site you have.
2. Clarify with your shop or PSP provider whether agentic orders are technically identifiable — and if not, what labelling is planned.
3. Decide whether you want to allow agents, and document the decision. Either answer is defensible; only the unanswered question is not.