What Muse does — and OpenAI's narrower checkout route
Muse runs in Meta's cloud, so it keeps working when the phone is off. It books flights and completes purchases without a person watching the screen. Forbes reported more than 10 million U.S. users within six weeks of the September launch. The figure is U.S.-only, and no euro pricing exists yet because there is no European launch.
OpenAI's approach to shopping inside ChatGPT has moved in a different direction. The company wound down Instant Checkout in favour of dedicated retailer apps inside ChatGPT. That is a narrower, app-by-app route than a persistent agent that completes purchases anywhere.
Two ways to handle an agent at checkout
Walmart chose the guest-list approach. An external agent gets its own isolated profile, a spending cap, and a step where a human has to approve the order. The bot can shop, but it cannot drain anything.
Amazon went the other way and blocks outside agents, Meta's Muse among them. The stated reason is control over the customer interaction, which is where Amazon's advertising and recommendation business lives.
Both positions are defensible from the retailer's side. Walmart treats an agent as a customer with a preloaded card. Amazon treats it as a crawler with a wallet. What neither wants is an unaudited process placing orders in someone else's name and then disputing the result.
Banks are the harder wall
Retailers can block a bot with a rule. Banks have to answer a different question: is this instruction from the account holder, or from a program holding the account holder's password? Those look identical on the wire. JPMorgan has told customers that external AI agents may not access accounts, and that sharing credentials with a third-party agent breaks its terms.
That position is not hostility to AI. It is fraud liability. Once credentials are shared, the bank loses the one signal its monitoring relies on, which is evidence that a human was present.
Why a European checkout resists automation
The U.S. model is simple: hand the agent your login and let it buy. That model does not map onto the EU, for a structural reason rather than a political one. The Second Payment Services Directive and its technical standards require strong customer authentication for electronic payments, built from two independent factors. A stored password is one factor. An agent that holds a password and nothing else cannot authenticate a payment on its own, so the flow either stops and asks the user, or it runs through a licensed payment initiation provider acting on explicit consent under open banking.
That changes the design brief. In Europe the agent is not a substitute for the customer's credentials. It is a client of a consent-based payment API, and the consent has to be logged, scoped and revocable under the GDPR. An agent that holds a year of purchase history for personalisation is also a data-minimisation problem for whoever operates it.
Transparency rules add a second layer. Since 2 August 2026, Article 50 of the EU AI Act requires that people be told when they are interacting with an AI system, and that synthetic content carry detectable markings. Enforcement of general-purpose model obligations is live as well, with administrative fines reaching €15 million or 3% of worldwide turnover. The deadline for standalone high-risk systems in Annex III was pushed to 2 December 2027 by the Digital Omnibus adopted in 2026.
One detail worth flagging for anyone building in this space: Article 50 governs disclosure to the person talking to the system. It does not settle whether a merchant must be told that the buyer on the other side is a program. That is left to contract terms and, for payments, to consent rules.
What one agent purchase costs in tokens
Inference is cheap enough that it is almost irrelevant to the argument, which is worth showing with numbers. The per-million-token prices below are illustrative list rates for the named models as of November 2026, not a quoted rate card. Take a purchase flow that burns 20,000 input tokens and 2,000 output tokens across browsing, comparison and checkout. That token count is our stated assumption, not a measurement; we do not run a shopping agent in production.
At DeepSeek V4.1 Flash rates ($0.15 per million input, $0.60 per million output) the flow costs about $0.0042. At Mistral Large 4 pricing ($0.68 / $2.09) it lands near $0.018. Claude Sonnet 5.5 at $2.00 / $10.00 comes to roughly $0.06, and Grok 4.7 at $2.00 / $6.00 to about $0.05.
For an illustrative basket in the low hundreds of dollars — use $100 to $300 as the stand-in range, not a measured average — that token bill is roughly 0.001% to 0.06% of the order. A model roughly fourteen times more expensive per token adds about six cents to a low-hundreds purchase. Whatever is slowing autonomous commerce down, it is not the price per token. It is credentials, liability, and returns.
What agent traffic can and cannot show
Agent traffic to retail checkout systems has been rising, but the public numbers are easy to overread. Traffic is not revenue. A handful of monthly readings is not a trend line, and a larger share of sessions does not mean the same share of baskets, since an agent that fails at authentication still counts as a visit. The stronger signal would be completed purchases with a named basket size, not page views; without that, the shift described here is about access and liability, not market share.
What a builder should take from the split
The Walmart design is the one that survives contact with fraud teams and data protection officers at the same time: an isolated identity for the agent, a hard budget, and a human approval step above a threshold. It is also the only design that produces an audit trail when a customer disputes an order.
In the EU, storing a card number or a banking password inside the agent is the wrong shortcut. Payment initiation through a licensed provider, with consent recorded per merchant and per amount, keeps the liability where the regulation expects it. Everything else is a workaround that a bank can void in a terms-of-service clause.
Can I use Meta's Muse in Europe?
Not on the basis of anything Meta has published. The more than 10 million user figure reported by Forbes covers the United States, and no euro subscription or commission terms have been released. Treat any EU date you see as unconfirmed until it appears in Meta's own documentation.
Does the EU AI Act ban shopping agents?
No. The obligations that apply today come from Article 50 transparency duties and from the general-purpose model rules enforced by the European AI Office, with fines up to €15 million or 3% of worldwide turnover. Standalone high-risk obligations under Annex III now apply from 2 December 2027 after the Digital Omnibus adopted in 2026.
Can an agent pay with my card in the EU without asking me each time?
Only through a consent-based route. Strong customer authentication requires two independent factors, so either the user approves the payment, or a licensed payment initiation provider initiates it under explicit, scoped consent. Handing an agent a stored password does not satisfy that test.
Meta's public disclosures stop at the U.S. border: more than 10 million U.S. users within six weeks, according to Forbes, an undisclosed commission on completed purchases, and no European price list.
