AI Shopping Is Becoming A Permissioned Channel

Amazon confirmed on 20 September that it had blocked Meta’s new Muse personal AI agent from browsing and buying on Amazon.com. The action came less than two weeks after Meta launched Muse with the ability to open a browser, fill forms and request approval before making purchases.

The dispute shows that agentic commerce depends on permission from the retailer whose catalogue, account data and checkout infrastructure an agent wants to use. For retail CMOs, that turns AI shopping from a product experiment into a channel-governance decision.

What Amazon Blocked

Amazon told GeekWire that Meta had not disclosed Muse’s planned access, that the agent did not identify itself while browsing, and that Amazon had not authorised it to enter customer accounts, collect data or process transactions. Users attempting to shop through Muse saw a notice saying the agent’s continued access violated Amazon’s conditions.

Meta says Muse runs in a dedicated secure virtual machine and that credentials are stored so the agent cannot see passwords or payment methods. Its launch also connected checkout to Stripe’s Link wallet, which can generate a one-time card and seek user approval before a purchase.

Those protections address the relationship between the user and the agent. Amazon’s objection concerns the separate relationship between the agent and the merchant. Its spokesperson said third-party applications should “operate openly and respect service provider decisions.” Customer permission does not automatically create retailer permission.

Agent Access Is A Commercial Decision

Amazon has an economic reason to control the interface. GeekWire reported that the company generated more than US$68 billion in advertising revenue last year. An outside agent that compares products, chooses an item and proceeds to checkout can reduce the browsing that exposes shoppers to sponsored listings, recommendations and Amazon’s own Alexa for Shopping assistant.

The retailer’s published agent terms require automated services to identify themselves in web requests and allow Amazon to limit their access. The immediate issue is security and consent, but the same gate controls which party sees intent data, shapes recommendations and receives attribution.

Only 16% of shoppers were comfortable allowing an AI assistant to find and buy products for them in Coveo research cited by Axios. Adoption remains early, yet the confrontation establishes the bargaining question: does an agent arrive as an authorised sales partner, or as an uninvited layer between retailer and customer?

Retailers Are Choosing Different Access Models

Amazon’s position is not the only model. Walmart is working with Google and OpenAI to surface its products inside their agentic shopping interfaces while developing its own assistant, Sparky. OpenAI says retailers including Target, Sephora, Lowe’s, Best Buy and Wayfair have integrated product data through its Agentic Commerce Protocol.

These are permissioned connections, with product feeds and conversion paths agreed in advance. Shopify’s Mani Fazeli described its approach as keeping “the merchant’s brand front and center.” The competitive difference is whether external agents participate through negotiated standards that preserve merchant controls.

That question matters in APAC, where brands often sell through dominant marketplaces as well as their own stores. A single policy will not fit every channel. Marketplace operators may protect advertising and customer data, while brand-owned stores may accept external-agent traffic for incremental discovery.

What Retail Leaders Should Decide Now

CMOs, commerce leaders and technology teams should define an agent-access policy before demand forces a hurried answer. It should specify approved agents, machine identification, permitted catalogue and account data, authentication, payment controls, returns responsibility, attribution windows and whether sponsored placement can influence recommendations.

They should also separate discovery from transaction authority. A brand may want broad product visibility in AI answers while limiting access to loyalty balances, order histories or checkout. That creates a practical ladder for testing value without surrendering the entire customer relationship.

Amazon’s block suggests that agentic commerce may develop through bilateral permissions and platform rules rather than frictionless access to every store. The executive decision is which agents the business will recognise, under what terms, and how much of the customer journey it is prepared to share.