Amazon and JPMorgan are Restricting AI Agents for Shopping and Banking

Amazon and JPMorgan Chase are restricting AI assistants that shop or move money on customers’ behalf, putting limits on users’ ability to hand everyday tasks to software. The restrictions target particular ways of accessing stores and accounts, while ordinary anti-bot checks are also stopping agents at businesses that welcome them.

One affected popular agentic system is Meta’s  Muse AI assistant which can open a browser, fill out forms and carry out multistep tasks. It runs on a dedicated cloud computer and asks users before sensitive actions, such as sending an email or making a purchase. A customer’s approval, however, leaves the assistant dependent on the business at the other end accepting its visit.

Amazon Cut Off Muse’s Shopping Access

Muse is technically able to complete an Amazon purchase after its release before being blocked. By the evening of September 20, users were instead seeing a notice that access by an unauthorized AI agent violated Amazon’s terms.

Amazon cut off Muse after unsuccessfully asking Meta to exclude its store. The retailer said Muse did not identify itself, Amazon had not agreed to participate, and the handling of customer credentials created privacy and security concerns.

Meta says Muse can use credentials held in secure storage without seeing passwords or payment details. Amazon’s objections also concern who is moving through its account pages and processing transactions without the retailer’s consent.

Banks Separate Reading from Moving Money

JPMorgan’s restriction concerns outside agents obtaining banking credentials and logging in to act for retail customers. A spokeswoman confirmed the policy to CNBC on October 13, while saying the bank supports secure, customer-authorized data sharing.

That permits narrower use: an assistant can obtain some information, such as an account balance, through a read-only connection. It cannot use the blocked route to send money or pay a bill. The bank supplied no timeline for safeguards that would allow agents to interact directly with accounts.

JPMorgan’s CFO Jeremy Barnum tied the blocks to credential sharing and screen scraping, where software retrieves information through account pages, in the bank’s recent earnings-call. He said the bank was blocking those services where it could to protect customers and itself.

Charles Schwab chief executive Rick Wurster separately said his company was blocking unauthorized agents attempting to move clients’ money, including Muse. He is concerned about lack of control over the actions and who would bear responsibility if something went wrong.

There is also competition over who handles customers’ cash. Agents could possibly move money around autonomously, like making deposits toward higher-yielding accounts. JPMorgan is already testing its own Smart Cash tool in a targeted pilot to predict cash flow and to move funds for higher yields. Barnum sees external agents as combining longstanding competition for yield-seeking deposits with competition from financial-technology companies. He also argues that deposit stability should rest on confidence in the bank, rather than a technical difficulty of moving funds.

A Failed Visit Can Have Different Causes

Yelp told TechCrunch it permits non-human traffic only through paid access to its data-licensing program. Delta said it had no partnership or integration enabling third-party agents to shop for or book flights on its digital platforms, and was taking steps against unauthorized automation.

Walmart told TechCrunch that reported Muse purchase failures were unintentional and not due to blocks, as it is partnering with Meta for the assistant: an interrupted check might fail its purpose and eject the agent from the payment system before it  successfully completes a purchase. Personal agents like Muse might trigger safeguards intended to stop fraud and hacking.

Zillow permits ai agents acting for consumers, although anti-scraping defenses sometimes block them. And eBay says it restricts unauthorized agents and actions, including scraping and model training, rather than prohibiting every third-party shopping agent from completing purchases.

Direct Connections Can Carry a Purchase Through

Shopify’s live Shop MCP connection gives compatible agents tools to search its catalog of more than 1 billion products, build carts and check out on Shopify stores. MCP is the connection standard the assistant uses to call those tools.

Previously, an agent could find a product in Shopify Catalog but had to send the shopper to the store’s website to pay. With the new connection, a signed-in shopper can use saved addresses and Shop Pay payment details inside the assistant. The shopper confirms the total and payment method before the agent completes the purchase, and the merchant receives the order in its normal Shopify administration system.

AI developer Sierra and Meta are also collaborating to develop a standard for identifying agents and authorizing their access to businesses. Sierra published a draft of the Personal Agent Protocol, known as Poppy, on October 9. Under its design, a business lists the ways an agent can contact it and the actions available. The assistant identifies itself; when account access is needed, the customer signs in and approves limited access carried by a session token, a temporary credential.

The business still chooses whether the assistant uses its website, direct software interfaces or the business’s own agent, and what actions it permits. Sierra plans workshops and a reference implementation, sample software for developers, during the month after its October 9 publication.