Walmart wants to let the AI agents shop for you

Asking an artificial intelligence to compare milk, find the usual toilet paper and place the order is no longer just a demonstration. Walmart announced on 21 October an infrastructure allowing authorized AI agents to search for products and make purchases on behalf of a customer.

The project includes interesting protections, but it is not yet fully available. Consumer profiles are due to arrive next month, the technical gateway starts in beta and Walmart has not confirmed a launch in Canada.

A separate profile for each agent

The system called Agent Profiles must assign a separate profile to each authorized assistant. The customer will be able to view orders made by the agent, remove access and choose between two levels of control: approve each purchase or allow transactions to be made under a specified limit, e.g. $50, $100 or $200.

According to Walmart's official announcement (https://corporate.walmart.com/news/2026/10/Making-It-Posible-for-Agents-to-Shop-on-Behalf-of-Customers), human approval will remain the default rule. The agent will not receive card data and will not be able to add or modify a payment method. Walmart also claims to be able to require additional validation, even when a transaction meets the ceiling.

Automated controls will remain covered by the usual return process, including when an agent does not act as intended. However, this protection does not guarantee that an error will be detected before shipping.

A path reserved for software

Today, a general agent must often manipulate a site designed for a person: open pages, interpret buttons and cross payment. This method can be slow and fragile.

The Walmart Agent Gateway will instead provide a structured connection based on the Model Context Protocol, or MCP. The Verge reports that the gateway is offered in beta to companies and developers. It should allow compatible assistants to query the catalog and prepare transactions without simulating all clicks of a browser.

Walmart claims to be able to offer to the agent the usual items of a customer without passing on all its history. For example, the store could indicate the bread or milk brand sought while keeping other orders in its own systems. This is an affirmation of the planned architecture; no independent audit of this data separation has yet been published.

Does the user really keep control?

Permissions reduce the risk of an entirely unexpected purchase, but they do not guarantee the quality of the decision. An agent can keep a budget while selecting a bad format, an excessive quantity or a product placed more advantageously in the results.

A [academic study on automated shopping] (https://arxiv.org/abs/2508.02630) observed that agents based on language models were sensitive to the position of products, sponsored mentions and platform recommendations. Behaviours also varied from one model to another. Conferring a purchase to the AI therefore does not remove presentation biases; It can simply make them less visible.

Confidentiality will also depend on two companies. Walmart will apply its own privacy notice to the data received, while the company providing the agent will apply its policy to the information shared in its application. The client will have to understand both.

Prices, availability and Canada

Reuters indicates that the tools must be accessible to Walmart account holders without Walmart+ subscription. No separate price or user fee was reported.

Agent Profiles is to be launched in November 2026, while the gateway begins with a beta for companies and developers. The advertisement refers to Walmart.com and Sams Club, not Walmart.ca. Walmart has confirmed neither date, compatibility nor conditions specific to Canada. Therefore, it should not be assumed that Canadian customers will be able to use these functions as soon as they launch in the United States.

Limits and Considerations

The system has not yet undergone extensive public testing. It is not known which agents will be approved at the launch, how the disputes will be divided between Walmart and their supplier, and how far the ceilings will apply to taxes, substitutions, delivery fees or modifications after the order.

The modalities promise to block agents that exceed the client's intent, but the method of detection is not detailed. An authorization of $200 also remains a real authorization: it should be limited to repetitive and easy to verify purchases, not expensive, medical or sensitive products.

Conclusion

Walmart turns the AI agent into a new user of his online store. Separate profiles, default approval and ceilings provide a more conservative basis than full access to the account.

The real test will, however, begin when these agents have to manage substitutions, promotions, quantities and returns in everyday life. For Canadians, technology is to be monitored, but it is not yet a confirmed Walmart.ca function.