Walmart Plans to Create a Separate Database for AI: Commercializing "Agent"—First, Resolving Who Grants Authorization and Who Pays the Bill

It’s not hard to get an AI to find a carton of milk. But getting it to choose the right size, use the correct payment method, know when to stop when prices change, and provide a clear resolution process if it makes a mistake—that’s what comes close to a sustainable business.

Wal-Mart, Walmart (retailer) On October 21, details were announced regardingShopping AgentArrangements:Agent “Profiles” will create separate profiles for agents within user accounts and is expected to launch within the next month; the Walmart Agent Gateway, meanwhile, will begin with testing among businesses and developers. The term "separate" here refers to the distinction between operational identities and records—it does not mean that the AI will become an independent payer.Official Announcement from Walmart

This issue is worth noting because it brings Agent shopping into the realm of the permissions, records, and post-sale issues required for actual transactions.

Break Down Single Sign-On into Manageable Authorizations

According to the published design, users can approve purchases on a transaction-by-transaction basis or set spending limits; proxy purchase records are displayed separately, and users can delete proxy profiles. By default, purchases require approval, and Walmart may also require additional confirmation for purchases made within the spending limit. Proxies cannot access bank card information, nor can they add or modify payment methods.Authorization and Payment Rules

It aims to resolve a common issue of permission mismatch: a user may simply want to have the assistant restock items, but the system might require the assistant to have login credentials sufficient to manage the entire account. Only by separating the proxy’s actions from those of the main account can we discuss the scope of a specific authorization, how to revoke it, and who is responsible for restoring the process in the event of an error.

Spending limits can only control the amount; they cannot fully express purchasing intent. Purchasing the wrong size, placing duplicate orders, or substituting with unacceptable alternatives can all occur even within the spending limit. Therefore, when designing shopping bots, developers should treat “how much to spend” and “what to buy” as separate considerations. A single, explicit restocking instruction should not automatically become an authorization for ongoing purchases.

The gateway changes the transaction interface.

The Gateway page emphasizes direct access to product and fulfillment data, provides authentication, access control, and observable operations, and sets category and frequency limits. It remains in the early stages of invitation-only testing; some features may be subject to change and may not be available to all participants.Gateway Official Page

If these interfaces are successfully implemented, developers will be able to reduce their reliance on page layouts, button placements, and temporary pop-up windows. Retailers will also be able to identify agents and enforce rules at the system entry point, rather than having to speculate after the fact about what happened during a typical user session. These are benefits at the product architecture level and should not yet be considered as measured success rates or cost reductions.

More profound changes may occur in channel relationships. Users can express their needs through their preferred assistants, while retailers retain control over inventory, fulfillment, and order processing. For Walmart, this provides a way to continue facilitating transactions as the point of purchase shifts elsewhere; for third-party vendors, it means they do not have to rebuild the entire retail supply chain on their own. This represents an assessment of the division of labor in the industry; based on current announcements alone, it is not yet possible to deduce the fee or revenue-sharing models for either party.

To determine whether this is commercially viable, we must also calculate the costs associated with abnormal orders. If manual intervention is required every time a product is out of stock and needs to be substituted, or if users frequently overturn the system’s recommendations, the time saved on clicks may not necessarily offset the costs of customer service and returns. In addition to monitoring order conversion rates, the pilot program should also track instances of manual intervention, repeat purchases, and refund processing to determine whether this channel is worth expanding.

Order History and Returns Determine Whether Users Are Willing to Let Go

Walmart stated that its return policy still applies when agents do not act as expected. Gateway also plans to help agents identify items that users frequently purchase without disclosing their broader purchase history.Order and Data Arrangements

These two points correspond to different costs of trust. The former lets users know that there is still a way to resolve issues if something goes wrong; the latter attempts to narrow the trade-off between personalization and data exposure. Knowing “which brand of milk to buy this time” does not inherently require knowing what a household has purchased over the long term.

However, simply following the standard return process should not be interpreted as the platform promising unconditional compensation for all agent errors. For consumers, truly useful records should also answer the following questions: Who initiated the purchase? When was it approved? Did the final product or price change? And what stage is the refund process at? Complete transaction records are more effective at building trust than a simple “Task Completed” message.

Merchants' preparations will focus on product data.

From a merchant’s perspective, when agents take on more of the screening work, the comparability of product information becomes even more important. For the same everyday item, if details such as specifications, unit price, delivery areas, estimated arrival time, and return/exchange policies are scattered across different pages, it becomes easier for the assistant to overlook certain conditions when comparing products.

This provides brands and sellers with a more concrete direction than simply chasing traffic: first, ensure that product specifications, variants, and fulfillment conditions are accurate; then consider how to be understood by agents. More accurate data does not guarantee a recommendation, but it can reduce the likelihood of being mistakenly selected or overlooked due to incomplete information. For orders generated by agents, it’s also worth monitoring cancellation, replacement, and return rates separately—rather than focusing solely on the number of completed transactions.

For developers, pilot acceptance testing should also cover exception handling: whether the process stops when authorization expires, whether it reconfirms after inventory changes, whether it prevents duplicate orders when payment results are unclear, and whether it can resume after a task is interrupted. A single successful demonstration is unlikely to fully address these operational issues.

Walmart’s plan is still in its early stages, but it points to a clear direction: for shopping agents to be adopted for the long term, authorization, payment, and post-transaction traceability must all be sufficiently transparent. Consumers are generally willing to delegate tasks with clearly defined boundaries; the system must demonstrate that it can stay within those boundaries.