Walmart Is Turning AI Shopping Into Delegated Authority

On October 21, Walmart announced an Agent Gateway for approved AI assistants to search Walmart and Sam’s Club and complete transactions, plus Agent Profiles that will let customers give each agent separate approval rules and spending limits. The business consequence is not merely easier chatbot shopping: Walmart is creating an account-and-permission layer for a nonhuman buyer. Brands now have to decide whether their product data, price, availability and proof can win when software—not a web page—does the comparison.

A storefront designed for software

Walmart’s announcement contains two capabilities at different stages. The Agent Gateway is beginning as a beta for companies and developers, giving approved agents a direct route into Walmart and Sam’s Club instead of forcing them through web pages designed for human clicks. Agent Profiles are announced for launch within a month, not yet broadly deployed.

Reuters reported that the gateway is intended to connect assistants including ChatGPT, Gemini and Muse to search products and complete purchases. Walmart says the tools will be available to account holders without requiring Walmart+.

The distinction from ordinary conversational commerce is architectural. An agent using a browser can misread a page or stall at login. An agent-native connection returns product and transaction information in a form software can use reliably.

That changes what visibility depends on. Descriptive accuracy, variants, stock, delivery promises and return conditions become persuasive inputs because an agent may evaluate them before a customer sees the page.

The account becomes the control surface

Today, an agent may sign in as the customer, mixing human and machine purchases in one history. Walmart’s proposed Agent Profiles separate those actions. A customer will be able to authorize an individual agent, see its orders independently, revoke access and choose between approval for every purchase or a spending cap such as $50, $100 or $200.

The agent will not receive the customer’s payment-card information and cannot add or change payment methods. Walmart says its normal returns process will apply when an agent does not act as intended, and the company may still demand extra approval inside an authorized limit. It also says agents can be blocked for violating its terms.

This is delegated authority made visible. The agent receives a bounded job and budget, while the retailer owns the permission system and transaction record.

That separation could become commercially valuable beyond fraud prevention. Retailers can distinguish human purchases from delegated ones and study agent-driven baskets, repeat orders, returns and substitutions. Brands may learn whether agents favor price stability, delivery certainty or familiar specifications. Those advantages are unproven, but the infrastructure makes them measurable.

Walmart also promises agents the same prices humans see, with offers and fees disclosed before purchase. If maintained, price parity becomes a trust feature as companies decide whether software should receive different treatment.

Marketing must survive without the page

The marketing risk is compression. A page can stage a story through imagery, social proof and sequence. An agent may reduce the choice to a few attributes and a recommendation. Brand advantage must survive translation from experience into evidence.

Product teams should watch where differentiation is stored. A promise that exists only in a hero image may disappear. A structured, current claim supported by policies or customer outcomes is more likely to survive the agent’s first comparison. Brand memory and product truth now meet earlier in the transaction.

For OrionPilot, the relevant discipline is continuity between strategy, content, execution and performance learning. Businesses need a clear record of the promise they approved, the evidence supporting it and the result it produced. OrionPilot does not control Walmart’s gateway; its role is to help keep the marketing system coherent while the buying interface changes.

Permission does not remove agency risk

Spending limits solve only one class of problem. An agent can remain inside a $100 cap and still choose the wrong size, optimize the wrong preference or repeat a purchase the customer no longer wants. Walmart can isolate payment credentials and require approvals, but it cannot guarantee that every outside model interprets intent correctly.

Privacy remains divided. Walmart says an agent receives only task-relevant information allowed by the customer, while its provider’s privacy policy governs data shared with the app. That matters when shopping intent reveals household routines or recurring needs.

Businesses should watch four kinds of evidence after launch: how many customers create Agent Profiles; how often they require approval instead of granting a spending limit; whether agent-originated orders produce different basket sizes, substitutions or returns; and whether brands can measure incremental sales rather than purchases shifted from Walmart’s existing channels. Reliability, adoption and commercial lift remain unproven.

Walmart is not simply welcoming another source of traffic. It is defining how software earns the right to act inside a customer account. If customers accept that model, the decisive commerce asset may no longer be the page that wins attention. It may be the permission system that turns intent into an authorized purchase.