AI Agents in E-Commerce: The Challenges Blocking Agentic Shopping

October 2026: retailers pick sides

TechCrunch reported on October 21 that Amazon has started blocking AI agents from making purchases on behalf of shoppers, and that travel and booking sites have locked agents out entirely. The same day, Walmart CEO John Furner announced Agent Gateway, which TechCrunch describes as Walmart's MCP server: approved agents can search Walmart and Sam's Club inventory and buy for the user.

The Walmart terms are a useful map of the problem. According to TechCrunch, agents may not act beyond the shopper's intent, may not access payment information, and may access only the data a task needs; non-compliant agents get blocked. Agent Profiles, rolling out next month, give each agent its own identity in the customer's account, with per-transaction approval or a spending limit. Walmart's returns policy applies even when an agent acts incorrectly.

Blocking and gateways are two answers to the same question: how does a merchant tell a sanctioned agent from a bot it should refuse?

Challenge 1: access and bot defences

Checkout flows were built to stop automation. Fraud systems, CAPTCHAs, rate limits, and login challenges treat a shopping agent as a scraper, because at the network level it looks like one. Browser-driving agents work around this inconsistently and break when a page changes.

The durable fix is an explicit interface. A gateway such as Walmart's, or a site exposing tools directly through something like WebMCP, lets the merchant admit known agents on its own terms instead of guessing from traffic patterns.

Challenge 2: authorization and identity

The hardest problem is delegated authority. When an agent buys something, the merchant, the payment network, and the shopper all need to answer: did a person authorize this purchase, within what limits, and can that be proven later?

A login session cannot carry that. It proves who signed in, not what they allowed. The emerging pattern is a separate agent identity with its own scope, which is what Walmart's Agent Profiles do and what Mastercard says Agent Pay for Machines addresses: who authorized the agent to spend, how much, on what, and whom it pays. Scoped credentials also limit damage when an agent is manipulated, a risk that grows with how much private data the agent can reach.

Challenge 3: payments

Handing an agent raw card details is unacceptable for most shoppers and merchants. Networks are building alternatives. PYMNTS reported on October 20 that Mastercard opened early access to Agent Pay for Machines, which lets agents pay for digital services against a budget a business sets, building on x402, an open protocol for online payment requests. American Express released a merchant playbook for agentic commerce on October 6 and announced an intent to protect eligible card members and merchants from errors by registered agents; coverage on Yahoo Finance noted it is an announced intent, with terms not yet published.

The protocol landscape is fragmented, with card networks, payment processors, and AI platforms each backing their own specifications. Merchants should expect to support more than one for a while.

Challenge 4: catalog data agents can trust

A human shopper tolerates a vague product page. An agent comparing options across stores cannot. It needs current stock, exact variants, total price including shipping, and attributes stated as data, not marketing copy. Stale inventory produces failed orders; inconsistent sizing produces returns.

This is a retrieval problem. Agents that re-read full product pages for every comparison burn tokens and still miss details buried in prose. Structured, retrievable product knowledge, served through a tool, beats scraping, and the same logic applies on the agent side: stored product research and preferences should be retrieved, not regenerated. That is the pattern RDK is built for, cutting token spend 80 to 90 percent on repeated, reference-heavy work.

Challenge 5: returns, disputes, and liability

When an agent orders the wrong size or misreads a promotion, who absorbs the cost? Walmart's answer, per TechCrunch, is that its normal returns policy applies. Amex is signalling it may cover registered agent errors. Most merchants have no policy at all.

Until that settles, audit trails are the defence: a record of what the shopper asked for, what the agent was permitted to do, and what it did. Without one, every dispute is one party's word against a log nobody kept.

What merchants should do now

Decide deliberately whether to admit agents rather than letting fraud tooling decide by accident. If admitting them, expose a sanctioned path with scoped permissions and spending controls, publish clean structured catalog data, and log every agent action. The broader adoption barriers are covered in why AI agents are hard to adopt at scale.