AI Agents Can Shop. Who Handles Refunds?

Amazon’s Bedrock AgentCore Payments, developed with Coinbase and Stripe, lets AI agents find paid services, authenticate users and make payments with stablecoins, cards and other payment rails. An Oct. 6 review by the Reserve Bank of Australia found that liability, fulfillment and refunds remain unclear when agents buy goods or services.

The RBA’s payments consultation drew written submissions from 75 stakeholders, including merchants, payment providers and card issuers. Participants reported that existing chargeback rules may not assign responsibility when an agent acts outside a customer’s instructions. They also warned that agent-led purchases could increase merchant costs. One stakeholder cited reports of an additional charge of about 4% for purchases assisted by AI. The RBA described adoption as early and evidence of harm as limited.

Edgars Nemse, CEO of the GenLayer Foundation, told CryptoSlate that payment is easier to verify than fulfillment. “The money moved, or it didn’t,” he said. Determining whether a service was delivered as promised requires judgment. A signed mandate or spending limit can show what a customer authorized, but it cannot establish whether the result met the customer’s expectations.

Agents could file complaints at little or no cost, while merchants would need to provide receipts, delivery records and processor data. Nemse expects agents to “dispute far more often, because disputing costs them nothing.”

Mastercard and Datos projected 324 million chargebacks worldwide by 2028. A 2026 U.S. merchant benchmark put the average operational cost of a chargeback at $128, excluding the value of lost goods or services. Applying that U.S. figure to the global estimate as an illustration, a 5% increase would create 16.2 million additional cases and about $2.1 billion in operational costs. A 15% increase would create 48.6 million additional cases and about $6.2 billion.

Platforms currently provide much of the dispute infrastructure. Amazon blocked Meta’s Muse shopping agent, citing unauthorized access and its policies. Nemse described the dispute as a fight over control of the customer interface, purchase data and advertising space. He expects Google to favor its own shopping agents. Shopify has moved toward allowing browser-based AI shopping agents to reach checkout.

Agents that search the open internet could direct customers to smaller merchants instead of large platforms. Disputes may be harder to resolve when the purchase takes place outside a platform with established refund procedures. “Without it, your agent finds the small merchant, and you still go back to Amazon,” Nemse said.

Google’s AP2, Mastercard Agent Pay and Visa Intelligent Commerce focus on authorization and identity. Their tools include signed mandates, tokenized credentials, spending controls and agent identity checks. They can establish whether an agent was authorized to make a purchase, but they do not determine whether subjective work, such as a completed insurance claim or software task, met the customer’s expectations.

One proposed model uses AI validators to review evidence and reach a decision through consensus. The GenLayer Foundation estimates that routine cases could be resolved in about 30 minutes and fully escalated cases in about three hours. Deposits, fees or reputation penalties could discourage unsupported claims. Appeals could be available when a decision is disputed. An on-chain decision could control escrowed funds, but it would not automatically force an ordinary merchant to issue a refund through a card network.

A CI&T survey of 1,011 U.S. consumers found that 27% were comfortable with fully automated AI shopping. The RBA plans to announce its regulatory priorities by the end of 2026.