Agent Commerce Has No Chargeback Code

You call your bank to dispute a charge. The representative asks you to pick a reason: fraud, item not received, not as described, duplicate charge. You pick one, the bank files it under a four-digit reason code, and the dispute machinery kicks in.

Every one of those codes assumes a human authorized the transaction. What happens when an AI agent did?

The reason code problem

Visa and Mastercard each organize disputes into four categories: fraud, authorization, processing errors, and consumer disputes. Within those categories, the networks maintain dozens of individual reason codes. Between the two, there's a code for almost every scenario a human shopper can create.

"Not as described" covers the jacket that looked blue online but arrived grey. "Not received" covers the package that never showed. "Unauthorized" covers the stolen card number. Each code maps to a specific failure mode with a defined liability chain: who pays, who provides evidence, and what documentation resolves it.

But an agent hallucination creates a failure mode that doesn't fit any bucket. Your AI assistant booked a $400 hotel in Austin when you asked for one in Atlanta. The hotel delivered exactly what was ordered. The payment was authorized through the agent's scoped token, and the product matched the description.

Nothing went wrong except the agent's interpretation of your intent.

Under the current system, that dispute lands on the merchant. The merchant delivered the product that was ordered, at the price that was quoted, to the payment method that authorized it. The merchant did nothing wrong.

Three parties, no clear owner

Monica Eaton, founder and CEO of Chargebacks911, framed the gap directly: "In an agentic world, 'who authorized it?' doesn't have a clean answer."

The consumer didn't authorize the specific purchase. They authorized the agent to act on their behalf within some set of parameters. The agent interpreted those parameters and made a decision. The merchant fulfilled the order.

So who absorbs the loss? Today, by default, the bank does. As Backbase put it in an August analysis: "Your bank is liable for agents it never built."

The bank sits under the transaction rail. When a dispute doesn't fit an existing code, the bank eats the cost or forces it onto the merchant through the closest matching reason code, which misrepresents what actually happened.

The chargeback system was built for disputes between a human buyer and a human seller. Agent commerce inserts a third actor, the AI, and none of the existing liability frameworks account for it.

Three protocols, zero dispute resolution

Three competing standards are racing to define how agents pay: ACP from Stripe and OpenAI, AP2 from Google and the FIDO Alliance, and x402 from Coinbase and the Linux Foundation. Each takes a different approach to authorization, the question of how an agent proves it has permission to spend money.

ACP standardizes checkout flows between agents and merchants. AP2 defines a trust and authorization layer with 60-plus partners. x402 uses HTTP's 402 status code to make payments native to web requests.

All three solve authorization. None of them define what happens after the purchase goes wrong.

None of the protocols include a dispute resolution spec, a chargeback flow, or a refund mechanism. They tell the agent how to pay. They don't tell anyone what to do when the agent pays for the wrong thing.

Why this gap matters now

Agent-initiated transactions are still a rounding error in total payment volume. But the Anthropic commerce agent spec launched with Visa and Mastercard in September. Google's AP2 has 60-plus partners. ChatGPT's Instant Checkout runs on ACP.

The payment rails for agent commerce are being laid right now. The dispute rails haven't started.

Now that agents are executing real transactions through real card rails, this becomes a liability question the card networks, the protocol designers, and the banks all have to answer. Visa and Mastercard can add a reason code. The protocol builders can add a dispute layer. Or the banks can keep absorbing losses for a failure mode that has no name.

The first card network to publish an agent-specific reason code will signal that agentic commerce has moved from pilot to production. That code doesn't exist yet.