When a purchase goes wrong, the complaint arrives as a dispute code, a transaction ID, a merchant category, a timestamp, and a reason selected from a dropdown. The chargeback is one of the most structured forms of disagreement in commercial life.
That structure is now shaping how payment networks think about AI agents.
In June 2026, AP reported that Visa's collaboration with OpenAI would let users link Visa cards to ChatGPT so agents could shop on their behalf. The guardrails Visa described are worth reading closely: spending limits, required approval steps, approved merchants, fraud monitoring. Most early transactions would loop in a human, with the agent sending a notification for the consumer to approve the actual purchase.
Visa's Jack Forestell told AP something interesting about disputes. The straightforward cases are when the consumer didn't intend the purchase, or the merchant processed it incorrectly. The harder case, he said, is when consumer intent and merchant processing are both correct, but something in the middle caused a problem.
That something in the middle is the agent. And preventing that problem means building records that can distinguish what the agent did from what the consumer authorized.
Mastercard's Agent Pay announcement from April 2025 works the same seam from a different angle. Trusted AI agents must be registered and verified before making payments. Transactions facilitated by agents should be recognizable to consumers, issuers, and merchants. There's a process to clarify "unfamiliar or unrecognized" agentic transactions. The language throughout is about making agent activity legible after the fact, because someone will eventually ask.
The Agentic Commerce Protocol, maintained by OpenAI and Stripe, goes further into the structural details. Its order specification defines adjustment types including refund, credit, return, exchange, price adjustment, cancellation, and dispute. Its delegate payment mechanism issues scoped tokens with explicit bounds: a reason, a maximum amount, a currency, a checkout session, a merchant, an expiration. Stripe tags agentic-commerce orders with the originating agent's name, filterable in the dashboard. Every transaction carries its own provenance.
Agents can already navigate a checkout flow. All of this infrastructure exists for what comes after: someone files a dispute, and that dispute needs to land on a specific party with specific evidence. Who authorized this purchase. Within what limits. Through which agent. With what scope.
Payments are forcing structure onto agent activity because money, when it goes wrong, generates structured complaints. And structured complaints demand records that can tell the difference between performing an action and being authorized to perform it.