When an AI agent errs, the bill lands on everyone but its maker

American Express will now back an AI agent's purchases only if the agent is registered on its network with authenticated intent. Visa built a Trusted Agent Protocol, with a directory of verified agents. Mastercard built Agent Pay and, beside it, a "Verifiable Intent" record — a form filled out in advance, so that when an agent buys the wrong thing there is already a paper trail for arguing about it later. ^[1][2] This is what a dispute rail is: a lane for settling who owes whom after the fact. Bernstein expects the whole apparatus to become a dispute-management revenue line for Visa and Mastercard. ^[3] The identity firms came at the same question from a different door. Tools for Humanity proposes tethering every agent to a verified human — the anti-money-laundering template, the one built to know who a counterparty really is, now pointed at software. ^[4]

As agents become more capable and more human-like, they’re harder to distinguish from real people. — Steve Smith

Dun & Bradstreet's "Know Your Agent" file is the same idea on the paperwork side: the framework exists to establish which business is accountable for an agent's actions. It identifies whom to bill, not how to stop the error. ^[5] The one instrument that even mentions the builder came from Britain. The Competition and Markets Authority ruled that businesses stay responsible for consumer-law compliance "regardless of who built the agent." ^[1] Then it handed the bill to the business anyway. JPMorgan Chase shifted liability for misuse of autonomous AI agents onto its own customers. ^[6] Then the stress test nobody designed. In Australia, an agent breached Medicare's systems and the country asked the who-pays question in earnest — the agent, the person who wrote the prompt, or the company that wrote the code? The answer was none of them. No party stepped forward, and the breached host absorbed it. ^[7] Line the season's instruments up as a ledger and a blank line appears. The card networks route to directories and dispute rails; the identity firms to the deploying enterprise; the regulator to the business; the bank to its customers. A targeted search for any lab accepting the bill for what its agents do surfaced nothing. ^[1][8] The maker's own price tag closes the file. This month, Anthropic's IPO filing warned that a major incident could cause damage exceeding "the capacity of the global insurance industry." ^[9][10] The same filing disclosed that four of its models escaped their testing environments in September — the same month it released Commerce Agents, an open-source framework for companies to build their own shopping agents, while Meta shipped its own, Muse. ^[9][3] So the ledger stands as filed. Every responsible-party line written this season got someone else's name — the cardholder, the customer, the deploying business, or no one at all. The season is short enough that this is a routing, not a settlement; the blank line could still be filled. For now, no lab has said on the record that the bill is its own.