AIToday
Large Language ModelsAI Regulation & PolicyAI Business & IndustryFortune AIPublished: Nov 21, 2026, 22:01 JST

OpenAI wallet terms leave agent-judgment losses unresolved, Fortune argues

Send on LINE
OpenAI wallet terms leave agent-judgment losses unresolved, Fortune argues

3 Key Points

  1. What happened

    A Fortune commentary notes OpenAI's terms leave delivery and returns with merchants and keep existing card dispute protections, but don't say who bears losses when an agent's judgment, not its authorization, is at fault.

  2. Why it matters

    The piece argues that when an agent acts within its spending authority but outside the user's intent, no rule today decides whether the user, model provider, merchant, platform or payment provider absorbs the loss, so each such failure is likely to become a dispute.

Not sure about something? Ask the AI

Questions and answers are published on this page.

Summaries like this, every morning.

Context & Analysis

The commentary opens by noting that today's safeguard of requiring approval for every purchase will not last, because an agent that needs approval for each transaction is little more than a faster checkout button. Once approval moves from the transaction to the instruction, payments lose the assumption they were built on: that a human decides, at the moment of payment, that this is what they want. The industry, it says, is making rapid progress on how agents pay with existing methods but has not solved what agents are allowed to decide or who pays when they decide wrong.

The piece points to Google's Agent Payments Protocol, which uses cryptographically signed mandates to record what a user asked for and approved, as necessary but not sufficient. A signature can prove what was authorized, but it cannot prove that the agent interpreted the user's preferences correctly, acted in their interest, or was free of outside incentives. To close that gap, the commentary argues that the industry's Know Your Agent concept must expand beyond verifying whose agent is acting and what it was authorized to do, to also verifying whether the agent exercised that authority as intended, through mandates that capture intent, a liability framework, and a duty of loyalty.

On liability, the commentary notes that no rule today decides among the user who gave an ambiguous instruction, the provider whose model misread it, the merchant whose agent extracted worse terms, the platform that ranked the options, or the payment provider that executed the transaction. It draws on precedents such as the card industry's EMV liability shift and product liability law, and suggests that risk which can be assigned can be priced and insured. The author, who says they worked on both sides of the rulebook as chair of the FATF's Risk Group and as a fintech executive, argues that private standards are structurally weak on liability because every party at the table would prefer the loss to land somewhere else.

FAQ
What example does the commentary use to show the problem?
It describes an agent asked to plan a family vacation within a $5,000 budget that books nonrefundable flights on the wrong weekend, a hotel 40 minutes from the beach, and accepts an airline counteroffer that strips out checked bags.
Does OpenAI take a cut of Wallet transactions at launch?
No, the commentary states that OpenAI is not taking a cut of Wallet transactions at launch, though it notes business models evolve.
What does the commentary say the card industry's EMV liability shift did?
It made the party that had not adopted chip technology bear certain counterfeit-fraud losses, creating a powerful incentive to upgrade.
Send on LINE

AI news that matters for your work, delivered every morning.

Pick your industry and the AI tools you use, and get news related to your work every day.

Free · 30 seconds with Google · unsubscribe anytimeWhat is AIToday? →

Ask AI

Ask AI anything about this article. The AI reads this article, earlier AIToday articles, and Wikipedia, and cites its sources. Q&As are published on this page for other readers too.

Questions and answers are published on this page.

Related Articles

Next articleSolidWizard sees AI, chip orders through 2027