ChatGPT Wallet Can Spend Your Money. Nobody Knows Who Pays When It's Wrong
OpenAI's ChatGPT Wallet lets agents pay for users — with approval per purchase. A Fortune commentary argues the real gap is liability when approval moves to the instruction.
By Grace Kim
3 min read

What's News
- OpenAI launched ChatGPT Wallet last week, its second attempt at agentic payments after shelving in-chat checkout in March.
- Every Wallet purchase currently requires human approval.
- Google's Agent Payments Protocol uses cryptographically signed mandates to record user approvals.
- OpenAI takes no cut of Wallet transactions at launch.
- The EMV liability shift is cited as a precedent for pre-assigning losses.
OpenAI launched ChatGPT Wallet last week — its second attempt at agentic payments — and every purchase an agent makes still needs a human to approve it. That safeguard won't last, argues a new Fortune commentary from a former FATF Risk Group chair, and the industry has no answer for what happens when it goes.
OpenAI shelved in-chat checkout earlier this year, TechCrunch reported in March. Wallet is the retry: agents that pay merchants on a user's behalf, with approval required per transaction. The commentary's core argument is blunt. "An agent that needs your approval for every purchase is a faster checkout button, not an agent." 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.
What happens when a purchase is authorized, delivered — and wrong?
The author's thought experiment: ask your agent to plan a family vacation within a $5,000 budget. It books nonrefundable flights on the wrong weekend, a hotel 40 minutes from the beach, and accepts a counteroffer from the airline's own agent that quietly strips out checked bags. Every payment was authorized. Every merchant delivered exactly what was ordered. The trip is still nothing like what you asked for.
OpenAI's terms leave delivery and returns with merchants and preserve existing card dispute protections. But agents introduce a different failure mode: a properly authorized, properly fulfilled purchase that didn't reflect the user's instructions. "Authorization to spend is not authorization to judge," the piece reads. "Today's infrastructure only knows how to verify the first."
What is the agentic stack missing?
Google has started building the identity layer. Its Agent Payments Protocol uses cryptographically signed mandates to record what a user asked for and approved. That's necessary but not sufficient, the commentary argues: a signature proves what was authorized, not that the agent interpreted the user's preferences correctly or was free of outside incentives. The fix is a broader version of what the industry calls Know Your Agent — verifying not just whose agent is acting, but whether it exercised its authority as intended. Three layers are needed.
- Mandates that capture intent. Today's permissions center on price limits, timing and product parameters. A real mandate encodes priorities, trade-offs the user rejects, and escalation conditions — "nothing nonrefundable without asking me." A spending limit tells the system how much. A mandate tells it what for.
- A liability framework. When an agent acts inside its spending authority but outside its mandate, the loss could land on the user, the model provider, the merchant's agent, the ranking platform or the payment provider. No rule decides among them today. The precedent exists: the card industry's EMV liability shift made non-adopters of chip technology bear counterfeit-fraud losses, forcing upgrades. Assigned risk can be priced and insured.
- A duty of loyalty. OpenAI isn't taking a cut of Wallet transactions at launch, but business models evolve. Once a platform earns merchant fees, paid placement or advertising, its agent faces a conflict the user can't see — the same conflict financial regulation addressed for brokers through disclosure and best-interest obligations.
Who writes the rules?
AI labs, payment companies and commerce platforms are writing agentic protocols right now. Private standards handle identity and payment flows well, because everyone benefits from interoperability. They're structurally weak on liability, because every party at the table would prefer the loss to land somewhere else.
The author writes from experience on both sides — as chair of the FATF's Risk Group and as a fintech executive: the rules that worked were the ones that settled, in advance, who absorbs failure. "Systems that answer that question scale. Systems that leave it open stall, or have the answer imposed by regulators after a visible failure."
Card payments became global infrastructure because people believed someone would make it right when something went wrong. Agents need the same promise. Whoever makes it credible will do more than enable agentic payments — they'll help write the rules of the new economy.
Source: fortune.com
More from Grace Kim
Show full bio
Market editor covering people, launches and day-to-day practice in entrepreneurship, startups and venture funding at Business Bearings.
2,582 articles