AI agent payments took a step forward last week when OpenAI launched ChatGPT Wallet. Its agents can now pay merchants for a user, though a person still has to approve each purchase. It’s OpenAI’s second try at agentic payments, after it shelved in-chat checkout earlier this year.
That approval step won’t last. An agent that needs sign-off on every purchase is just a faster checkout button. Once people approve the instruction instead of the transaction, payments lose their old assumption that a person chose, at the instant of payment, to buy exactly this.
Approved, delivered and still wrong
Picture asking an agent to plan a family vacation on a $5,000 budget. It reserves nonrefundable flights on the wrong weekend and a hotel 40 minutes from the beach. It also accepts a counteroffer from the airline’s own agent that quietly drops checked bags.
Every payment was authorized and every merchant delivered what was ordered. The trip still looks nothing like the request.
Under OpenAI’s terms, merchants handle delivery and returns, and current card dispute protections stay in place. Those may give some remedy, but they don’t answer who absorbs the loss when the agent’s own judgment failed. Permission to spend isn’t permission to judge, and today’s systems can only verify the first.
What the agent payment stack lacks
The industry is working on verifiable mandates. Google’s Agent Payments Protocol relies on cryptographically signed mandates that log the user’s request and approval. A signature proves what was authorized. It can’t show that the agent understood the user’s wishes, served their interests or had no outside incentive.
The idea of Know Your Agent today means checking whose agent is acting and what it may do. It also needs to cover whether the agent used that authority as intended. That takes three layers.
Mandates that capture intent
Today’s agent permissions mostly cover price caps, timing and product specs. A fuller mandate would record the user’s priorities, the trade-offs they accept or reject, and the conditions that require a check-in, such as “nothing nonrefundable without asking me.” Merchant agents, payment providers and any later arbiter would need to verify those limits without seeing the user’s private instructions. A spending limit says how much. A mandate says what for.
A liability framework
Say an agent respects its spending authority but breaks its mandate. The loss might fall on a user with an ambiguous instruction, a provider whose model misread it, a merchant agent that squeezed out worse terms, a platform that ranked the options, or the payment firm that processed it. No rule picks among them, so each failure is likely to become a dispute.
Settling it in advance would build consumer trust. One yardstick is who could have prevented the harm most cheaply. The card industry’s EMV liability shift put certain counterfeit-fraud losses on whoever hadn’t moved to chip technology, giving it a strong reason to upgrade. Product liability law assigns defects along the supply chain. Clear assignment also builds a market, because a risk with an owner can be priced and insured.
A duty of loyalty
OpenAI takes no share of Wallet transactions at launch, yet business models shift. When a platform starts earning from merchants, whether via fees, paid placement or ads, its agent faces a conflict of interest its user can’t see. Financial rules tackled the same issue for brokers and advisers through disclosure duties and best-interest obligations. Consumer agents need enforceable standards covering whom they serve, how paid influence is revealed and how an outside party can check that the mandate was followed.
Who writes the rules for agent payments
AI labs, payment companies and commerce platforms are writing the agent protocols at this moment. Private standards work for identity, credentials and payment flows, where everyone benefits from systems that interoperate. Liability is their weak spot, since each party would rather the loss fall on someone else.
Systems that settle in advance who absorbs failure tend to scale. Those that leave it open stall, or end up with regulators dictating an answer once a failure becomes public, typically shaped by the previous crisis.
Cards grew into global infrastructure because users trusted that someone would fix things after a problem. Agents need that same assurance spanning consumers, AI platforms, merchants and banks. Companies that build mandates, liability rules and loyalty into their agents early are the likelier ones to win consumer trust with the wallet.
Source: Fortune