
Agentic payments are purchases an AI agent completes on a person's behalf. Since November 13, 2026, they run through ChatGPT. OpenAI's ChatGPT Wallet, built with Visa, Mastercard, and American Express, lets its agents pay with a token instead of a card number, starting with Pro subscribers in the US. Every purchase still waits for Face ID.
That is the right place to start. If an agent buys the wrong thing, card rules hold the cardholder responsible, and no network has written a rule for agent mistakes yet. Approving each purchase contains that risk. But nobody hands over their weekly shop one fingerprint at a time. What comes next is the mandate: a standing instruction that sets what an agent may buy, how much it may spend, where, and until when—with an off switch. If banks don't build it into their own apps, people will set those rules inside ChatGPT, and the bank will only see the charge.
What ChatGPT Wallet gets right
Start with what happens when an agent gets it wrong. Under Visa's Core Rules, the cardholder is responsible for an agent's actions as if they had made the purchase themselves. Fraud is covered. A stolen card used through an agent still falls under existing protections. Mistakes are not. Neither Visa nor Mastercard has a binding dispute rule for the case where the agent, not a fraudster, buys the wrong thing. OpenAI's own terms send most payment and order disputes to the merchant or the card issuer.
The ways an agent can get it wrong are already documented. Fraud teams describe prompt injection hidden in listings, ads, and reviews that steers an agent towards another seller, a bigger quantity, or a higher price. Meanwhile, the signals fraud models rely on disappear when software does the shopping: mouse movement, device, location. Riskified found that up to 34 percent of the data its models use was missing from ChatGPT's Instant Checkout.
So OpenAI made the person the control. Before each purchase, they see the items, the total, and the card, then approve with Face ID or a fingerprint. If a material detail changes before checkout, the wallet asks again. It's the same on-screen confirmation that earns trust in healthcare conversations, applied to money. Each approval leaves a record, and for now that record is the clearest proof of what the person authorised.
Our fintech clients have landed in the same place. Most of the ones I've worked with aren't ready to give an agent that kind of autonomy. Whenever the question came up, we arrived at the same pattern: the agent does the work—prepares the transaction, fills in the form—and the person confirms it. Two things hold it there. The first is regulation. In some markets, the law doesn't yet cover payments an agent makes on its own, and banks are limited in what they can switch on. The second is trust. Few people are ready to hand their money to an agent. Showing where the data comes from helps, but it isn't enough. Trust has to be built: simple tasks first, each one confirmed, with control the person loosens over time.
Per-purchase approval isn't caution for its own sake. It is the evidence layer the rules haven't built yet.
OpenAI says spending limits, purchase categories, and time periods will come later. No date yet.
OpenAI says spending limits, purchase categories, and time periods will come later. No date yet.
Where per-purchase approval breaks
The trouble with asking people to approve everything is that they do. Before Anthropic changed the default, Claude Code users approved 97 percent of permission prompts. A biometric check is the same dialog with a better lock. By the tenth purchase, Face ID stops being a decision and becomes a gesture.
It also limits what an agent is for. Approving a pair of headphones works. Approving every item in a weekly grocery order, a subscription renewal, or a train ticket that sells out at 6 a.m. does not. The more useful the task, the less it tolerates a person at every step.
And the evidence weakens with volume. A thousand approvals show that a person pressed a button. They don't show what the person wanted. If a prompt injection pushed the agent to a different seller, the approval proves consent to a cart, not to the reasoning behind it.
Per-purchase approval protects the first purchase. It can't carry the hundredth.
How mandates scale agentic payments
A mandate is a standing instruction a person gives their agent: what it may buy, how much it may spend, where, and until when. British readers already know the word from Direct Debit mandates. The difference is choice. A Direct Debit lets one company collect one kind of payment. An agent mandate lets software decide what to buy, inside limits the person set.
It moves the decision from the moment of purchase to the moment of setup. The person decides once: this agent may buy groceries and household goods, up to $150 a week, from these merchants, until the end of the month. Purchases inside the rules go through. Anything outside them comes back for approval. A mandate is still an explicit yes. It is given once, with the full picture, instead of a hundred times mid-task.
The building blocks for AI agent payments already exist. Google's Agent Payments Protocol, now being standardised at the FIDO Alliance, lets a person sign a mandate upfront so an agent can complete a purchase later without them present. Mastercard's Agentic Tokens are scoped to one agent and a set of merchants, and can be revoked in real time. Fintechs ship the controls too. Lightspark lets payments under a threshold run automatically and holds anything above it. Robinhood's agentic card comes with monthly limits and purchase alerts.
None of them describe how trust grows. That is the design problem. Banks solved a version of it long ago: a new customer gets a small credit limit, builds a history, and earns a bigger one. Agents need the same path. Start with per-purchase approval. Use that history to suggest the first narrow mandate: you've approved every grocery order this month, so should your agent handle the ones under $80? Widen it while the record stays clean. Pull it back the moment something looks wrong.
The path won't move at the same speed everywhere. Where the law keeps a person on the final step, it stops at approval for now. And every step has to explain itself. A mandate people can't read is one they won't grant. Show what the agent can spend, where, and why it's asking for more.
That turns approval history from a log into an asset. Every Face ID tap today is data for tomorrow's mandate. Right now, that data sits with OpenAI.
Credit limits grow with history. Agent mandates should too.
Credit limits grow with history. Agent mandates should too.
Who sets the rules keeps the customer
Large institutions see the same risks. So far, their answer is no. On November 11, United Airlines barred unauthorised third-party agents from booking or paying for travel. A traveller's permission, United says, does not replace the airline's approval. Three days earlier, five UK lenders restricted AI agents from customer accounts, including HSBC, Barclays, and Lloyds. In the US, Morgan Stanley began blocking third-party agents, while JPMorgan, Bank of America, and Citi restricted them. The reasons overlap. UK banks point to unclear liability for an agent's mistakes and the risk of automated switching draining deposits. US banks say they can't tell a customer from a bot logging in with the customer's password.
Blocking buys time. It doesn't answer the question, and it closes only one door. ChatGPT Wallet pays through card network tokens, so the bank still approves the charge. It doesn't set the rules the agent followed. Spending authority is moving to the agent platform, which decides what an agent may buy, from whom, and up to what amount before the network sees the transaction. Merchants face the same squeeze, as Amazon's standoff with Meta's Muse showed.
If that layer lives in ChatGPT, the bank becomes the rail underneath. It approves charges it can't explain and handles disputes it didn't design. But banks hold what an agent platform doesn't: the account, the transaction history, and the dispute process. In DBS's pilot with Visa, the bank checks whether an agent's action fits the customer's permissions before money moves. That points to the right model: the bank as the place where the mandate is set, not only enforced.
The bank that designs the mandate keeps the relationship. The bank that enforces someone else's becomes infrastructure.
In three weeks, four US banks, five UK lenders, and United Airlines restricted outside AI agents.
In three weeks, four US banks, five UK lenders, and United Airlines restricted outside AI agents.
What banks should design first
Start with the card. It doesn't need to be a special agent card. An ordinary card the customer opens and sets aside for the agent will do, as long as its controls are good enough. It's the easiest first step I know. Most banks offer a daily cap and a limit on online payments. Not every bank offers limits by merchant category, and where it does, the setting is often buried in the app. That is the gap. An agent that only buys groceries should run on a card that can't pay anywhere else. Put those limits on the bank's side, where the agent can't change them. Limits inside ChatGPT, once OpenAI ships them, will be set and enforced by the same company that runs the agent. A bank-side limit is a second lock, held by someone else.
Then the agent list. One screen showing which agents can use which card, what each may do, and what each has spent. Today, few customers could answer a simple question: who can spend my money?
Then the first mandate. Build it from what the bank already sees: past purchases by merchant and category. Walmart's agent profiles let customers pick limits such as $50, $100, or $200. A bank can go further and suggest a limit the customer has already lived with.
Then the off switch. One tap to pause or revoke an agent, effective immediately, across every merchant.
Finally, the receipt. Every agent purchase stamped with the mandate it ran under, readable by the customer and usable in a dispute. These are the action logs we flagged after OpenAI's DevDay, now with money attached.
Face ID got agentic payments started. Whoever designs what replaces it decides where the customer goes next.
Where should your customers set the limits?
Discuss with your AI.

Victor has 13+ years of experience in fintech, retail, healthcare, and digital product design. He is also a part of our internal R&D stream, exploring how AI can enhance design workflows and help teams move from concept to prototype faster.
