Agentic payments are payments made by an AI agent on behalf of a person or a business. Instead of the person clicking “buy”, the agent finds what it was asked for, checks the price against limits its owner set and pays with a credential issued for that job. In the main schemes, the card number or bank details never pass through the AI itself.

How an agentic payment works

Three things have to be in place. The owner has to give the agent authority, usually with a spending limit. The agent needs a way to pay that does not expose the underlying card or account. And the merchant and the bank have to be able to tell that a known agent is at the checkout, rather than the cardholder or a hostile bot.

Card networks handle the second part with tokens, the same technology that sits behind Apple Pay and Google Pay. Mastercard Agent Pay, announced in April 2025, issues what it calls agentic tokens: stand-ins for a card that belong to one agent and mark each payment as started by an agent. Visa launched its equivalent, Visa Intelligent Commerce, the same month. Wallets take a simpler route. Stripe’s Link can give an agent a single-use virtual card that only works for the purchase the user approved.

The third part is newer. Merchants spent years building defences against automated traffic, and a shopping agent looks a lot like a bot. Visa’s Trusted Agent Protocol, released with Cloudflare on 14 October 2025, lets approved agents sign their requests so a merchant’s checkout can let them through and keep blocking everything else.

The protocols and who is behind them

Several groups have published standards for how agents and merchants talk to each other. Google’s Agent Payments Protocol, or AP2, appeared in September 2025 with more than 60 merchants and payment companies signed up, TechCrunch reported. It records the user’s authority in two signed mandates. The intent mandate says what the user wants, such as a flight under a set price. The cart mandate approves a specific basket. If a purchase is disputed, those signed records show what the user actually agreed to. In January 2026 Google added the Universal Commerce Protocol, built with Shopify, Etsy, Wayfair and Target, so that people can buy eligible products inside Gemini and Search’s AI Mode while the retailer stays the seller of record.

OpenAI and Stripe released the Agentic Commerce Protocol on 29 September 2025, together with Instant Checkout, which let US ChatGPT users buy from Etsy sellers without leaving the chat. Stripe passes the payment to the merchant as a Shared Payment Token, so ChatGPT never sees the card. Few merchants took it up. In March 2026 OpenAI said Instant Checkout was moving into retailer apps connected to ChatGPT, Search Engine Land reported. The protocol is still in use.

Coinbase’s x402, launched in May 2025, is built for a different kind of payment. It revives a web status code that was reserved decades ago and never used, 402 Payment Required, so a website or an API can ask for payment and an agent can pay in a stablecoin as part of the same request. It suits small payments between machines, such as an agent paying per query for data, more than shopping. For how those coins work, see What is a stablecoin?

Where agentic payments are running

Most live services still ask a person to approve each purchase. Meta’s Muse agent can buy from businesses across the web through Stripe’s Link wallet, but the user confirms the total in the chat first, as Fintechbits reported. In France, Cleverbridge completed a live agent payment in September 2026 with Visa and Revolut, using a Visa Payment Passkey so the cardholder could approve it under Europe’s authentication rules (our report). In the UK, GoCardless processed a monthly charity donation set up through an AI agent over Direct Debit, inside the Financial Conduct Authority’s AI Live Testing programme (our report).

Business spending is the other early market. Companies already give employees cards with budgets and expense rules, and agents fit the same model. Corpay’s Agent Card lets AI agents create virtual cards and pay suppliers within limits set in advance (our report), and Mercury now issues virtual cards to agents alongside staff cards (our report).

The networks are also working on the fraud side. On 30 September 2026 Mastercard said it was testing a score in the US that estimates how likely it is that an agent, not a person, started a card purchase. The aim is to stop banks declining legitimate agent purchases that look odd to fraud models trained on human habits (our report).

Who pays when an agent gets it wrong

The rules on unauthorised payments were written with people in mind. In the US, Regulation E says a transfer is not unauthorised if it was made by someone the consumer gave access to, until the consumer tells the bank that person is no longer allowed. The rule does not say whether handing a payment credential to an AI agent counts the same way, or who carries the loss when the agent buys something its owner did not intend.

Europe has a practical hurdle as well. Under PSD2, most online card payments need strong customer authentication, two separate checks that only the cardholder can pass, and an agent cannot pass them alone. Today the cardholder approves each agent purchase with a passkey. If banks start accepting standing limits for trusted agents, agents could buy without that prompt. In the UK the government has named agentic payments a priority for safe use of AI in financial services, and the FCA’s live testing lets firms try these flows with real customers under supervision.

Much of the protocol design is aimed at disputes. Signed mandates, agent tokens and single-use cards all leave a record of what the owner approved, which is the evidence a bank or merchant will ask for when a customer says the agent bought the wrong thing. Purchases made through Muse are also covered by Link’s purchase protections, including returns on eligible orders, according to Meta.

Questions about agentic payments

Are agentic payments a new way to move money?

No. The money moves over existing systems: cards, bank transfers and, in some cases, stablecoins. What is new is that software starts the payment, and the credentials and checks built around that.

Does the AI agent see my card number?

Not in the main schemes. Card networks give agents tokens, wallets give them single-use cards, and Stripe’s Shared Payment Token passes the payment to the merchant without revealing the card to the agent.

Can an agent buy things without asking me?

Some protocols allow it within limits set in advance. AP2, for example, lets an agent create the cart mandate itself if the user has given detailed instructions. As of October 2026, most live consumer services still ask for approval of each purchase.

How is an agentic payment different from a subscription?

A subscription charges the same amount on a schedule agreed at sign-up. An agent decides what to buy and when, within the limits it was given, so each payment can be different.

Are agentic payments regulated?

They fall under the existing payment rules, such as Regulation E in the US and PSD2 in Europe. Those rules do not mention AI agents, so how authorisation and refunds apply is still being worked out, including through the FCA’s live testing in the UK.