Agentic payments, explained: how AI agents pay and who is behind it

Agentic payments explained in one line: an AI agent, not you, clicks “buy,” but only inside limits you set in advance, using a card token or a stablecoin wallet it has been allowed to use. Six competing protocols from OpenAI and Stripe, Google, Visa, Mastercard and Coinbase define how that works, and most of it is still early, uneven and partly in pilot.

How we checked this

We used the official pages and specifications for the six protocols most often named in this space (ACP, UCP, AP2, Visa Trusted Agent Protocol, Mastercard Agent Pay and x402), plus OpenAI’s Instant Checkout announcement. Market figures come from an IMF paper (2026), Deloitte’s 2026 stablecoin prediction, and industry guides from Airwallex, Checkout.com and Eco; vendor figures are labeled as vendor claims. Facts checked on September 24, 2026. For who supports which protocol, see the protocol tracker.

What are agentic payments, concretely?

Today most online payments work like this: you find a product, you type or autofill your card, you press a button. You are the one who decides and the one who clicks. An agentic payment moves both of those steps to software. Airwallex’s guide puts it plainly: agentic payments are “transactions that an AI agent initiates, evaluates, and completes on a buyer’s behalf, operating inside spending rules the buyer sets in advance, rather than a person manually approving each purchase” (Airwallex). The rules are things like a budget, delivery preferences and approved stores.

The IMF describes the wider category, agentic AI, as “autonomous systems that sense their environment, set goals, and perform multistep tasks with little human input,” and gives an example that makes it concrete: a system that tracks a household’s propane tank, finds the cheapest supplier and arranges the refill on its own (IMF, 2026). That is the shift in one sentence: from “click to pay” to “decide to pay.”

It helps to break one purchase into four parts:

1. Intent. You state a goal, such as “trail-running shoes under $150 that arrive Friday” (Eco’s example, Eco).
2. Agent. An assistant such as ChatGPT or Gemini, or a dedicated shopping agent, compares options and picks one.
3. Merchant. The store receives an order that did not come from a human clicking on its website, and has to decide whether to trust it.
4. Rail. Money moves, over a card network or over a stablecoin network.

Every protocol in this hub is an attempt to standardize one or more of those hand-offs. Some focus on the conversation between agent and store, some on proving the agent is legitimate, and some on the money itself.

How the money moves: the flow in plain language

Here is the typical sequence, with the caveat that each company implements it differently.

First, you give the agent permission. That can be a saved card in an app, a wallet with a balance, or a one-time approval screen. Second, the agent talks to the store. Instead of reading a web page built for humans, it uses a structured interface. Google’s Universal Commerce Protocol, for example, returns a checkout object with line items, prices, totals and a status such as ready_for_complete (Google Developers Blog).

Third, the payment credential changes shape. On card rails, the common pattern is to hand the store a limited token rather than your real card number. Eco, a payments company, describes ACP’s version as a Shared Payment Token “bound to a specific merchant and a specific dollar amount, time-bounded, single-use” (Eco). Mastercard’s program uses “Mastercard Agentic Tokens,” which it says build on the same tokenization already used for contactless phone payments and card-on-file (Mastercard, April 2025).

On stablecoin rails, the agent signs a payment from a wallet. With x402, a server answers a request with the HTTP status “402 Payment Required,” listing a price and accepted tokens; the agent sends a signed payment in a header and retries (Coinbase, May 6, 2025).

Fourth, the store fulfills the order and handles support, refunds and returns much as it does now. That last step matters more than it sounds, because it is where most problems will surface.

The six protocols in one table

Every row below comes only from the protocol owner’s own page or repository. For deeper side-by-sides, see ACP vs UCP: OpenAI’s and Google’s checkout protocols compared and the protocol tracker.

The x402 guide covers the protocol in more detail.

How the pieces fit: ACP and UCP are mostly about the conversation between an agent and a store (catalog, cart, checkout). AP2 is about proof: who authorized what. Visa’s and Mastercard’s programs sit on the card side, giving agents recognizable identities and limited tokens. x402 is a payment method in its own right, designed for machines paying machines. They overlap, and the owners say they intend to interoperate: Visa says its protocol is meant to “complement” ACP and that it is working with Coinbase on x402 interoperability (Visa), and Coinbase describes x402 running inside Google’s AP2 (Coinbase).

What an agent can buy today, and where

This is the question readers ask first, and it is the one where official pages age fastest.

ChatGPT. When OpenAI announced Instant Checkout, it said US ChatGPT Plus, Pro and Free users could buy directly from US Etsy sellers in chat, with over a million Shopify merchants “coming soon,” and that it supported single-item purchases, with multi-item carts to follow (OpenAI). OpenAI also said merchants pay “a small fee on completed purchases,” that the service is free for users, and that Instant Checkout items are not preferred in product results. Check whether a buy button actually appears in ChatGPT before you count on it.

Google. Google says UCP lets businesses show offerings at “shopping touchpoints across consumer interfaces such as AI Mode in Google Search and Gemini app, and others in the future,” with the business remaining the merchant of record (Google Developers Blog).

Card networks. Visa and Mastercard describe their programs as infrastructure for merchants, issuers and AI platforms rather than consumer apps you sign up for. For what changes on your card statement and how to spot an agent purchase, see Visa Trusted Agent and Mastercard Agent Pay: what changes for cardholders.

Amazon. Amazon runs its own agent feature, covered separately in Amazon “Buy for Me”: how it works, what it can spend, how to cancel.

Machine-to-machine. The most active use of stablecoin rails so far is not people buying shoes. Eco reports that by late April 2026, x402 had 69,000 active agents, 165 million transactions and roughly $50 million in cumulative volume, an average of about 30 cents per call (Eco). Those are Eco’s figures, not audited ones, but the order of magnitude tells you the use case: tiny payments for data and API calls.

Cards or stablecoins: why both exist

It can look strange that two very different money systems are being wired into the same agents. The reason is that they solve different problems.

Cards already have what shoppers and stores rely on: fraud monitoring, disputes and chargebacks, and near-universal acceptance. The card networks are extending that rather than replacing it. Deloitte notes that “major providers have launched virtual cards for agents that contain built-in fraud protection, spending limits, and merchant category restrictions,” and that many of these still run on traditional rails (Deloitte). For a typical retail purchase with a human in the loop somewhere, cards fit well.

Stablecoins are digital tokens designed to track a currency, most often the US dollar. Their pitch for agents is about friction. Deloitte’s analysis says traditional payment systems “may obstruct agents with paywalls that require users to enter card credentials or complete human authentication,” while stablecoins “can be held in preauthorized wallets and transferred programmatically on 24/7 payment rails” (Deloitte). Coinbase makes the same argument about micropayments that card fees make impractical (Coinbase).

The tradeoff runs the other way too. Coinbase describes stablecoin settlement as “instant, final.” For a merchant, that means lower fraud risk. For a shopper, “final” means there is no chargeback button in the way card users are used to. Checkout.com, which processes both, expects stablecoins to be “complementing existing card and bank infrastructures” rather than replacing them (Checkout.com), and Deloitte expects a “hybrid model.” We explain stablecoins as a payment rail here, not as something to buy or hold.

If you want the per-request mechanics, read x402 in plain language: how agents pay per request in USDC.

Who is liable when it goes wrong?

Short answer: it is not settled, and the people building the protocols say so. Google’s AP2 announcement lists “accountability: determining accountability if a fraudulent or incorrect transaction occurs” as one of the core problems the protocol has to address, alongside proving the agent had authority and that the request reflects the user’s true intent (Google Cloud).

What the official pages do make clear is who the seller is. ACP’s repository says an AI platform can let users transact with businesses “without being the merchant of record” (ACP on GitHub), and Google says that under UCP “you remain the Merchant of Record” (Google Developers Blog). In practice that means the store, not ChatGPT or Gemini, is generally who you deal with for refunds and returns, much like buying from a third-party seller on a marketplace.

The harder question is whether “my agent chose it” counts as “I authorized it.” With a card, your existing dispute rights come from your card issuer and the network rules, and how they apply to agent purchases depends on the facts, the card type and the terms you accepted. With a stablecoin payment, there may be no dispute mechanism at all. This is exactly where consumer trust is thin: in a Checkout.com survey cited by Airwallex, a quarter of consumers said they would stop using an agent if a purchase turned out to be hard to dispute (Airwallex).

We cover buyer scenarios in detail in An AI agent bought the wrong thing. Who pays? Refunds, chargebacks and liability. This hub is not legal advice; for a disputed purchase of meaningful size, talk to your card issuer first and a consumer-protection attorney if that fails.

How big is this, really?

Numbers in this space are mostly forecasts and surveys from companies that sell agentic payment services. Here is where each figure comes from.

Read these with care. An $8 billion 2026 estimate and a $3–5 trillion 2030 forecast are hundreds of times apart, which tells you how uncertain the path is. Eco itself says none of the live deployments it reviewed “have replaced human checkout at meaningful volume yet.” The IMF is more measured than the vendors, framing the benefits as expected: lower costs, better liquidity, less fraud, alongside the need for oversight (IMF, 2026). Our reading: traffic and experiments are real and growing; agent-completed purchases are still small.

What it means for a shopper

For most people, the first contact will be an assistant offering to finish a purchase you were already researching. A few practical points follow from the sources above.

The agent is not the store. If the item is wrong, you will usually deal with the merchant, since both ACP and UCP keep the business as merchant of record. Keep the order confirmation the way you would for any online order.

Limits are your main safety tool. Airwallex’s definition of an agentic payment depends on “spending rules the buyer sets in advance,” and card programs advertise spending limits and merchant restrictions. Set them before you need them: see Spending limits for AI agents: how to set them in ChatGPT, Gemini and your card.

Pay with something that has dispute rights if the purchase matters to you. A card generally gives you a route to dispute; a stablecoin payment generally does not.

Be wary of anything calling itself a shopping agent that asks for your card number or a crypto transfer outside a known app. The legitimate protocols are built around tokens and approvals precisely so the agent does not need your raw card details. Our guide Fake AI shopping agents: how to tell a real one from a scam covers the warning signs.

What it means for a freelancer

If you sell services, digital products or data, the part of this that may reach you first is machine-to-machine payment. x402 was built so that “APIs, apps, and AI agents” can pay per request over the web (Coinbase). A freelancer who publishes a dataset, a paid newsletter archive or a small API could, in principle, charge an agent a few cents per access instead of selling subscriptions. The Eco figure of about 30 cents per x402 call shows that this is where the protocol is being used.

There are real costs to weigh. Getting paid in a stablecoin means holding a wallet, understanding how and when you convert to your local currency, and keeping records for tax. Settlement is final, so a mistake in your pricing or a bad actor on the other side is harder to unwind. None of this is advice to accept crypto; it is an option with tradeoffs, and a tax professional in your country is the right person to ask about reporting.

On the buying side, you are also a shopper. If you let an agent handle recurring purchases for your business (software, supplies, travel), the same rules apply: set limits, pay with a method that has dispute rights, and keep receipts. For multi-currency setups, our Money without borders guides go further.

What it means for a small shop

Small merchants are being courted by several protocols at once, and the official pages make a consistent promise: you keep control. OpenAI says sellers keep “full control of their payments, systems, and customer relationships” (OpenAI). Google says that with UCP “you own your business logic, and you remain the Merchant of Record,” with an embedded option for a customized checkout. Visa describes “no-code functionality for merchants to securely identify agents with an intent to buy” (Visa). Those are claims from the companies building the systems; how they play out depends on your platform.

For most small shops, the realistic path is through the platform you already use rather than implementing a protocol yourself. Google lists Shopify and Etsy among UCP’s collaborators, and OpenAI’s launch named Etsy sellers and Shopify merchants. That means your commerce platform’s settings, not a developer, are likely where agent sales get switched on. Our step-by-step guide is Accepting payments from AI agents as a small business, without code.

Budget for the fee. OpenAI says merchants pay a fee on completed Instant Checkout purchases. And prepare for disputes. If shoppers are uneasy about agent purchases, the store that makes returns easy will be the one they trust.

What to watch out for

• Stale pages. Launch posts stay online long after products change. Check the date and current status of any “buy in chat” feature before relying on it; our Instant Checkout note above is a live example.
• Vendor numbers presented as market facts. Most figures in this field come from companies that sell agentic payment services. Look for who ran the survey and who benefits.
• Protocol names as trust signals. A site saying it “supports ACP” or “uses x402” tells you nothing about whether it is honest. Scammers can borrow names.
• Final payments. Stablecoin settlement is described by its backers as final. That is good for merchants and risky for buyers who are used to chargebacks.
• Manipulated agents. An agent reading websites can be misled by content planted on those sites. Keep spending limits low and approvals on for anything you would not want bought by mistake.
• Account security. An agent with payment permission is only as safe as the account it runs in. Our 12-point account protection checklist is a good start.

Go deeper

All guides in this cluster live on the Agentic payments category page.

• Agentic commerce protocol tracker: ACP, UCP, AP2, TAP, Agent Pay, x402 — who supports what
• ACP vs UCP: OpenAI’s and Google’s checkout protocols compared
• Visa Trusted Agent and Mastercard Agent Pay: what changes for cardholders
• x402 in plain language: how agents pay per request in USDC
• An AI agent bought the wrong thing. Who pays? Refunds, chargebacks and liability
• Spending limits for AI agents: how to set them in ChatGPT, Gemini and your card
• Amazon “Buy for Me”: how it works, what it can spend, how to cancel
• Accepting payments from AI agents as a small business, without code