Explained to kids!
Agentic commerce is when a computer helper, called an AI agent, does the shopping for you: it finds the thing, compares prices and even pays. But it can only pay with your permission, for what you said, up to the amount you allowed. The shop needs to check that the helper is real, and that you really asked for it.
Let’s see it in real life…
Nils lives with his mom, Sara, on the fourth floor. Jonas, the teenager from the second floor, runs errands for the whole building on Saturdays.
Mom is working this morning, so she writes a note: “Bread, milk, apples. If strawberries cost less than 3 euros, one box. No sweets. Not more than 15 euros. Bring the receipt.” She signs it and puts 15 euros in an envelope.
Nils adds in big letters: “THE GREEN APPLES.”
At the grocery, Mr. Amadi knows Jonas. He still reads the note. “Signed by Sara. Fine.”
Jonas picks the bread, the milk, the green apples. The strawberries cost 3.50. He leaves them.
At the counter, Jonas reaches for a chocolate bar. Mr. Amadi taps the note: “No sweets. Not on this list.”
Jonas laughs and puts it back. The total is 11.20 euros. He pays from the envelope and keeps the receipt.
Back upstairs, Nils checks the bag against the note. Everything matches. Mom checks the receipt: 11.20, change 3.80. Correct.
“Why did Mr. Amadi read the note if he knows Jonas?” asks Nils.
“Because knowing Jonas is not the same as knowing what I asked for. The note is the proof. Without it, Jonas is just a boy with my money.”
That is agentic commerce.
Tiny recap
- It means… a helper shops and pays for you, with a note that says exactly what it may buy and how much it may spend.
- It helps when… you have no time, or when the helper can compare prices faster than you.
- Watch out for… helpers without a note. A shop should check who the helper is and what it was allowed to do, before taking the money.
…detailed for specialists
Agentic commerce is commerce in which an AI agent, acting on a person’s instructions, searches, selects, negotiates and pays for goods or services, with the person present or not at the moment of payment. The payment runs on existing rails, cards, wallets, bank transfers or stablecoins, but with new elements: proof that the agent is who it claims to be, and proof of what the person authorized it to do.
It matters in 2026 because the schemes, the big PSPs and the AI platforms shipped their frameworks between April 2025 and mid-2026, because live agent-executed transactions have cleared in the US and Europe, and because none of the fraud, authentication and dispute rules PSPs and PayFacs run today were written for a buyer that is a program.
What is agentic commerce?
In traditional e-commerce a human browses, fills a cart and clicks pay. In agentic commerce the human states an intent (“find white running shoes under $120 and buy them when in stock”), and an agent, in a chat assistant, a browser or a merchant app, does the rest. The agent needs three things today’s checkout does not provide: a way for the merchant to recognize it as a legitimate agent rather than a bot, a credential to pay without holding the person’s card, and a record of what it was allowed to buy.
That record is the mandate, the signed note in the kids’ story. Google’s Agent Payments Protocol defines an Intent Mandate, which captures the user’s request and its limits, and a Cart Mandate, signed when a specific cart is approved, both cryptographically signed with verifiable credentials. The credential is usually a scoped token: a payment token restricted to one merchant, one amount or one agent, derived from the same network tokens that secure wallets and card-on-file today. Agentic commerce is the consumer face of what we called the Internet of Payment: objects and programs that pay, with humans setting the rules.
Who is building the rails?
- Visa Intelligent Commerce launched in 2025. Its Trusted Agent Protocol, introduced in October 2025, helps merchants tell malicious bots apart from legitimate AI agents, and by December 2025 Visa had completed hundreds of agent-initiated transactions with more than 100 partners. In July 2026, 31 European issuers completed live agent-executed transactions, authenticated with Visa Payment Passkeys to meet strong customer authentication.
- Mastercard Agent Pay, announced in April 2025, introduced Agentic Tokens, which require agents to be registered and verified, let users define what an agent may buy, and make agent purchases identifiable to consumers, merchants and issuers. Mastercard said all US cardholders would be enabled for Agent Pay by the 2025 holiday season, with Citi and U.S. Bank first.
- Stripe and OpenAI’s Agentic Commerce Protocol (ACP) powers Instant Checkout in ChatGPT. Stripe issues a Shared Payment Token scoped to a specific merchant and cart total; ChatGPT passes it to the merchant, who remains responsible for accepting the order, charging, tax and fulfillment. Etsy sellers went live on September 29, 2025, with over a million Shopify merchants to follow.
- Google’s AP2, released in September 2025 with more than 60 organizations including Adyen, American Express, Mastercard, PayPal, Revolut and Worldpay, is payment-method agnostic: cards, stablecoins and real-time bank transfers, with an x402 extension for crypto.
- PayPal adopted ACP in October 2025 to put its wallet, with buyer and seller protections and dispute resolution, inside ChatGPT, routing its merchant network through one ACP server.
Stripe, Adyen, Worldpay and PayPal are all connectors a platform can already run; what changes is the data each transaction carries.
How does an agentic payment work?
- Intent and consent. The user gives the agent a task and limits, recorded as a signed intent with price ceiling, timing and merchant scope.
- Agent identification. The agent presents a verifiable identity to the merchant, through Visa’s Trusted Agent Protocol, Mastercard’s agent registration or an AP2 credential. The merchant decides whether to serve it.
- Cart and approval. The agent assembles a cart. With the human present, the user approves it, producing a cart mandate. With the human absent, the pre-signed intent authorizes the cart if it meets the conditions.
- Credential issuance. A scoped token is created for that merchant and amount: an Agentic Token, a Shared Payment Token or a network token with agent indicators. The agent never holds the PAN.
- Authorization. The merchant’s PSP sends the authorization with the token, the agent identifier and, where supported, the intent reference. In Europe, strong customer authentication still applies: Visa’s pilots used passkeys bound to the user, which fits the delegated and exemption logic we cover in 3-D Secure and SCA exemptions.
- Fulfillment, receipt and evidence. The merchant fulfills, the receipt goes back to the user, and the mandate, the token and the authorization are stored together for the dispute that may come.
Agentic commerce in numbers (2026)
- Traffic: Adobe measured generative-AI referrals to US retail sites up 693% year over year during the 2025 holiday season, on $257.8 billion of online spend.
- Usage: Visa’s October 2025 survey found 47% of US shoppers use AI tools for at least one shopping task, and Visa predicts millions of consumers will let agents complete purchases by the 2026 holiday season.
- Purchases: Worldpay’s October 2026 report finds that 31% of consumers who shopped with AI completed a purchase or payment through it, and that consumers expect agents to handle 11–13% of their purchases today and 17% in five years.
- Trust: in a Harris Poll for Visa in May 2026, 23% of US consumers trust generative AI to handle payments on their behalf, while 61% would trust Visa to handle agentic transactions.
- Reach: PayPal brings over 400 million users and tens of millions of merchants to ChatGPT checkout, with merchant catalogs due in 2026.
Who is liable when an agent buys the wrong thing?
Nobody has settled it. The mandate frameworks are designed to answer the question with evidence: a signed intent proves what the user asked, a cart mandate proves what was shown, the token proves which agent paid. But as of October 2026, Visa and Mastercard publish no dispute reason code specific to agent purchases, as we noted in AI in payments: agentic commerce and the acceptance problem. A “my agent bought this without my consent” claim will arrive as fraud or as a cardholder dispute, and the merchant will defend it with the mandate, if it kept one. The same tools that already fight bots and account takeover, described in AI vs fraud in payments, now have to distinguish a verified agent from a scripted one, and an authorized agent from an authorized agent exceeding its mandate.
Two other questions are open. Agent-to-agent payments, where a program buys an API call from another program, are too small and frequent for card economics; AP2’s x402 extension and the stablecoin rails behind it are the current answer, and AP2 also names real-time bank transfers, which puts instant payment systems among the candidate rails. And in Europe, strong customer authentication applies to the human, not the agent, which makes a passkey-bound mandate the only compliant path for now.
Agentic commerce: what changes for PSPs, PayFacs and platforms
Onboarding and merchant readiness
Merchants need machine-readable catalogs, an ACP or AP2 endpoint, and a policy on which agents they serve. Platforms will be asked to provide this per merchant, as they provided hosted checkout pages. Agents become a new counterparty to onboard: who registered it, with which scheme, under which terms.
Authorization and token management
Agent transactions arrive with new fields: agent identifier, token type, mandate or intent reference, human-present or not. Authorization messages must carry them, retries must keep them, and issuers will use them to approve or decline. Scoped tokens are network tokens with constraints, which is why Network Tokenization becomes a prerequisite rather than an optimization.
Routing and retries
An agent retries a declined purchase faster and more often than a person. Agent-initiated payments need their own retry limits, inside the schemes’ reattempt rules, and their own acquirer preferences where one acquirer supports agent indicators and another does not.
Events and evidence
A mandate nobody can find in six months is useless. Each agent payment should raise events the operator’s systems can store and replay: intent received, cart approved, token issued, authorized, captured, disputed. Signed webhooks and a queryable API, the subject of Developer & Event infrastructure, are how a platform keeps the proof attached to the payment across several providers.
Disputes, pricing and reporting
Dispute teams need the mandate and the agent identity on the transaction record, not in a separate tool. Pricing teams must decide whether agent traffic is priced like card-not-present e-commerce or as a new channel with its own fraud profile. Reporting needs an agent dimension, so a merchant sees what share of its sales, declines and disputes came through agents, which is where Operations & Data comes in. PSPs whose stack predates tokens and event streams will find this the moment to upgrade a legacy platform.
NORBr provides the white-label payment infrastructure to route, onboard, pay out and analyze, under the operator’s own brand; it does not issue agentic tokens or verify agents, it keeps what the schemes and agent providers return attached to the right payment.
Is agentic commerce worth preparing for in 2026?
Yes, and the preparation is less about AI than about data discipline. The protocols, the tokens, the first banks and merchants are live; consumer trust is the limiting factor: a quarter of US consumers trust an AI to pay for them, two thirds trust their card network to supervise it. The platforms that win the next two years are those whose transaction records carry an agent identity and a mandate reference end to end, and whose merchants can prove, when a dispute lands, what the person asked for. Arthur Bedel explains the stakes in our #TheShot episode agentic commerce: when an AI agent pays, and the Payments Library report Agentic Commerce and Payments 2026 goes further. For the market where it is live first, see our guide to payment methods in the United States.
Frequently Asked Questions
What is agentic commerce?
Agentic commerce is shopping done by an AI agent on a person’s instructions: the agent searches, compares, selects and pays, within limits the person set. It runs on existing payment rails, with two additions: proof that the agent is legitimate, and a signed mandate proving what the person authorized it to buy.
How does an AI agent pay without my card number?
With a scoped token. The card network or PSP issues a token limited to one merchant, one amount or one agent, such as Mastercard’s Agentic Tokens or Stripe’s Shared Payment Tokens, built on network tokenization. The agent presents the token; the real card number never leaves the issuer and the token vault.
Who is liable if an AI agent makes a wrong purchase?
It is not settled. As of late 2026, Visa and Mastercard have no dispute reason code specific to agent purchases, so claims arrive as fraud or as a chargeback. The protocols answer with evidence: an intent mandate, a cart mandate and the agent’s identity. Merchants and platforms that keep them can defend the transaction.
What is the difference between Visa Trusted Agent Protocol, Mastercard Agent Pay and Google AP2?
Visa’s Trusted Agent Protocol lets merchants recognize verified agents; Mastercard Agent Pay issues Agentic Tokens to registered agents; Google’s AP2 is an open, payment-method-agnostic protocol built on signed mandates. Stripe and OpenAI’s ACP connects merchants to ChatGPT checkout. They overlap and are designed to coexist on the same transaction.
Does strong customer authentication apply to AI agents in Europe?
Yes, to the human behind the agent. PSD2 requires strong customer authentication of the payer, so Visa’s European pilots authenticated agent transactions with payment passkeys bound to the user. Exemptions and merchant-initiated logic, and the PSD3 and PSR rules that follow, will shape how unattended agent purchases are approved.
