A glowing authorization token passing from a human hand to a network of machine nodes, in a dark technical illustration.

What Is an Agentic Payment?

Last Updated: November 10, 2026 • Autonomous Agentic Commerce Spec v2026-08-25

Production Validated

What Is an Agentic Payment?

An agentic payment is a payment that an AI agent initiates and completes on behalf of a person or business, inside limits that person approved in advance, where the payment carries verifiable evidence of that approval. The agent does not borrow the person’s card number. It acts under a mandate — a recorded statement of what it may buy, from whom, and up to how much — and it pays with a credential scoped to that mandate. Every party downstream (the merchant, the payment network, the issuer) can check the payment against the approval instead of taking the agent’s word for it.

That is the whole definition. Everything else in this piece is unpacking it.

The anatomy of an agentic payment

Payment Model Initiating Entity Authorization Mechanism Contract Formation Moment Human In-The-Loop Primary Protocol Standard
Traditional Ecommerce Human Shopper. Manual CVV / 3DS SMS OTP input. Clicking ‘Place Order’ on merchant website. Mandatory at checkout moment. Card-Not-Present ISO 8583.
Recurring Subscription Merchant Billing Engine. Stored card token from initial purchase. Initial terms of service agreement. None (automated batch schedule). Card-on-File (CoF) Gateway Billing.
Scheduled Automatic Bill Pay Bank Account Service. Pre-authorized ACH debit agreement. Signed bank authorization mandate. None (calendar trigger). NACHA ACH / SEPA Direct Debit.
Agentic Payment Autonomous AI Agent. Cryptographic FIDO AP2 delegation mandate. Algorithmic acceptance of verified quote. Delegated pre-authorization with budget caps. Universal Commerce Protocol (UCP).
Machine Micropayments Autonomous Cloud Workload. Ephemeral API wallet signature. Programmatic HTTP status negotiation. Zero human presence. Layer-2 Lightning / Stripe Projects.

Six parts have to be present. Remove one and you have something else — usually a demo.

  • A principal. The person or business whose money it is and whose instruction starts the chain. Agentic payments are delegated spending, not independent spending: there is always a principal.
  • An agent. The software that finds the item, agrees the price and executes the payment. The agent works inside the delegation; it does not hold funds of its own.
  • A mandate. The approved limits, recorded before the payment happens: what may be bought, from which merchants or categories, the maximum amount, and an expiry or condition. In Google’s Agent Payments Protocol (AP2) the mandate is a cryptographically signed object — an Intent Mandate for the instruction, a Cart Mandate for the exact approved basket. Other systems record the same facts in their own form.
  • Authorization evidence. The proof that travels with the payment: the signed mandate, or a token that could only have been issued under one. This is the piece that replaces the signals card systems used to rely on — a person, a device, a typed card number, a clicked Buy button.
  • A scoped credential. The agent pays with a token bound to the mandate — typically restricted to a merchant, an amount or a short validity window — rather than the underlying card number. Stripe’s Shared Payment Tokens in the Agentic Commerce Protocol and Mastercard’s Agentic Tokens are the current card-world examples.
  • Settlement. The money actually moves, on ordinary rails: a card network, a bank transfer, a stablecoin rail. Execution of the payment and settlement of the payment are different events, and the definition only requires the first to be agentic.

The layer map

Most confusion in this market comes from comparing systems that sit on different layers of the same stack. A working map, from the buyer down:

  • Agent layer — the assistant or purchasing agent that receives the instruction and does the shopping.
  • Permission layer — where the mandate lives: what was approved, by whom, with what limits. AP2 works here.
  • Checkout layer — where agent and merchant agree the order: cart, price, tax, shipping, and the terms of the transaction. This is where the Universal Commerce Protocol sits, alongside the Agentic Commerce Protocol (ACP). UCP is a checkout protocol: it standardizes how an agent buys from a merchant, with the merchant remaining seller of record. Its full treatment is in the definitive guide to the Universal Commerce Protocol.
  • Credential and network layer — the scoped token is presented and the transaction is authorized over payment networks and rails: Visa, Mastercard and their agent programs, or stablecoin protocols such as x402 for machine-to-machine payments.
  • Settlement — funds move between issuer, acquirer and merchant, on the same rails card payments have always used.

A single purchase can touch every layer: an AP2 mandate records the approval, UCP runs the checkout with the merchant, a network token carries the credential, and the merchant settles on ordinary card rails. None of the pieces replaces the others.

Where UCP sits — and where it does not

UCP is in the checkout layer, and the distinction matters for this definition. UCP is not a payment network: it does not authorize transactions against accounts, it does not move money, and it issues no credentials of its own. When an agent checks out through UCP, the payment still rides a payment method — a wallet or card credential — over a network that settles it.

What the checkout layer contributes to an agentic payment is the agreement: a machine-readable order that both sides can verify against the mandate before money moves. The consent side of that agreement — when an agent must stop and ask — is covered in Consent Management in Agentic Commerce and UCP AI Consent: When Agents Must Ask Permission. The credential’s own lifecycle, from issuance to expiry, is covered in UCP Token Lifecycle.

Not this: three things an agentic payment is not

It is not autopay. A subscription or a stored recurring charge executes a fixed instruction the human wrote once: this amount, this merchant, this date. No decision is delegated, because no decision remains. In an agentic payment the human approves limits and the agent makes the decision the instruction left open — which item, at which price, from which seller, and whether to buy at all. The decision is the difference.

It is not one-click checkout. One-click removes typing, not the human. The person is present, sees the order and confirms it. Some agentic checkouts also ask for a confirmation tap — Google’s and OpenAI’s both launched that way — but the category does not require the human to be present at the moment of payment. A mandate signed in advance can authorize a purchase that completes while the principal is asleep, which is exactly the case one-click cannot express.

It is not a human paying with a saved card through a chatbot. If a person pastes or autofills their own card details into a chat window and presses pay, the payment is an ordinary card payment with an unusual front end. Nothing has been delegated, nothing is scoped, and no authorization evidence exists beyond the card transaction itself.

The neighbouring terms

The vocabulary is unsettled, and the honest move is to map it rather than crown a winner:

  • Agentic commerce is the whole act: discovery, comparison, negotiation, checkout and payment, performed by an agent. An agentic payment is the payment leg of that act — the moment value moves. Every agentic payment inside a shopping flow is part of agentic commerce; not every part of agentic commerce is a payment.
  • Agentic checkout is the shopper-visible end of the checkout layer: the buy surface inside a chat or AI search result.
  • Agent payments / AI agent payments are loose synonyms for agentic payment, used interchangeably by most writers. Autonomous payment usually means the same thing with the emphasis on the human-not-present case.
  • Machine payments is the neighbouring market rather than a synonym: agents paying other software per API call, per crawl or per second of compute, typically in fractions of a cent on stablecoin rails. The authorization problem is the same; the ticket size and the rails are not.

What makes one real instead of a demo

A working agentic payment, as opposed to a staged one, has four properties. They are properties of the transaction, not of any one vendor’s stack:

  • The mandate is verifiable by parties who were not in the room. The merchant and the issuer can check the authorization evidence themselves, after the fact. A screenshot of a chat is not evidence; a signed mandate or a token that only exists because a mandate was approved is.
  • The delegation is revocable. The principal can withdraw or narrow the mandate, and the revocation takes effect on payments not yet executed — the same property agent compliance auditing depends on when records have to be produced later.
  • Replays fail. A captured credential or a repeated instruction cannot spend twice. At the checkout layer this is ordinary engineering — idempotency keys on agent transactions — but without it a scoped token is just a card number with extra steps.
  • Liability has an address. Authorization, authenticity, accountability: Google framed these as the three questions an agent payment system must answer when AP2 was announced in September 2025. When an agent buys the wrong thing inside its mandate, someone specific answers for the loss. Who that is — principal, agent platform, merchant, issuer — is being settled program by program, and any definition that pretends it is already settled everywhere is ahead of the facts.

The short version

An agentic payment is a payment an AI agent initiates and completes on behalf of a person, inside limits that person approved, carrying proof of that approval. Autopay executes an instruction; an agentic payment exercises a delegation. UCP’s place in it is the checkout — the layer where the order is agreed — while the mandate sits above it and the networks and rails sit below.

Frequently Asked Questions

What is the formal legal and technical definition of an ‘Agentic Payment’?

An Agentic Payment is an autonomous financial transaction initiated and completed by an artificial intelligence agent on behalf of a human principal, utilizing cryptographic proof of delegation without requiring human form-filling at execution.

How is a legally binding contract formed when neither buyer nor seller is a human?

Under commercial contract law, the agent operates under a valid legal power-of-attorney grant created by the human principal’s cryptographic delegation token, forming a binding contract upon algorithmic acceptance of the merchant’s signed quote.

How does an agentic payment differ from a traditional saved-card recurring charge?

A recurring charge repeats fixed amounts on a set calendar date, whereas an agentic payment dynamically determines what, when, and from whom to buy based on changing real-world conditions within authorized spending limits.


by

Tags:

Comments

Leave a Reply

Your email address will not be published. Required fields are marked *