← Articles

Payment Infrastructure for AI Agents: The Complete Guide (2026)

Published July 29, 2026Updated July 29, 2026Guide

Key takeaways

  • Payment infrastructure for AI agents has four layers: verified agent identity, user intent and consent, payment credentials, and checkout execution.
  • A payment credential alone is not a purchase. A purchase is a confirmed order at a merchant, so execution is the layer that most often fails.
  • Consent should be a structured record of specific terms that the user approves before the purchase, and the system should enforce those terms during execution.
  • Industry efforts address different layers: Web Bot Auth and the Visa Trusted Agent Protocol for identity, ACP and UCP for merchant checkout, AP2 for authorization records, and Visa Intelligent Commerce and Mastercard Agent Pay for network credentials.
  • When agent purchases run on the user's own card with clear authorization, the user keeps the normal dispute rights of that card.

Payment infrastructure for AI agents is the layer that turns an agent's purchase decision into a completed, authorized transaction at a real merchant. It has four parts: proof that the agent is a known actor, a record of what the user approved, a chargeable credential the agent cannot read, and execution of the merchant checkout. This guide describes each layer, the industry protocols that touch it, and the questions to ask when you evaluate a provider.

Why do AI agents need their own payment infrastructure?

AI agents need their own payment infrastructure because the web's payment stack assumes a human at a keyboard. The checkout form, the fraud model, and the consent model all presume that a person reads the screen and clicks the button. An agent breaks each of those assumptions.

A checkout is a form built for a person. An agent can choose the correct product and still fail at address autocomplete, a guest or account choice, a one-time passcode sent to the user's phone, or a bot challenge that fires because the traffic does not look human.

  • Credentials: An agent should not hold a raw card number. If the agent can read the number, its logs, its prompt history, and anyone who compromises it can read it too.
  • Consent: A chat message that says "yes, buy it" is not an authorization record. It has no amount, no merchant, no line items, and no way to prove later what the user agreed to.
  • Identity: Merchants and CDNs cannot tell a legitimate shopping agent from a scraper or a card tester, so the safe default for them is to block it.
  • Execution: Even with a valid credential and a clear instruction, some system must complete the merchant's checkout and return an order confirmation.

What are the layers of agent payment infrastructure?

Agent payment infrastructure has four layers: identity, intent and consent, credentials, and execution. Each layer has a different failure mode, and many products in this market cover only one or two of them. When you evaluate a stack, check each layer separately.

Layer 1: Identity

The identity layer answers the merchant's first question: is this traffic legitimate? Merchants use bot detection because anonymous automated traffic is often hostile. An agent with no verifiable identity looks the same as an attacker, and bot detection treats it as one.

The emerging practice is cryptographic identity. In Cloudflare's Web Bot Auth proposal, an agent signs each HTTP request with HTTP Message Signatures (RFC 9421) and publishes its public keys in a directory, so a site can verify which agent sent the request. The Visa Trusted Agent Protocol, developed with Cloudflare, builds on the same signatures so merchants can recognize registered agents. A payment layer for agents should sign its requests in this way, so merchants do not have to guess from traffic patterns.

Layer 2: Intent and consent

The consent layer answers the question: what exactly did the user approve? A strong answer is a structured record, often called an intent object or a mandate, that names the merchant, the items, the quantities, the price ceiling, and the fulfillment details. The user approves it before anything executes.

The record works as a receipt that the user sees before the purchase. Its value comes from enforcement: the system should check the terms throughout execution, not once at the start. This is more specific than a card spend limit, which can see an amount and a merchant category but not a named merchant or an item.

Layer 3: Credentials

The credential layer answers the question: what gets charged, and who can see it? There are two common models.

  • The user's own card: The user adds an existing card, and a PCI DSS compliant vault replaces the number with a token. The agent never sees the card number. The user spends their own money, so no stored balance is involved.
  • An issued virtual card: A card issuer creates a card for one purchase or one program. This model suits funded programs, where the platform pays instead of the end user, and it brings the obligations of the issuing program with it.

Card networks also offer agent-specific credentials. Visa Intelligent Commerce and Mastercard Agent Pay use network tokens, which can carry a one-time cryptogram per transaction, so an agent-initiated payment is identifiable to the merchant and the issuer.

The difference that matters for product teams is the compliance burden. A model that holds stored value can raise money-transmission and customer identification questions. A model that charges the user's own card through a tokenized credential avoids most of them.

Layer 4: Execution

The execution layer answers the question: was the order actually placed? This layer is easy to underestimate. An authorization is not a purchase. A purchase is a confirmed order at a merchant, with a total the user recognizes and a fulfillment path that works.

Execution can use a merchant's structured interface where one exists, such as an ACP or UCP endpoint or an MCP tool. Where none exists, execution must complete the merchant's existing checkout. In both cases, the system should check each step against the approved terms, so a raised price, a substituted item, or an unexpected fee stops the purchase instead of going through.

How do the layers compare?

Each layer answers a different question, uses different standards, and fails in a different way.

LayerQuestion it answersRelevant standards and programsTypical failure without it
IdentityIs this agent a known, accountable actor?Web Bot Auth, HTTP Message Signatures (RFC 9421), Visa Trusted Agent ProtocolBot detection blocks the agent
Intent and consentWhat exactly did the user approve?Intent objects, AP2 mandates, passkeys (FIDO2 and WebAuthn)Disputes with no record of what was approved
CredentialsWhat gets charged, and who can see it?Tokenization, network tokens, Visa Intelligent Commerce, Mastercard Agent PayCard numbers exposed to the agent and its logs
ExecutionWas the order placed?ACP, UCP, MCP, or the merchant's existing checkoutThe agent has a card but no order is placed

How do users approve what an AI agent is about to buy?

Users approve an agent purchase through a scoped prompt that shows the merchant, the line items, and the total, and they confirm it with a biometric or a passkey. The pattern is familiar from Apple Pay, where a payment sheet shows the merchant and total and Face ID or Touch ID confirms it.

Two properties make the approval work. It should render inside the product the user already uses, so the user does not meet an unfamiliar brand at the moment of decision. And it should not redirect the user to a hosted payment page, because a redirect breaks the conversation and adds steps.

Who is liable when an AI agent buys something?

When an agent buys with the user's own card under a clear, recorded authorization, the purchase is a cardholder-authorized card transaction, and the user keeps the normal dispute and chargeback rights of that card. Card network rules and applicable regulation set the details, and they continue to change for agent-initiated payments.

This differs from a stored-value balance or a prepaid instrument, where the protections can differ from those of the user's own card. The authorization record also helps the other side of the transaction. Issuers and merchants often see thin context on an agent purchase. A record of the approved merchant, items, and ceiling gives them a better signal to tell legitimate purchases from fraud.

This page describes general practice. It is not legal or compliance advice. Liability for a specific transaction depends on network rules, regulation, and your agreements.

How is payment infrastructure different from a virtual card API?

A virtual card API gives an agent a chargeable number, which covers the credential layer and no other layer. For some uses, such as paying for an API subscription or a developer tool, that is enough.

It is not enough when the purchase involves a consumer merchant. A card number does not know what the user approved, cannot prove to a CDN that the agent is legitimate, and cannot complete a checkout form.

  • A card API: Issuance and spend controls. It suits machine-to-machine spend where there is no consumer checkout to complete.
  • A protocol: A shared format for agents and merchants to exchange offers and orders. It is necessary, but it executes nothing on its own and works only where the merchant has adopted it.
  • A PSP: Merchant-side acceptance. A payment service provider serves the merchant that receives money, not the agent platform that spends it.
  • Agentic commerce middleware: One layer that covers identity, consent, credentials, and execution together, and works whether or not the merchant has adopted a protocol.

How do the agentic commerce protocols fit in?

The agentic commerce protocols each cover one part of the stack, so they complement each other instead of competing.

  • MCP (Model Context Protocol): An open standard, introduced by Anthropic, that lets an agent discover and call tools on external servers. A merchant can expose search or ordering as MCP tools.
  • ACP (Agentic Commerce Protocol): An open specification from OpenAI and Stripe that defines how an agent reads a merchant's products and completes checkout.
  • UCP (Universal Commerce Protocol): An open specification introduced by Google with commerce partners including Shopify, with a similar merchant-side role.
  • AP2 (Agent Payments Protocol): An open protocol from Google that describes how an agent carries a user's authorization for a payment, with signed records called mandates.
  • Visa Intelligent Commerce and Mastercard Agent Pay: Card network programs for agent-initiated payments, based on network tokens and agent recognition.

Protocol adoption by merchants takes years. A payment layer for agents should use the protocol a merchant supports and still complete the purchase at merchants that support none.

How should you evaluate an agent payment provider?

You evaluate an agent payment provider by checking each of the four layers with evidence, not with a feature list. These five questions find the gaps quickly.

  1. 1

    Does it execute, or only issue? Ask for a live purchase at a consumer merchant, end to end, with an order confirmation from the merchant.

  2. 2

    What is the user consent model? Look for a structured, line-item authorization that the user approves before execution and that the system enforces during execution.

  3. 3

    Which network programs does it support? Ask which network programs, such as Visa Intelligent Commerce or Mastercard Agent Pay, are live and which are in progress.

  4. 4

    Does the user stay in your product? If the user sees a third-party brand or a redirect to a hosted page, conversion usually drops.

  5. 5

    Who carries the regulatory burden? Prepaid balances and stored value can bring money-transmission and customer identification obligations. A model that charges the user's own card avoids most of them.

Ask one more question if your agent buys from consumer merchants: what happens when the merchant blocks the request? A provider should have an answer for bot detection, one-time passcodes, and sign-in walls.

What does a full-stack agent look like once payments are solved?

Once payments are solved, an agent moves from advice to transaction. A shopping agent places the order instead of a list of links. A concierge agent books the table instead of naming three restaurants. A travel agent books the flight and the hotel instead of an itinerary.

The agent then has a completion event. That changes retention, because a user whose order arrived comes back, and it changes the business model, because a transaction can earn revenue in ways a recommendation cannot.

What are the steps to add payments to an AI agent?

To add payments to an AI agent, cover each of the four layers in order, from a tokenized credential to a reported outcome.

  1. 1

    Choose a credential model: Use the user's own card through a tokenized credential, so the agent never sees a card number. Add issued cards only if you fund purchases yourself.

  2. 2

    Define the approval record: Turn each request into structured terms: merchant, items, quantities, price ceiling, fulfillment, and expiry.

  3. 3

    Put the approval in your own UI: Confirm with a biometric or passkey, without a redirect.

  4. 4

    Give the agent a verifiable identity: Sign requests with HTTP Message Signatures so merchants and CDNs can recognize the agent.

  5. 5

    Execute against the approved terms: Use a merchant protocol where one exists, complete the checkout where none exists, and stop if the terms change.

  6. 6

    Report the outcome: Return the receipt, order status, and any failure to the user in the same conversation.

Agents can already decide what to buy. The missing part is the layer underneath: verified identity, enforceable consent, safe credentials, and execution that completes at the merchant. A team can build each layer itself or use infrastructure that provides them together.

Frequently asked questions

How do I let my users buy things through my AI agent?

To let users buy through an AI agent, you need four things: a tokenized payment credential the agent cannot read, a structured approval of each purchase that the user confirms, a verifiable agent identity for merchants, and a way to complete the merchant checkout. You can build these yourself or use payment infrastructure for agents that provides them together.

What is payment infrastructure for AI agents?

Payment infrastructure for AI agents is the layer that converts an agent's purchase decision into a completed transaction at a real merchant. It covers verified agent identity, a user-approved record of the purchase terms, a tokenized credential the agent never sees, and execution of the merchant checkout.

Can an AI agent use my real credit card?

Yes, an AI agent can use your real credit card through a token. A PCI DSS compliant vault replaces the card number with a token, so the agent never sees the number. You approve each purchase with a scoped prompt and a biometric or passkey, and you keep the normal dispute rights of your card.

What are Visa Intelligent Commerce and Mastercard Agent Pay?

Visa Intelligent Commerce and Mastercard Agent Pay are card network programs for payments that AI agents initiate. Both use network tokens, which can carry a one-time cryptogram for each transaction, so merchants and issuers can recognize an agent-initiated payment.

Why is issuing a virtual card not enough for agent payments?

A virtual card covers the credential layer only. It does not know what the user approved, cannot prove to a merchant or CDN that the agent is legitimate, and cannot complete a merchant checkout. Many agent purchases fail after the card exists, at the checkout page.

Do users need to pass KYC to let an agent spend for them?

Usually not, when the agent charges the user's own card. The user spends their own money on a card their bank already issued, and no stored balance is involved. Models that hold user funds or issue new cards can bring customer identification obligations.

How do you keep an agent from spending more than the user approved?

You keep an agent within the approved amount by recording the purchase as structured terms (merchant, items, quantities, and a price ceiling) and checking those terms throughout execution. If the merchant changes the price or substitutes an item, execution stops and the user gets a new approval.

How do the ACP, UCP, and AP2 protocols relate to each other?

ACP, from OpenAI and Stripe, and UCP, introduced by Google with partners including Shopify, are merchant-side protocols for catalog and checkout. AP2, from Google, describes how an agent carries the user's authorization for a payment. They cover different parts of a purchase and can be used together.

Related articles