Agentic Payments Explained: When an AI Agent Pays for You

The short answer: agentic payments are purchases made by an AI agent on your behalf — an assistant that finds a product, builds a basket and pays, within limits you set. The card networks and big tech are building the plumbing now: Visa announced Visa Intelligent Commerce in April 2025, using tokenised card credentials and spending limits you control, and Google published the Agent Payments Protocol (AP2) in September 2025, an open standard that records what you asked for and what you approved as signed "mandates". For you, the practical questions are simple: what can the agent spend, how do you prove what you authorised, and who fixes it when it buys the wrong thing.

"Agentic payments" has become one of the most searched phrases in payments this year, mostly by people trying to work out whether it is a real change or a buzzword. It is real in the sense that the largest payment companies are shipping standards for it. It is early in the sense that the consumer protections, dispute habits and bank controls around it are still being worked out.

What an agentic payment actually is

Every card payment today assumes that a human is the one clicking "buy". An agentic payment breaks that assumption. You tell an AI assistant what you want — "book the cheapest refundable train to Manchester on Friday morning", "reorder washing powder when it is under £10" — and the agent completes the purchase.

Google's AP2 announcement sets out why that needs new plumbing. It names three questions current payment systems cannot answer when an agent is in the middle: authorisation (proving you gave the agent authority for this purchase), authenticity (letting a merchant confirm the request reflects your real intent) and accountability (working out who is responsible if the transaction is fraudulent or wrong).

How the main approaches work

Visa Intelligent Commerce. Visa's April 30, 2025 announcement opened its network to developers building agents that can find and buy. It relies on tokenised digital credentials — a stand-in for your real card number — and lets consumers set spending limits and conditions for agent transactions. Visa's own summary: "Each consumer sets the limits, and Visa helps manage the rest." The release names Anthropic, IBM, Microsoft, Mistral AI, OpenAI, Perplexity, Samsung and Stripe among its collaborators.

Google's Agent Payments Protocol (AP2). Announced on 16 September 2025, AP2 is an open protocol designed to work as an extension of Google's Agent2Agent protocol and alongside the Model Context Protocol. Google says it is payment-agnostic, supporting "credit and debit cards to stablecoins and real-time bank transfers", and that more than 60 organisations are collaborating on it, including Adyen, American Express, Coinbase, Mastercard, PayPal, Revolut and Worldpay.

The core idea in AP2 is the mandate — a cryptographically signed record of your instructions:

• Intent Mandate: captures what you asked for and the rules the agent must follow, such as price limits and timing
• Cart Mandate: created when a specific basket is approved, recording the exact items and price, so "what you see is what you pay for"

Google describes the chain from intent to cart to payment as a "non-repudiable audit trail". In plain terms: if there is a dispute, there is a signed record of what you asked for and what was approved.

What changes for you as a customer

Your card number travels less. Network tokens mean the agent and merchant handle a substitute credential rather than your real card number, which limits the damage if one of them is breached.

Limits move to the front. Instead of a single credit limit, you set rules for the agent: a maximum per purchase, a list of merchants or categories, a time window. Those controls are the whole safety model, so they are worth setting deliberately rather than accepting the defaults.

Approval may happen before you see the payment. In delegated mode you approve the rules, not each payment. That is efficient for routine purchases and risky for anything expensive or non-refundable.

Disputes depend on evidence. Signed mandates should make it clearer what you did and did not authorise. How card issuers, merchants and regulators treat an agent's mistake — a wrong size, a non-refundable fare, a duplicate order — is still being settled in practice. Keep the agent's confirmations, just as you would keep an email receipt.

A worked example

Say you ask an assistant to keep your household stocked and allow it up to £60 a month on cleaning products from two named retailers. Under an AP2-style flow you would sign an Intent Mandate once, recording that budget, those merchants and the product types. Each time stock runs low, the agent builds a basket, a Cart Mandate records the exact items and price, and the payment runs on a tokenised card credential rather than your card number. If one month the agent orders a twelve-pack instead of a two-pack, the signed records show what your rules allowed and what was actually bought — which is the evidence you, the merchant and your bank would need to sort it out. If the agent tries to spend £80, a properly configured limit stops it before the money moves. The value of the system is less in the clever shopping than in that paper trail and those hard limits.

A safe set-up for agent payments

• Use a dedicated card or virtual card for agents. A separate virtual card with its own limit means a misbehaving agent can only spend what that card allows, and you can freeze it without touching your main account. Our virtual card set-up guide walks through the options
• Set per-transaction and monthly caps in your banking app and, where offered, in the agent itself
• Keep high-value and non-refundable purchases human-approved. Delegate the washing powder, not the flights
• Turn on instant notifications for every card the agent can use
• Review what the agent can see. An agent with access to your email and calendar knows a great deal; what your AI bank knows about you covers the data side
• Know your dispute route before you need it: how AI fraud detection works in banking explains why a legitimate agent purchase can still be flagged

For businesses: agents as customers, and as staff

Two shifts matter for companies. First, some of your customers' purchases will arrive via agents, so checkout flows that rely on a human reading the page will need structured product data and clear refund terms an agent can parse. Second, companies are starting to give agents spending power internally — booking travel, paying SaaS invoices. The control model is the same one finance teams already use for staff: issue each agent its own virtual card with merchant restrictions and a hard limit, and reconcile the spend like any other card programme. Our guide to the best virtual card providers compares issuers that support this.

What is still unresolved

The standards are new and adoption is uneven. Not every bank exposes the controls an agent needs; not every merchant will accept agent-initiated payments; and the question of liability when an agent misreads an instruction has not been tested at scale. The AI part also brings its own failure modes, which we cover in when bank AI gets it wrong. Treat agentic payments today the way people treated online banking in its first years: useful for routine tasks, with tight limits and a human check on anything that matters.

For the wider picture of how AI already shapes your bank account, read your bank is now an algorithm and how AI banking works.

This is general information, not financial advice. Features, partners and protections described here are taken from the companies' own announcements and may change; check your bank's and the agent provider's current terms before letting software spend your money.