AI agents pay: How agents are changing digital payments

How AI Agents Are Changing Digital Payments and Online Transactions

Rate this post

Tell an assistant to reorder the printer toner when it runs low, within $80, from any approved supplier. A few days later the toner arrives. You never opened a checkout page, never typed a card number, and never saw the confirmation email until it was already done.

Payments were designed on one assumption: a human is clicking the buy button. Fraud systems, card rules, checkout pages and dispute processes all lean on it. AI agents break that assumption, and the industry has spent the last eighteen months writing new rules for software that spends money on someone’s behalf.

AI agents in payments are software agents that can discover products, compare options and complete purchases for a person or a business within limits that person has set in advance. This article covers how the new payment standards work, who has launched what, and where the unresolved problems sit.

Quick answer

AI agents in payments are autonomous software that shop and pay on a user’s behalf within preset limits. Card networks and payment companies, including Visa, Mastercard, Stripe, Google and OpenAI, have built protocols that give agents tokenised credentials, verifiable authorisation and identity checks. Most remain in pilot or early rollout, and liability rules are still unsettled.

How does an AI agent make a payment?

An AI agent makes a payment by presenting a credential the user has authorised for that purpose, usually a tokenised card or a single-use token, together with proof that the user approved the purchase and its limits. The merchant and the payment network check both before accepting the transaction.

The key concept is a mandate: a record of what the user allowed the agent to do, such as spend up to a set amount, on certain merchant categories, within a time window. Google’s Agent Payments Protocol (AP2) formalises this with signed mandates that tie a checkout to the user’s stated intent. Mastercard’s Verifiable Intent plays a similar role on its network.

Tokenisation is the second piece. Instead of handing a real card number to an agent, the network issues a restricted token that works only for the agent, under the agreed conditions. If the token leaks, the damage is bounded by the limits attached to it.

Who has launched agentic payment systems?

Visa, Mastercard, Stripe, Google, OpenAI and Coinbase have each launched protocols or programmes for agent payments since 2025. The efforts overlap more than they compete, and several are designed to work together.

Initiative Backer What it does
Visa Intelligent Commerce Visa Tokenised credentials and authorisation for agents
Mastercard Agent Pay Mastercard Agent registration and agentic tokens; Verifiable Intent added in March 2026
Agentic Commerce Protocol (ACP) OpenAI and Stripe Checkout between agents and merchants, announced September 2025
Agent Payments Protocol (AP2) Google, now stewarded by FIDO Alliance Signed mandates for agent purchases
x402 Coinbase Machine-to-machine payments over HTTP
Machine Payments Protocol Stripe with Tempo Open standard for programmatic payments, March 2026

Visa and OpenAI announced an integration of Visa Intelligent Commerce into OpenAI experiences on June 10, 2026, with user-defined controls such as spending caps, merchant categories and human approval. Stripe launched its Agentic Commerce Suite in December 2025, letting merchants make products discoverable to agents through its existing integration.

Are agentic payments live or still in pilot?

Agentic payments are in early rollout, not mainstream use. Card networks describe deployments as pilots or phased launches, and one tracker reported in late October 2026 that only Stripe’s suite was self-serve for merchants, on preview terms. Expect fast change, and treat any claim of full availability with caution.

Mastercard stated in March 2026 that its Agent Pay pilot was not a commercial rollout at that stage, and Visa described its programme as in the process of deployment. Individual banks have completed live end-to-end agent payments, but those are demonstrations of feasibility, not proof of scale.

Forecasts of agentic commerce volumes, often quoted in the trillions of dollars by 2030, are projections by consultants and vendors, and they depend on assumptions about consumer trust that remain unproven. A cautious reading of the evidence is that infrastructure is arriving faster than demand.

How do you tell a trusted AI agent from a malicious bot?

Merchants and networks are building identity and verification layers, sometimes called Know Your Agent, so that a legitimate agent can prove who it is and who authorised it. Without this, fraud systems built to block bots will also block good agents.

This is a real problem for merchants. For years, fraud tools treated automated traffic as a threat. Now some automated traffic is a paying customer’s assistant, and the signals that separate it from an attacker are new. Mastercard and Visa have partnered with FIS so that issuing banks can use agent authentication from both networks.

Expect merchants to adopt allow-lists of registered agents and to demand signed proof of the user’s mandate. Smaller merchants will depend on their payment provider to do this for them, which is another reason the processor’s roadmap matters when choosing one.

Who is liable when an AI agent makes a bad purchase?

Liability for agent purchases is not settled. Existing dispute rules assume a human cardholder either authorised a transaction or did not, and an agent acting within a mandate sits awkwardly between those categories. Networks and regulators are still working out who bears the loss when an agent misreads an instruction.

Consider the cases. If the agent followed the mandate precisely but the user’s instruction was sloppy, the user probably bears the loss. If the agent exceeded its mandate, the blame shifts toward the agent’s developer. If a third party hijacked the agent through hidden instructions in a webpage, the picture is murkier still.

In the ACP model, OpenAI states that it is not the merchant of record, and Google says its brand partners remain the merchant of record in its programme. That keeps refunds and disputes with the merchant, a familiar structure, but it does not answer the agent-error question.

What should merchants and payment companies do now?

  • Ask your payment provider which agent protocols it supports, and on what timetable.
  • Make product data machine-readable: accurate prices, stock, delivery times and return policies in structured form.
  • Review fraud rules so that registered, mandated agents are not blocked outright.
  • Update terms of sale and refund policies to address orders placed by agents.
  • Log the agent’s identity and mandate evidence with each order, for use in any dispute.
  • Start with a narrow pilot, such as repeat purchases of consumables, where errors are cheap.

For payment companies, the strategic question is whether to build for one network’s standard or for interoperability. The protocols above are converging, and aligning with open, multi-network standards looks lower risk than betting on one.

What risks do AI agents in payments create?

The main risks are agent error, hijacked instructions, over-broad permissions and a lack of clear recourse. These risks are manageable with tight limits, but they grow with each added permission.

  • Misread intent: an agent buys the wrong item or the wrong quantity.
  • Prompt injection: hidden instructions in content the agent reads redirect its spending.
  • Excess authority: wide limits on a credential turn a small error into a large loss.
  • Weak recourse: if the dispute process does not recognise agent purchases, the cardholder may be left out of pocket.

Keep limits low at first. A weekly cap of a modest amount, restricted to named merchants, gives you the convenience with little exposure. Raise it as the system proves reliable.

Frequently asked questions

What is agentic commerce?

Agentic commerce is shopping and buying carried out by AI agents on behalf of a person or business. The agent finds products, compares options and completes the purchase within limits the user set. It depends on new payment protocols that let merchants and networks verify the agent and the user’s authorisation.

Is it safe to let an AI agent use my credit card?

It is safer when the agent uses a restricted token with spending caps, approved merchant categories and a human approval step for large purchases. Giving an agent a real card number with no limits is risky. Check whether your card issuer supports agent-specific controls before you try it.

What is the Agent Payments Protocol (AP2)?

AP2 is an open protocol created by Google for agent-led payments. It uses signed records called mandates to show what the user authorised and to bind a checkout to that authorisation. Google contributed the standard to the FIDO Alliance for stewardship in April 2026.

How are Visa and Mastercard approaching AI agent payments?

Visa runs Intelligent Commerce, built on tokenised credentials, real-time authorisation and a Trusted Agent Protocol for identity. Mastercard runs Agent Pay, built on agentic tokens, with Verifiable Intent added as a record of user authorisation. Both are rolling out in phases, working with AI companies and banks.

Who is responsible if an AI agent buys the wrong thing?

No universal rule exists yet. Responsibility depends on whether the agent followed the user’s mandate, exceeded it, or was manipulated by a third party. In current models the merchant remains the merchant of record, so refunds usually go through the merchant, but disputes over agent errors are unresolved.

Related reading

This article is general information, not legal, financial or investment advice. Rules, prices and product features differ by country and change often, so check current details with a qualified adviser or the provider. See our disclaimer. Last reviewed November 2, 2026.

Published by The Funding Journal Editorial. Practical writing on fintech, crowdfunding, investment platforms, SaaS and AI, and the stories of founders who raised capital.

Similar Posts