For years, the question about AI agents in commerce was whether they would be able to buy on our behalf. That question has been answered: yes, and the infrastructure is being built now. What is still unclear is how they will pay, and that is where a protocol race is under way that every bank and issuer should watch closely.

The map, simplified

Without going into every technical detail, three layers are being defined in parallel:

Card networks. Mastercard launched Agent Pay, which lets verified agents transact on behalf of a cardholder using agent-specific tokens, an extension of its tokenization service. Visa introduced its Trusted Agent Protocol, which verifies the agent’s identity through message signatures and an agent directory, and gives merchants an on-ramp to accept agent-initiated payments. Both follow the same logic: the merchant must be able to tell a legitimate agent from any bot, and the issuer must know on whose behalf it acts.

Processors and platforms. Stripe is pushing protocols that let an agent complete a purchase or pay for a service programmatically, even for very small amounts and at machine speed. OpenAI and Stripe defined an agentic commerce protocol so agents can buy within a conversation.

Big tech. Google published AP2, an open protocol that aims to standardize how an agent expresses the user’s purchase intent and authorization, regardless of the payment method.

The practical conclusion is that they are not mutually exclusive. A real agent-initiated purchase will likely touch two or three of these layers: the protocol the agent uses to express intent, the processor that executes it, and the network that settles and protects it.

The bank’s part

On this map, the bank or issuer rarely makes the headlines, but it carries the final risk. If an agent buys something the customer did not want, the dispute lands at the bank. So it pays to prepare on three fronts:

Identity and mandate. The issuer needs to know that a transaction was initiated by an agent, which agent, and under which customer mandate: amounts, merchants and validity. Network agent tokens point in that direction, but the bank must be able to read and enforce that mandate.

Authorization experience. Customers must be able to see, authorize and revoke their agents from the bank’s app, as clearly as they manage their cards today. It is probably the most underrated digital banking feature of the next two years.

Disputes and fraud. Today’s fraud models are trained on human behavior. Agents buy differently: faster, at any hour, with repetitive patterns. Rules and models need tuning so they neither decline legitimate transactions nor approve the wrong ones.

Preparing without marrying a protocol

With so many competing standards, the temptation is to wait for a winner. That would be a mistake: the bank that waits will be late, and the one that bets everything on a single protocol may be tied to the wrong one.

The alternative is architectural: build an in-house layer for agent mandates and authorization, decoupled from any specific protocol, and connect external protocols to it through adapters. The bank defines its rules once and can add or swap protocols as the market settles.

Agentic commerce will not wait for the industry to agree. The banks that design today how to recognize, authorize and control agents will be the ones that capture that volume.


I help banks and fintechs design the payments and identity architecture to operate with AI agents. See services.

Further reading: Juniper Research on Visa, Mastercard and Stripe’s agentic commerce push.