Agentic commerce means AI agents that find, choose and pay for things on a customer's behalf without the customer approving each step. Since 2025 the card networks, Google and Stripe have published the first frameworks for how these payments are identified, authorised and settled. For processors, issuers and acquirers, the effects on authorisation, fraud and liability are already visible.
The infrastructure being built
Visa Intelligent Commerce gives AI agents tokenised card credentials with spending controls set by the customer. Visa's Trusted Agent Protocol, announced in October 2025, uses cryptographic signatures so that a merchant can tell a legitimate shopping agent from a malicious bot. In April 2026 Visa extended its Agentic Ready testing programme for issuers to more than 85 partners in Asia Pacific and Latin America.
Mastercard Agent Pay, announced in April 2025, registers AI agents and issues them tokenised credentials so that issuers and merchants can recognise agent-initiated transactions and apply the limits the customer has set.
Google's Agent Payments Protocol (AP2), released in September 2025 with more than 60 partners, is an open protocol rather than a network product. It defines signed mandates that record what the user asked for and what the agent proposes to buy, so there is evidence of the user's instructions if a dispute arises.
Stripe launched the Agentic Commerce Protocol with OpenAI in September 2025 to let ChatGPT users complete purchases inside the chat, and in March 2026 introduced the Machine Payments Protocol for payments between software agents and services. PayPal has also opened merchant catalogues and checkout to shopping agents.
What changes in authorisation
Today's authorisation logic assumes a person who can respond to friction such as a 3DS challenge, a one-time passcode or a biometric prompt. An agent working at machine speed cannot. The new frameworks rely instead on registered agent credentials, limits on what the agent may buy (merchant category, maximum value, frequency) and cryptographic proof that the agent is acting within the customer's instructions.
For issuers, agent traffic looks different from human traffic. Adobe reported a 4,700% year-on-year rise in visits from generative AI tools to US retail websites in mid-2025. An agent making several purchases in quick succession from different merchants may simply be following its instructions, which is not in itself a fraud pattern. Fraud models trained on human behaviour will need recalibrating.
Fraud liability in agent-initiated payments
Liability for agent-initiated fraud is not settled. Under current scheme rules, card-not-present fraud without 3DS generally falls on the merchant or acquirer. When the agent's authority was itself set up fraudulently, for example through account takeover or by tricking the customer, the chain of liability becomes much less clear. Both Visa and Mastercard have said they will adapt their rules for agent payments, and acquirers should follow those changes closely.
What acquirers and processors should do
Acquirers and processors should check four things. Can their systems handle the credential formats and message fields the new protocols use? Are their fraud models ready for traffic that differs from human behaviour in speed, timing and merchant mix? Do merchant contracts and scheme registrations need updating for agent payments? And do risk teams have a way to assess the exposure created when an agent's authority is set up?
In the GCC, where card-on-file and wallet payments are growing quickly, payment institutions licensed by the CBUAE, SAMA or QCB should raise agent authentication with their regulator early, before launching commercially. Supervisory expectations for agent payments are still taking shape.