Amex Releases Technical Framework for AI Agent Payment Authentication

The system revolves around a Create Intent API that records what a customer tells an autonomous agent to buy, including which merchant to use, a spending ceiling, and other parameters. Once the agent logs the instructions, the customer reviews and digitally signs them. American Express then generates an Intent ID alongside a Proof of Intent Token that encodes the approved constraints.

The agent uses that token to obtain a limited-scope payment credential rather than exposing the underlying card number. Both the Intent ID and the token accompany the transaction through every stage — authorisation, fraud screening, and any chargeback or dispute that follows.

Hannah Bozian, who leads agentic partnerships and strategy at the card network, explained to TechCrunch that the design shifts intent capture upstream so risk evaluation happens before a charge hits the account instead of after. She offered a practical scenario: a cardholder asks the agent to purchase white Nike trainers in size nine, the agent tries to buy Adidas shoes instead, and Amex blocks the attempt when the credential is validated because it violates the signed intent.

The October release builds on promises the company made in April when it launched its ACE developer kit and Agent Purchase Protection programme. Under that programme, eligible cardholders who authorise an agent and supply authenticated intent data to Amex would gain protection against charges triggered by agent mistakes.

The April announcement described the protection as a future offering with final terms still under negotiation, and two of the five services in the developer kit remained under development at the time.

Amex can deploy this infrastructure rapidly because it runs a closed-loop model, acting simultaneously as issuer, network, and acquirer. That gives the company complete visibility across a transaction in a way that open-loop schemes cannot achieve without coordinating multiple participants.

The broader card industry is tackling the same challenge through EMVCo's draft Intent Services framework, which closed its public consultation period on 30 September, and through a voluntary principles document that six global banks including Bank of America released last month.

All three initiatives share a common direction: liability will rest with whichever party possesses an authenticated, cryptographically signed record of what the customer actually approved. That creates an immediate operational question for issuers and processors using Visa and Mastercard networks. Any bank or processor whose agent partners cannot capture, sign, and transmit intent data through the authorisation flow will end up in a weaker position when disputes arise as agentic transactions grow.

The regulatory landscape remains unsettled. Neither the FCA nor the Payment Systems Regulator has published specific guidance on AI agent transactions, though the PSR's current rules for authorised push payment fraud reimbursement already emphasise whether the payer gave genuine, informed consent. An authenticated intent record of the type Amex describes would carry significant weight in that assessment.

European regulators implementing PSD3 will confront the same definitional challenge: when does an agent's action legally constitute the customer's consent?

For merchants, a payment credential tied to a specific signed intent offers a materially stronger position in disputes than a stored card number submitted by an automated process. The commercial case for accepting intent-scoped credentials is tangible, and acquirers will need to determine how quickly they invest in the acceptance infrastructure to support them.