American Express published its Intent Intelligence framework on 13 October, setting out the technical specification for recording, authenticating and transmitting what a customer instructed an AI agent to purchase before any charge is processed. The move makes Amex the first major card network to publish detailed plumbing for agentic commerce, and it has direct implications for how liability will be allocated across the payments stack as autonomous purchasing agents proliferate.
How the framework works
The architecture centres on a Create Intent API that captures a customer’s instructions and constraints, covering the merchant, a maximum budget and other preferences. The customer then authenticates and signs off on what the agent recorded. Amex issues an Intent ID and a Proof of Intent Token encoding those high-level constraints. The agent exchanges the token for a scoped payment credential used at checkout in place of the actual card number, and the Intent ID and token travel with the transaction through authorisation, fraud decisioning and any subsequent dispute process.
Hannah Bozian, Amex’s VP of agentic partnerships and strategy, described the rationale to TechCrunch as pulling intent earlier into the transaction flow so that risk is assessed before a charge lands rather than after. Her illustrative example: a customer instructs an agent to buy white Nike trainers in a size nine, the agent attempts to purchase Adidas instead, and Amex can block the transaction at the point of credential validation.
The framework extends commitments Amex made in April when it launched its ACE developer kit and Agent Purchase Protection, under which eligible cardholders who authorise an agent and pass authenticated intent to Amex would be protected from charges caused by agent error. The April release described that protection as a future commitment with terms still to be finalised, and two of the kit’s five services remained in development at that point.
Market context and regulatory read-across
Amex can move quickly here because it operates a closed-loop model as issuer, network and acquirer simultaneously, giving it end-to-end visibility over a transaction that a four-party scheme cannot replicate without coordination across multiple counterparties. That same four-party challenge is being addressed at an industry level through EMVCo‘s draft Intent Services framework, whose public comment window closed on 30 September, and through a voluntary principles paper published last month by a group of six global banks including Bank of America.
The direction of travel across all three workstreams is consistent: liability will follow whoever holds an authenticated, signed record of what the customer actually authorised. That creates an immediate practical question for issuers and processors operating on Visa and Mastercard rails. Any bank or processor whose agent partners cannot capture, cryptographically sign and pass intent data through the authorisation chain will find itself in a weaker dispute position as agentic transactions scale.
The regulatory frame is still forming. Neither the FCA nor the Payment Systems Regulator has yet issued specific guidance on AI agent transactions, though the PSR’s existing authorised push payment fraud reimbursement rules place significant weight on whether a payer gave genuine, informed consent. An authenticated intent record of the kind Amex describes would be a compelling artefact in that context. European regulators working through PSD3 implementation will face the same definitional question: at what point does an agent’s action constitute the customer’s consent?
For merchants, a payment credential scoped to a specific intent represents a meaningfully stronger dispute position than a stored card number submitted by an automated process. The commercial incentive to accept intent-scoped credentials is real, and acquirers will need to decide how quickly they build the acceptance infrastructure.
The post Amex Publishes Intent-Authentication Framework for AI Agent Payments appeared first on The Fintech Times.
