The Pulse
Amex Ties AI-Agent Purchases to Customer-Approved Intent
American Express published its Intent Intelligence framework on October 13, 2026, proposing authenticated purchase instructions and Proof of Intent Tokens for AI-agent transactions. The framework is part of Amex’s ACE Developer Kit and is d
AI.info Team · Written with AI ·
“Intent Intelligence is designed to provide these guardrails through data sharing that enables payment credential scoping and user sign-off.”
Inderpreet Hundal and Gianna Rotondo, American Express
American Express published its Intent Intelligence framework on October 13, outlining a way to bind an AI agent’s purchase to customer-approved instructions before checkout. The proposal records a customer’s constraints, authenticates their approval, then carries that authorization through payment using identifiers and tokens. Amex presents it as a recommended approach for agentic commerce, not a claim that the system is already in wide use.
Approval becomes a payment boundary
The framework starts with a Create Intent API that captures a customer’s instructions, including a maximum spend, merchant preferences and other purchase conditions. The customer authenticates their identity and signs off on the intent the agent has recorded. Amex says the sign-off may use a passkey, or an issuer may allow another party, such as an agent or wallet provider, to handle that step.
After approval, the system creates an Intent ID for the purchase-intent record and a Proof of Intent Token, a JSON Web Token that represents the intent’s high-level constraints. The Proof of Intent Token can be used to obtain a separate, scoped digital payment credential for checkout, instead of giving the agent the customer’s card number. Amex says the credential can be limited by the approved intent and that the token can expire on a date the customer confirms.
The token may be single-use or support multiple credentials for purchases across one or more merchants, according to the framework. The intent record can also be updated or canceled, and Amex proposes using its identifiers later to check whether a payment request matches the authorization. Those functions are meant to provide information for authorization, fraud assessment and dispute handling; the framework does not promise that every mistaken or disputed purchase will be prevented.
Budgets, merchants and item conditions
The framework describes a range of details an intent may carry: a purchase summary, maximum amount, expiry date, preferred merchant and conditions linking items in an order. It also allows an intent to set a maximum number of transactions, which Amex says can determine how many payment credentials are issued. For privacy, the document says personal details unrelated to the purchase should be left out of the intent data.
Amex gives examples ranging from a broad spending limit for a task to a specific, time-limited instruction to buy a particular item if it becomes available. The company also distinguishes that kind of open-ended intent from a cart the customer has reviewed and approved. That distinction matters because an agent may have to choose the product and merchant itself in one case, while in another it may be able to change a cart after approval.
Amex wants other payment partners involved
Intent Intelligence is a service within Amex’s Agentic Commerce Experiences, or ACE, Developer Kit. The company says the framework is designed to work across different agentic-commerce arrangements, in which an agent, a token vault or a digital wallet may collect, store or transmit the customer’s intent. In each version, Amex argues that issuers should receive authenticated intent before a purchase so they can assess the request against the customer’s authorization.
In its October 13 announcement, Amex said it was engaging with EMVCo on card-based agentic transactions and supporting the x402 Foundation’s work on internet-native payment standards. It also said ACE integrations are under development. That leaves a practical gap between a published framework and a working network of agents, wallets, merchants and issuers that consistently passes the same authorization details.
A framework, not evidence of fewer disputes
The proposal addresses a specific weakness in agent-led shopping: an agent can misread an instruction, select the wrong product or act outside a customer’s intended limits. Recording the user’s approval before payment gives issuers a reference point for evaluating the eventual transaction. Whether it stops errors depends on how accurately agents translate natural-language requests into structured constraints and whether payment partners enforce those constraints at checkout.
Amex frames the design as a way to reduce risk and make later disputes easier to assess, but its published framework does not provide measured results from live deployments. Its near-term test is whether partners adopt the intent fields and credential controls consistently enough for the customer’s approved limits to follow a purchase from instruction through payment.