American Express Wants to Check What You Meant Before an AI Agent Buys It

American Express published a framework on Tuesday for a problem that barely existed two years ago: an AI agent that buys the wrong thing on your card. The Intent Intelligence framework captures what a customer actually authorized before an AI agent completes a purchase, then carries that record through the transaction as a control — and Amex is pitching it as a candidate industry standard for agentic commerce.

The mechanics run in four steps. A Create Intent API first records the customer's purchase instructions and constraints: merchant name, maximum budget, product preferences and other conditions. The customer then authenticates their identity and signs off on the intent the agent captured. Amex issues an Intent ID plus a Proof of Intent Token carrying the high-level constraints, and that token is used to obtain a secure digital payment credential the agent presents at checkout instead of the customer's actual card number. The same Intent ID and token can later support fraud prevention, authorization decisioning and dispute resolution.

Hannah Bozian, Amex's VP of Agentic Partnerships and Strategy, offered a concrete failure case: a customer asks an agent for white Nike sneakers in size 9, and the agent submits a request to buy shoes from Adidas. Because the approved intent names Nike, Amex could stop the transaction before it settles. "Our approach differs from the implementations of others in that it pulls intent earlier into the transaction flow," she told TechCrunch. "That helps mitigate some risk before the risk occurs."

The framework is part of the Agentic Commerce Experiences (ACE) Developer Kit Amex launched in April, which now exposes Account Enablement, Intent Intelligence and Payment Credentials specifications, with Agent Registration and Cart Context still in development. Amex says it is engaging with EMVCo on card-based agentic transactions, supporting the x402 Foundation's work on internet-native payment standards, and plans to offer Amex Agent Purchase Protection for eligible purchases made by registered agents with authenticated intent.

Amex is building ahead of the evidence. The company said it has not yet seen a meaningful increase in disputes tied to agentic commerce, but its own Trendex survey of 2,005 US adults found 66% are likely to use AI agents to research the best price on a product or service over the next 12 months, and 61% would use agents for recommendations. Comfort drops sharply with stakes: 52% would let an agent buy everyday low-cost items, but only 15% would trust one with high-value purchases like electronics and appliances.

Bozian expects that boundary to move. Agents today ask for consent before nearly every purchase, but she anticipates more autonomous behavior in categories where the only real constraint is price and size — reordering paper towels, for instance. In those cases a machine-side check on whether the transaction matches what was authorized becomes the practical safeguard, and a log of which agent was allowed to do what, and when, becomes the practical tool for catching fraud committed through an agent rather than by one.

The deeper point is that payments infrastructure is quietly becoming the compliance layer of the agent economy. Model providers can align an agent's behavior, but only the card network can decide who bears the loss when behavior drifts. Whoever standardizes how intent is captured, scoped and evidenced will effectively write the dispute rules for a new class of transactions — and Amex, as both an issuer and a network, is trying to write them first.