Amex introduces a framework to keep AI purchases within customer-approved limits

American Express introduced its Intent Intelligence Framework on October 13, 2026, aiming to stop AI shopping agents from spending outside a customer’s instructions. The approach puts authenticated purchase limits into the payment process. It is a preventive measure, not a response to a reported surge: Amex says it has not seen agent-related disputes increase.

The framework joins Amex’s Agentic Commerce Experiences developer toolkit, launched in April. Its purpose is to carry the customer’s authorized instructions through the purchase and what follows, rather than leaving those instructions only inside the conversation with an agent.

Turning a shopping request into payment rules

Amex’s design covers different levels of delegation. A customer might give an agent a broad task with a spending ceiling, authorize a specific purchase when an item returns to stock, or approve a finished cart. Those choices leave different amounts of work—and room for error—to the agent.

The agent typically translates the request into structured instructions, asking follow-up questions if something is unclear. The Create Intent API, a software interface, shares the confirmed purchase instructions, including the merchant, maximum budget and other preferences. The customer then authenticates their identity and signs off on the instructions captured by the agent.

Sign-off can use a passkey, or the card issuer can delegate it to a third party such as an agent or wallet provider. After authentication, the system issues an Intent ID identifying the purchase record and a Proof of Intent Token carrying its high-level constraints.

That token can obtain a payment credential restricted by the approved instructions. The agent uses the secure digital credential at checkout instead of the customer’s actual card number. Amex describes these controls as a way to constrain spending and supply fraud-risk signals before the purchase is executed.

Our approach differs from the implementations of others in that it pulls intent earlier into the transaction flow. And this is important because it helps mitigate some risk before the risk occurs,

Hannah Bozian, Amex VP of Agentic Partnerships and Strategy, speaking to TechCrunch

Approval does not eliminate the agent’s work

Amex’s technical explanation identifies risks even after customer approval. An agent could change an approved cart before checking out, or make a mistake while filling a cart the customer previously reviewed. With a broad request, it must also interpret the desired goods, find an appropriate merchant and execute only when the conditions are met.

The framework allows more detailed boundaries than a price ceiling. Depending on the customer’s instructions and the implementation, controls can include:

• An expiry date confirmed by the user, with single-use or multi-use tokens supporting purchases across one or more merchants.
• A maximum number of purchases, mapped to the maximum number of payment credentials that can be issued.
• Linked-item conditions—for example, buying flight insurance only if the flight tickets were purchased.

The services also support updating and canceling authenticated instructions. Responsibility for collecting, storing and sharing them can shift among agents, wallet providers and intermediaries. Amex says purchase summaries should exclude personal details irrelevant to the transaction, such as why a customer is buying a flight.

The approval record works both ways

Amex says issuers can use the confirmed intent record to evaluate later disputes against what the cardholder authorized. That makes the record useful beyond blocking an incorrect purchase. The company also identifies the risk of customers fraudulently disputing agent purchases they did approve, giving issuers another reason to retain the authorization evidence.

Amex told TechCrunch it is discussing implementation of the toolkit with partners across the payments ecosystem. The company hopes the framework becomes an industry standard and plans to keep developing the toolkit as AI-led commerce evolves.