Amex introduces intent framework to rein in AI shopping agents

American Express is introducing a framework meant to stop AI shopping agents from completing purchases that a customer did not actually authorize. The company’s Intent Intelligence Framework is designed to capture a buyer’s instructions, authenticate them, and carry those constraints through the transaction.

That matters as personal AI agents begin to research products, compare options and, in some cases, buy on a user’s behalf. For consumers and small businesses, the question is no longer just whether an agent can find the right item, but whether it can be trusted not to buy the wrong one.

What Amex is adding to agentic checkout

According to TechCrunch, Amex’s new framework is built around the idea that a transaction involving an AI agent has three parties: the consumer, the merchant and the agent. The company wants the customer’s original purchase intent to remain visible throughout the payment flow, rather than being treated as a loose instruction given at the start.

The process begins with a Create Intent API. Amex says this captures the customer’s purchase directions and any limits attached to them, such as a merchant name, a maximum budget or product preferences. The exact way that information is collected may vary by AI model, but Amex’s goal is to support different agentic commerce approaches.

The customer then authenticates their identity and approves the intent that the agent has captured. After that, the system creates two pieces of transaction data: an Intent ID, which identifies the approved instruction, and a Proof of Intent Token, which contains the customer’s high-level constraints.

That token is then used to obtain a secure digital payment credential for checkout. Amex says the agent uses that credential instead of the customer’s actual card number, with the aim of keeping the purchase aligned with what the user approved.

The safety problem Amex is trying to solve

Large language models do not always behave predictably, and that creates an obvious risk when they are allowed to take action in commerce. A shopping assistant that misunderstands a prompt, substitutes the wrong product or ignores a constraint could create a real charge rather than just a bad answer.

Amex told TechCrunch it has not yet seen a rise in disputes tied to agentic commerce. Even so, the company says it is trying to prepare before this kind of buying becomes more common.

Hannah Bozian, Amex’s VP of Agentic Partnerships and Strategy, told TechCrunch that the company’s approach pulls intent earlier into the transaction flow. She gave the example of a customer asking an agent to buy white Nike sneakers in size 9, only for the agent to submit a request for Adidas shoes. Under Amex’s framework, that mismatch could stop the purchase from going through.

The broader point is that consent at the moment of purchase may not be enough if agents become more autonomous. Bozian also described a scenario where a user allows an agent to reorder household items, with constraints focused more on price and size than on a specific merchant or brand.

Why creators, freelancers and small businesses should care

For many Emonarc readers, AI agents are becoming part of everyday work: researching suppliers, preparing proposals, managing files or handling repetitive admin. We have already seen more products move from simple chat interfaces toward agents that perform multi-step tasks, as with enterprise agents such as Google’s Gemini agent for work across apps and systems designed to coordinate many agent actions, like Claude’s dynamic workflows.

Payments raise the stakes. If an agent can spend money, the user needs more than a general “yes” button. They need clear limits, a record of what was approved and a way for the payment system to recognize when the agent’s action no longer matches the instruction.

For a solo business owner, that could matter when an AI assistant is asked to buy office supplies, book services or source materials. For a small merchant, it could affect how agent-driven purchases are verified and how disputes are handled if a buyer later says the agent made a mistake.

Amex says the Intent ID and token can also help with fraud prevention and dispute resolution because they provide a record of what was authorized, by which agent and when.

What happens next

The framework is part of Amex’s Agentic Commerce Experiences, or ACE Developer Toolkit, which the company launched in April. TechCrunch reports that ACE now includes the intent framework, along with agentic tokens containing Amex payment credentials and other developer kit components.

Amex says it is already talking with partners across the payments ecosystem about implementing ACE. The company also says it plans to keep developing the toolkit as agentic commerce evolves.

The open question is whether the framework becomes a broader standard, as Amex hopes. For now, it is a sign that payment companies are preparing for a version of online shopping where AI agents do more than recommend products — they may also be trusted to complete the checkout, but only within boundaries the customer can prove they set.