American Express adds intent checks to its AI commerce toolkit

Consent before checkout

An AI agent can misunderstand a request or act unpredictably. Amex’s framework is meant to catch that gap before payment: a customer sets the intent, confirms their identity and approves what the agent recorded. The system then creates an Intent ID and a Proof of Intent Token containing the key restrictions.

The agent uses protected digital payment details obtained with the token rather than the customer’s real credit card number. Amex says the token can also help prevent fraud and resolve disputes later in the transaction.

The framework’s Create Intent API captures the customer’s request and constraints, which might include:

The way those details are collected may vary by model. Amex says it wants the system to support different forms of agent-led commerce.

The distinction matters when an agent gets a request wrong. If a customer asks for white Nike sneakers in size nine but the agent tries to buy Adidas shoes, Amex says the payment can be blocked because the purchase does not match the approved intent.

Guardrails for more autonomous agents

Today, many agents ask customers to approve a purchase. Amex is also preparing for a less supervised model: a customer might authorize automatic orders for household goods such as paper towels or toilet paper, where price and package size matter more than brand or seller. In that case, an automated check could confirm that the order still fits the original instructions.

The company says it has not yet seen an increase in disputes tied to AI-agent commerce. Its case for building protections now rests instead on signs of consumer interest. In a recent Amex Trendex study, 66% of surveyed US consumers said they were willing to use AI agents to find a better price for a product or service. Another 61% said they might ask an agent to recommend products or services in the next 12 months.

Amex has named the framework part of its ACE Developer Toolkit, which launched in April. ACE includes tokens for AI agents containing Amex payment information and other developer tools. The company is discussing adoption with partners in the payments industry and plans to develop ACE as agent-led commerce changes.

A record of which agent received permission to buy, and when, could also help investigate fraud committed through an agent. Hannah Bowzian, Amex’s vice president of partnerships and AI agent strategy, told TechCrunch that the company wants to keep agent-led shopping convenient, reliable and secure for cardholders. She is due to discuss the framework at TechCrunch Disrupt 2026 on October 13.

The unresolved boundary

I think the interesting test is not whether the framework can reject an obvious mismatch like Nike versus Adidas. It is how precisely a customer’s intent can be recorded when the request leaves room for judgment: a preferred brand, an acceptable substitute, or a budget that changes with availability. Amex says the system must be flexible across models, but the announcement does not say how those boundaries will be represented or who decides when an agent has crossed them.

That question becomes more consequential as agents move from asking for approval on each purchase to acting on standing instructions. A token can carry constraints through checkout; it cannot, by itself, settle what a customer meant.