Amex to agents: prove you meant to buy that

The question nobody in agentic commerce wants to answer is who pays when the agent buys the wrong thing. American Express would like the answer to be nobody, because the purchase never goes through.

On October 13, Amex introduced what it calls the Intent Intelligence Framework, presented by Hannah Bozian, its VP of Agentic Partnerships and Strategy, at TechCrunch Disrupt. The idea is straightforward. Before an agent completes a purchase, the framework captures what the customer actually asked for, gets the customer to authenticate and sign off on it, and then carries that intent through the transaction so the agent cannot quietly substitute its own judgment.

The mechanics, per Amex: a Create Intent API records the purchase instructions along with constraints like the merchant name, maximum budget, and product preferences. How the intent gets collected can vary by AI model; Amex wants the framework flexible enough to cover every agentic commerce model. Once the customer approves, the system issues an Intent ID and a Proof of Intent Token carrying the high-level constraints. That token is then used to obtain a secure digital payment credential, which the agent spends at checkout instead of the customer's real card number. The Intent ID and token stick around for fraud prevention and dispute resolution.

Bozian's pitch is about timing. "Our approach differs from the implementations of others in that it pulls intent earlier into the transaction flow," she told TechCrunch. "It helps mitigate some risk before the risk occurs." The example: a customer asks an agent for white Nike sneakers in size 9, and the agent submits a purchase request for Adidas. Under the framework, Amex can kill the transaction because the intent and the request do not match. The failure mode being designed against is the ordinary one: large language models are nondeterministic, and an agent that mostly does what you asked will occasionally do something adjacent.

Two things make this interesting beyond the press release. First, Amex says it has not yet seen an increase in disputes from agentic commerce. This is preemptive infrastructure, built before the problem arrives rather than after. Amex's own Trendex survey found 66 percent of US consumers say they are likely to use AI agents for price research and 61 percent for product recommendations in the next year. The company is betting the wave is coming and wants its rails to be the trusted ones when it does.

Second, the framework assumes a future most agent builders are already designing for: agents acting with real autonomy. Bozian's example is the household staple reorder, paper towels or toilet paper, where the only constraints are price and size rather than merchant or brand. In that world, asking the human to approve every purchase does not scale, so the check has to live in the transaction itself. Amex also keeps a log of what was authorized by which agent and when, which is the part that matters when something goes wrong and everyone points at everyone else.

The honest caveats. This is a framework announcement, not a shipped product. It sits inside the Agentic Commerce Experiences toolkit Amex launched in April, and Amex says it is in discussions with payments-ecosystem partners about implementation. Public docs, pricing, and partner terms were not part of the announcement, and no independent assessment of the scheme exists yet. Amex hopes it becomes an industry standard. Hope is not a standard.

If you build agents that touch money, the takeaway does not depend on Amex winning. Design the purchase flow so intent is captured and verified before checkout, not reconstructed after. Keep a per-agent authorization log. And assume the dispute wave Amex is building ahead of will arrive on your timeline too.

What remains unknown: the API surface, the partner list, and whether merchants and rival card networks adopt Amex's framing or build their own. The framework's value depends entirely on who implements it.

Related: know your agent before it spends on payment identity for agents, and Robinhood's real-money agents for what happens when agents touch live funds.