Amex builds guardrails for AI shopping agents before they misfire

American Express has built a framework that captures and authenticates a customer's purchase intent before an AI agent is allowed to complete a transaction — and it wants the rest of the payments industry to adopt it.

The company calls it the "Intent Intelligence Framework." It is Amex's answer to a problem that barely exists today but could scale fast: large language models are nondeterministic, and AI agents acting on a shopper's behalf are prone to doing what they were not asked to do — including making unauthorized purchases.

"These new transactions will have a consumer, a merchant, and an agent," Hannah Bozian, Amex's VP of Agentic Partnerships and Strategy, told TechCrunch. She said the company wanted agentic transactions to remain "seamless, trusted, and secure, the way our card members expect them to be."

Amex says it has not yet seen an uptick in disputes tied to agentic commerce. It is building the guardrails anyway, before this category of shopping grows.

Why does Amex think anyone wants this?

The consumer signal is real, at least by Amex's own polling. In a recent Amex Trendex report, 66% of surveyed U.S. consumers said they're likely to use AI agents to research the best price for a product or service over the next 12 months. Another 61% said they'd likely use an agent for a product or service recommendation in that window.

Payments networks face a structural question as agents start acting for shoppers. Card rules, fraud models, and dispute processes were all built around a human holding the plastic. When a bot holds a token instead, the industry needs a way to prove what the human actually asked for — and to carry that proof through the entire transaction lifecycle, from purchase to everything that follows.

That is the problem Amex says its framework solves.

How does the framework actually work?

The mechanics are concrete, and they start before any money moves:

• A Create Intent API captures the customer's purchase instructions plus any constraints they included — merchant name, maximum budget, product preferences.
• The customer then authenticates their identity and signs off on the intent the agent captured.
• The system generates an Intent ID, a unique identifier, alongside a Proof of Intent Token encoding the customer's high-level constraints.
• That token lets the agent obtain a secure digital payment credential it uses at checkout — not the customer's real credit card number.

How intent gets collected will vary depending on the AI model involved, but Amex designed the framework to be flexible enough to support all agentic commerce models, Bozian said.

The Intent ID and token aren't one-and-done artifacts. Amex says they can be used across the transaction process, giving companies a shared reference point for fraud prevention and dispute resolution. A log of what was authorized, by which agent, and when could also help address fraud committed in the agent's name, Bozian noted.

What does this prevent in practice?

Bozian gave the example that makes the risk obvious. A customer asks an agent to buy white Nike sneakers in a size 9. The agent submits a purchase request for Adidas shoes instead. Under Amex's framework, the transaction can be stopped before it goes through, because the agent made a mistake — the purchase doesn't match the intent the customer signed.

"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," Bozian said.

That timing is the design's core claim. Many agents today ask for customer consent at the moment of purchase. Amex is betting that agents will soon act far more autonomously, and that consent-at-checkout won't be enough.

Bozian pointed to household staples as the near-term use case. A consumer might let an agent re-order paper towels or toilet paper, where the only real constraints are price and size — not merchant or brand. In those scenarios, she said, it makes sense to have technology performing an extra check in the loop, verifying that each transaction still matches what the customer authorized.

Where does this fit in Amex's bigger agentic push?

The framework isn't a standalone product. It slots into the ACE Developer Toolkit — Amex's Agentic Commerce Experiences — which the company launched in April and which now includes the intent framework among its services.

Amex is already in discussions with partners across the payments ecosystem to implement ACE, including the agentic tokens that hold Amex payment credentials and other parts of the developer kit. The company says it plans to keep developing ACE as agentic commerce evolves — and it is explicitly hoping the Intent Intelligence Framework becomes an industry standard, not just an Amex feature.

Bozian will lay out the framework in more detail on October 13 at TechCrunch Disrupt 2026's Industry Stage, where she'll discuss what it means for agentic commerce.

What's at stake for the industry?

If AI agents become a mainstream checkout lane, the network that defines how intent is captured, authenticated, and enforced could shape the plumbing for everyone else. Amex hasn't seen the disputes yet, but it is moving early — building the verification layer before autonomous shopping becomes routine rather than courting regulators and angry card members after it does.