Amex Builds Guardrails to Stop AI Shopping Agents From Buying the Wrong Thing

Picture asking a personal assistant to grab white sneakers in your size, only to watch them come back with the wrong brand entirely. Now imagine that assistant has your credit card, and it moves at the speed of software. That is the quiet risk sitting underneath the rush toward AI-powered shopping, and it is exactly the problem American Express says it is trying to solve.

AI agents, the software tools that can browse, compare prices, and complete purchases on a person's behalf, are moving from novelty to habit. But these systems run on large language models, which are nondeterministic by nature. In plain terms, that means they do not always do the same thing twice, even when given the same instructions. For a chatbot answering trivia questions, that quirk is harmless. For a bot authorized to spend your money, it is a liability.

Amex's answer is something it calls the Intent Intelligence Framework, announced this week and described by the company as an attempt to set an industry standard for how AI agents handle money. The idea is straightforward even if the mechanics are technical: capture exactly what a customer wants to buy, lock that intent in before the transaction happens, and make sure every step afterward stays true to it.

"These new transactions will have a consumer, a merchant, and an agent," Hannah Bozian, Amex's VP of Agentic Partnerships and Strategy, told TechCrunch. Her goal, she said, is to keep agentic transactions "seamless, trusted, and secure, the way our card members expect them to be." That phrase, agentic transaction, is becoming shorthand across the payments industry for any purchase where software, not a human finger on a keyboard, initiates the final click.

How the Framework Actually Works

The system starts with what Amex calls a Create Intent API. Think of it as a detailed shopping list the customer hands over before the agent does anything. That list can include the merchant's name, a maximum budget, product specifications, whatever constraints matter to the buyer. The customer then authenticates their identity and formally signs off on that captured intent, much like confirming an order before it ships.

From there, Amex generates an Intent ID, a unique tag for that specific request, along with something called a Proof of Intent Token. This token holds the customer's high-level constraints and is used to generate a secure digital payment credential. Crucially, the agent never touches the customer's actual card number. It uses this token instead, which keeps the real payment details shielded while still tying the purchase back to what was actually authorized.

Bozian offered a concrete example of why this sequencing matters. Say a customer asks an agent to buy white Nike sneakers in a size 9, but the agent, through some glitch in reasoning, submits a request for Adidas shoes instead. Because Amex's framework checks the request against the recorded intent before the purchase clears, that mismatched transaction can be stopped before money moves. "Our approach differs from the implementations of others in that it pulls intent earlier into the transaction flow," Bozian said. "This is important because it helps mitigate some risk before the risk occurs."

That distinction, catching errors before they happen rather than disputing them afterward, is the heart of what Amex is pitching. It is the difference between a smoke detector and a fire extinguisher. Both matter, but one saves you a lot more cleanup.

The framework also anticipates a future where agents act with less human oversight than they do today. Many current shopping bots still pause and ask for a final yes before checkout. But Bozian pointed to routine purchases, reordering paper towels or toilet paper, for instance, where price and size are the only real variables and brand loyalty barely factors in. In those lower-stakes, repetitive scenarios, Amex sees room for agents to act more autonomously, provided there is still a verification layer quietly checking that the purchase lines up with what was approved. The Intent ID and its accompanying token double as a running log, Bozian noted, which could help sort out fraud claims or disputes involving an agent's actions down the line.

This isn't Amex's first move into agentic commerce. The Intent Intelligence Framework builds on the company's Agentic Commerce Experiences toolkit, known as ACE, which launched back in April. Amex says it has not yet seen a spike in disputes tied to AI-driven purchases, but the company wants to get ahead of the curve rather than scramble to catch up once problems surface. That instinct lines up with consumer appetite: in a recent Amex Trendex survey, 66% of U.S. consumers said they were likely to use AI agents to research the best price for a product over the next year, and 61% said they would use one for product recommendations. People are already leaning into this technology. The infrastructure to keep it safe is still catching up.

Amex says it is in active talks with partners across the payments ecosystem to bring ACE, including the agentic tokens carrying its payment credentials, into wider use. The company plans to keep developing the toolkit as agentic commerce matures, and Bozian is set to discuss the framework in more detail at TechCrunch Disrupt 2026.

Why It Matters

Financial fraud protections have always lagged a step behind new technology, and agentic commerce is no exception. What makes this moment different is scale and speed. A single confused AI agent might misplace one order. A fleet of them, deployed across millions of accounts, could misplace thousands simultaneously, and do it fast enough that traditional fraud review processes struggle to keep pace. Building verification into the front end of a transaction, rather than relying purely on after-the-fact dispute resolution, is a sensible hedge against that risk. Whether other card issuers and merchants adopt something similar, or whether Amex's framework becomes the de facto standard it hopes for, will say a lot about how responsibly this next wave of automated spending gets handled.