Amex's Intent Intelligence Framework would check an AI agent's order against what the customer approved

A customer asks an agent for white Nike sneakers in a size 9, and the agent submits an order for Adidas ^[9]. Hannah Bozian, Amex's VP of Agentic Partnerships and Strategy, used that example with TechCrunch and said Amex could stop the transaction because the agent made a mistake ^[9]. The risk comes from the models. TechCrunch wrote that large language models are nondeterministic, so bots are prone to doing what they are not supposed to ^[10].

The stop works because the instruction is on record before the agent shops. "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 ^[15]. She also described the parties involved. "These new transactions will have a consumer, a merchant, and an agent," she told TechCrunch ^[11].

The check is only as specific as the instruction behind it. Bozian noted that many agents today ask for consent before buying, while Amex foresees agents acting more autonomously in some cases ^[16]. Her example was re-ordering household staples such as paper towels, where the only real constraint is price and size, not merchant or brand ^[16]. An intent check on that order would pass any brand at the right price and size, and that is what the customer asked for.

The check also starts after a human step. The customer authenticates and signs off on the intent the agent captured ^[12], and Bozian said the way intent is collected may vary by AI model ^[17]. The framework can compare an order with the signed intent. It cannot tell whether the signed intent matches what the customer meant. In our reading of the flow Amex described, an agent that records only "shoes" and a customer who taps approve would let an Adidas order through. How much of the protection is real depends on how closely customers read the sign-off, and TechCrunch's report does not include a figure on that.

We think teams should require specific, signed constraints before giving an agent standing permission to buy, at the cost of more questions at setup. Two variables sort the cases: how specific the signed instruction is, and whether a person confirms each purchase or the agent acts on standing permission.

- Specific instruction, standing permission: the Nike case, with no person to catch the Adidas order. The intent check matters most here. - Specific instruction, person confirms each time: the person and the check cover the same mistakes. - Loose instruction, person confirms each time: the person is the control. - Loose instruction, standing permission: the paper towel case. The price and size limits are the entire control, so they deserve the most care.