Editorial illustration for Amex introduces framework to prevent AI agents from unauthorized purchases
Amex Shields Cards From Unauthorized AI Agent Purchases
Amex introduces framework to prevent AI agents from unauthorized purchases
American Express wants to settle a question that's about to get expensive for somebody: when an AI agent buys the wrong thing, who's on the hook. The company is rolling out what it calls the Intent Intelligence Framework, built to capture and verify what a customer actually asked for before an AI agent hands over the card number. The problem is baked into how large language models work.
They're nondeterministic by design, which means an agent told to book a flight or order groceries can occasionally go off script and complete a purchase nobody approved. Amex's fix is meant to keep the customer's original instructions attached to the transaction from the moment it starts to the moment it settles, rather than letting that intent get lost once an agent takes over.
The framework builds on Amex's Agentic Commerce Experiences toolkit, launched in April, and arrives before any real wave of disputes has hit the company's books. That timing is deliberate. Amex is betting agentic shopping is about to scale fast, and that card members, merchants, and the agents acting between them will need a shared standard for proving what was actually authorized.
Then an Intent ID, a unique identifier representing the intent, is created, along with a Proof of Intent Token, which includes the customer’s high-level constraints. This token is used to obtain a secure digital payment credential that the agent will use at checkout instead of the customer’s actual credit card number. In doing so, Amex says, the purchase will remain secure while being aligned with what the customer authorized the agent to do.
Why this matters
Amex's intent framework is a tell: the card networks don't trust AI agents yet, and they're probably right not to. A Create Intent API that logs what a customer actually wanted before an agent spends money on their behalf is a patch over a real gap. Large language models don't execute instructions the way traditional software does, and "nondeterministic" is a polite word for "sometimes does the wrong thing with your credit card." For founders building shopping agents, this is a preview of the compliance layer you'll need before banks let your product touch real transactions.
For developers, it's a signal that intent verification, not just authentication, is becoming a required piece of agentic commerce infrastructure. Amex's own numbers, consumers already leaning on AI for price research and recommendations, show the demand is there well ahead of the trust infrastructure. That gap is the opportunity and the risk.
Whoever builds the standard for proving what a user actually meant to buy, before the agent clicks "purchase," is building a piece of plumbing every agentic commerce product will eventually need.
Related Reading
LLMs & Generative AI ChatGPT's 'Nerdy' tweak rewards goblin metaphors in answers, study finds
LLMs & Generative AI Google to End Free Access to Gemini AI for Non-Subscribers
LLMs & Generative AI AI Engineers Face Rising Costs, Need New Strategies for Efficiency
From Archives Anthropic Cuts Claude Reasoning Token Theft by Up to 99%
From Archives Longtime Wikipedian's "Wake-Up Call" Notes Encyclopedia's Decline Since 2022
Common Questions Answered
What is American Express's Intent Intelligence Framework designed to prevent?
The Intent Intelligence Framework is designed to prevent AI agents from making unauthorized or incorrect purchases on behalf of customers. It captures and verifies what a customer actually requested before the AI agent accesses payment information, addressing the inherent nondeterministic nature of large language models that can cause them to execute instructions incorrectly.
How does the Proof of Intent Token protect customers during AI-authorized purchases?
The Proof of Intent Token includes the customer's high-level constraints and is used to obtain a secure digital payment credential instead of sharing the actual credit card number with the AI agent. This ensures the purchase remains secure while being aligned with what the customer originally authorized the agent to do.
Why are large language models considered nondeterministic when handling purchase requests?
Large language models are nondeterministic by design, meaning they don't execute instructions the way traditional software does and can sometimes produce unpredictable results. An AI agent told to book a flight or order groceries can occasionally misinterpret the request and make the wrong purchase, which is why Amex's framework is necessary to verify intent before payment.
What role does the Intent ID play in Amex's framework?
The Intent ID is a unique identifier created to represent what the customer actually asked for before an AI agent processes a transaction. It works in conjunction with the Proof of Intent Token to ensure that the payment credential issued to the agent is specifically tied to the customer's original authorized intent.
Further Reading
- Amex introduces a framework to prevent AI agents from making unauthorized purchases - TechCrunch
- American Express Is Telling Merchants How to Prepare for Agent Shopping — and Offering to Pay When It Goes Wrong - Yahoo Finance
- American Express Debuts Agentic Commerce Experiences (ACE) Developer Kit and Announces Industry-First Protection for Registered Agent Purchases - American Express
- Agentic commerce liability is still being written - Worldpay
- Who Pays When Your AI Agent Buys The Wrong Thing? - Forbes
