ANALYSIS American Express is making the first major card-network move to impose transactional discipline on AI agents, publishing a framework that embeds customer intent as a verifiable control before an agent can spend. The bet arrives alongside survey data showing that governance-as-document already fails inside enterprises, raising the question of whether governance-as-infrastructure can succeed where policy PDFs do not.
Why it matters
AI agents that can browse, compare, and buy on a user's behalf are no longer hypothetical. American Express published its Intent Intelligence Framework on October 13, describing it as a recommended approach to capturing and authenticating a customer's purchase intent before an AI agent completes a transaction5,6. The company made the framework publicly available, aiming to help advance a common industry approach for intent in agentic commerce. Amex chairs the EMVCo agentic commerce working group4, positioning the framework not as a proprietary feature but as a candidate standard for the broader payments ecosystem.
The big picture
The framework rests on a specific technical sequence. It begins with a Create Intent API that records a customer's purchase instructions and constraints, including merchant name, maximum budget, and product preferences2,3. The customer then verifies their identity and reviews and approves the purchase intent gathered by the agent. After approval, an Intent ID and a Proof of Intent Token are created. Using that token, the agent receives a secure digital payment credential in place of the actual card number. The Intent ID and token persist through subsequent steps of the purchase journey, supporting fraud prevention, authorization decisioning, and dispute resolution.
"These new transactions will have a consumer, a merchant, and an agent," said Hannah Bozian, Amex's VP of Agentic Partnerships and Strategy1. She said the company wanted to ensure that "agentic transactions are still seamless, trusted, and secure, the way our card members expect them to be". Bozian described the approach as bringing intent earlier into the transaction flow, mitigating risk before it occurs, citing the example of blocking an agent that tries to buy the wrong brand of sneakers.
ANALYSIS The architecture's distinguishing feature is temporal: it makes authorized purchase intent available to the issuer before the agent executes the purchase, not after a dispute is filed. That inversion matters because the traditional chargeback process assumes a human clicked "buy." When an autonomous agent acts on interpreted instructions, the evidentiary basis for disputes collapses. Amex is, in effect, creating a pre-transaction audit trail that can answer the question regulators and arbitrators will inevitably ask: did the customer actually want this?
The framework extends Amex's Agentic Commerce Experiences (ACE) Developer Toolkit, which launched in April. Amex said it has not yet seen a rise in disputes from agentic commerce but is acting proactively, and is in talks with payments partners to implement ACE.
ANALYSIS The governance gap Amex is trying to close at the transaction layer mirrors a broader pattern inside enterprises. A Zapier survey of 548 U.S.-based directors, VPs, and C-suite executives at companies with 500 or more employees found that 79% say their employees are actively bypassing or working around governance processes to deploy or modify AI-powered workflows faster9,10. Some 91% of executive leaders say their organizations have AI governance policies, yet 41% say their organization's goal is moving fast and refining governance as they go, even if it risks compliance.
The Zapier report's conclusion aligns with Amex's design philosophy: "The organizations that got ahead of this didn't do so by writing a better PDF but by building the governance into the tools themselves"11. ◆ Amex's Create Intent API and Proof of Intent Token are precisely that kind of embedded control. Rather than relying on a cardholder to remember spending limits or an agent developer to self-police, the framework makes the constraint a structural prerequisite for obtaining payment credentials.
Amex's vertically integrated model, serving simultaneously as issuer, acquirer, and network, gives it a structural advantage in deploying such a framework unilaterally. Bozian acknowledged the framework was initially developed with that model in mind but said it is flexible enough to work with other models. Whether Visa, Mastercard, or open-banking rails adopt compatible intent layers will determine whether the standard holds or fragments.
What's next
Amex said it is in active talks with payments partners to implement the ACE toolkit. The company's role as chair of the EMVCo agentic commerce working group gives it a formal channel to push the Intent Intelligence Framework toward broader adoption. Bozian framed the framework's key trait as "its flexibility", a signal that Amex expects to negotiate interoperability rather than dictate terms. The next test is whether competing networks build compatible intent layers or propose alternatives, a decision that will shape liability rules for every AI agent with a credit card on file.