The new trust layer in agentic commerce

American Express announced the Intent Intelligence Framework this week, a system designed to capture and authenticate the customer's purchase intent before an AI agent completes the transaction. The framework is part of a broader company effort to establish trust standards in what it calls "agentic commerce" — the set of automated financial interactions mediated by artificial intelligence.

The initiative addresses a problem that will become increasingly expensive as autonomous agents for purchasing expand: when a language model decides to buy something different from what the user requested, who bears the loss? Until now, there has been no industry standard for addressing this issue.

The nondeterminism problem

The heart of the problem lies in the nondeterministic nature of large language models. Unlike traditional scripts, an LLM can interpret similar instructions differently at different times. A customer might ask "buy something cheap for the house" and the agent could interpret it differently each time — sometimes buying a $20 accessory, other times a $500 appliance.

The Intent Intelligence Framework attempts to solve this by creating a cryptographic record of the customer's intent. The system works in steps: first, a customer instructs their shopping agent about what they want; then, the agent converts that intent into an "Intent ID" that is registered on American Express's network before any money is spent.

It's the financial equivalent of "measure twice, cut once" — and a direct response to the fact that digital payment infrastructure was designed for humans, not for autonomous systems that interpret instructions probabilistically.

From API to protection: the evolution of the ACE program

This October release was not American Express's first step into the agentic commerce space. In April 2026, the company had launched the Agentic Commerce Experiences (ACE) Developer Kit, a set of technical tools that allows developers to integrate American Express cards into AI interactions — including agent registration, payment tokens, and cart context.

The intent framework fits into this ecosystem as the missing verification layer. Without intent authentication, a registered agent could transact, but there was no guarantee it was doing what the customer actually wanted. The Create Intent API adds this layer of explicit validation.

The company also announced the Amex Agent Purchase Protection, a guarantee that covers costs resulting from errors by registered agents. In other words, American Express commits to absorbing the loss when an authorized agent makes a mistake — provided the customer registered their purchase intent and the agent sent that authenticated intent to the network.

What other networks are doing

American Express is not the only one trying to solve this problem. According to Worldpay, Visa, Mastercard, and American Express have all launched agentic transaction frameworks, and EMVCo is working on global specifications through a dedicated task force.

The crucial difference, however, lies in the approach. While most frameworks focus on the question "is this agent allowed to transact?", American Express is trying to answer "is this agent buying what the customer actually wanted?" — a distinction that separates authorization from intent.

The payments sector has a structural advantage that makes this distinction possible: American Express's closed-loop network, which processes both the payment and the merchant information. Open networks, such as those operated by Visa and Mastercard, have less visibility into what happens between the agent and the merchant — creating a coordination challenge that remains unresolved.

The market and trust

Worldpay found that 45% of consumers are ready to let an AI agent complete a purchase on their behalf, a number that rises to 54% among Gen Z members. But 95% also reported at least one concern about agentic commerce. Half said that having fraud protections in place would increase their trust.

This is the central tension: the convenience of autonomous agents for purchasing is undeniable, but consumer trust will not be built through marketing resources — it will be built through guarantees that if something goes wrong, someone will absorb the loss.

The question now is: when the first cases of "my agent bought what I wanted" versus "my agent bought something else" reach the courts, how will the burden of proof be determined? And who will pay the bill when an agent misinterprets an ambiguous instruction?

Sources: TechCrunch, American Express Newsroom, Worldpay

✓ Independent sources cross-checked and verified before publishing