American Express Is Building the Payment Rails for AI Agents — With Strings Attached

American Express has spent more than a century building trust between cardholders and merchants. Now it is attempting something considerably more ambitious: building the infrastructure that lets artificial intelligence agents spend money on your behalf, with guardrails baked into the transaction itself. The company's Agentic Commerce Experiences developer kit, known internally as ACE, represents one of the most concrete attempts yet to answer a question that the payments industry has been quietly dreading — when an AI makes a purchase, who is actually responsible for it?

The ACE system works by issuing what Amex calls "intent contracts" paired with single-use payment tokens. The idea is straightforward in principle: before an AI agent can complete a transaction, it must operate within a defined set of parameters that the user has pre-authorized. A single-use token then executes that specific transaction and expires. Nothing bleeds over. The agent cannot go rogue and book a first-class flight when you authorized economy. The constraint is structural, not just a matter of policy.

This is a meaningful departure from how most agentic commerce discussions have unfolded. Much of the industry conversation has centered on interoperability protocols — getting different AI systems and payment networks to speak the same language. Google's Agent Pay Protocol, known as AP2, is the most prominent example of this approach, and Amex is already a participant in that effort. But AP2 and similar frameworks largely address the plumbing. ACE is attempting to address something harder: the question of consent, scope, and auditability at the moment of transaction.

The challenge Amex is grappling with is not purely technical. It is fundamentally a trust architecture problem, and it sits at the intersection of behavioral economics, contract law, and systems design. When a human swipes a card, there is a clear principal — the cardholder — making a deliberate choice. When an AI agent executes a transaction on that cardholder's behalf, the chain of accountability becomes murky in ways that existing fraud frameworks were never designed to handle.

Intent contracts attempt to encode that chain directly into the payment event. Rather than relying on post-hoc dispute resolution — the chargeback system that has long been the financial industry's blunt instrument for handling contested transactions — ACE tries to make the authorization conditions machine-readable and enforceable before money moves. This is closer in spirit to a smart contract than to a traditional card authorization, even if it operates within Amex's proprietary network rather than on a blockchain.

That proprietary constraint is worth examining carefully. ACE currently functions only within the American Express network, which means its protections apply to a subset of the transactions an AI agent might attempt to make. For consumers who use multiple cards, or whose AI assistants operate across different payment rails, the coverage is incomplete. An agent authorized through ACE could still execute unconstrained transactions elsewhere. The system solves the problem it can reach, but the problem is larger than any single network.

There is also the question of the black box. Even with intent contracts and single-use tokens, the reasoning an AI agent uses to select a specific merchant, price point, or product remains largely opaque. A user might authorize an agent to "book a hotel in Chicago under $200 a night" and the agent might comply with the letter of that instruction while making choices — about cancellation policies, loyalty programs, or data-sharing agreements — that the user would have made differently. The token enforces the financial boundary. It does not enforce the judgment.

The deeper systemic consequence here involves what happens to consumer financial literacy and oversight as agentic commerce scales. If AI agents handle an increasing share of routine purchasing decisions, the feedback loop that normally connects spending to awareness begins to weaken. People learn their own financial habits partly through the friction of making purchases. Remove that friction entirely and you remove one of the mechanisms by which people develop preferences, notice patterns, and catch errors.

Amex's ACE framework is, in this sense, a bet that structured constraints can substitute for that friction without eliminating its protective function. Whether that bet pays off depends heavily on how transparent the intent contract layer becomes to ordinary users — not just to developers integrating the API. A system that is auditable in principle but opaque in practice offers weaker guarantees than its architecture implies.

The payments industry is moving toward agentic commerce whether or not the trust infrastructure is ready. Amex is at least asking the right questions early. The more interesting test will come when those intent contracts meet edge cases, disputes, and the inevitable gap between what a user meant to authorize and what an AI understood them to authorize. That gap has always existed in commerce. It has just never been automated at scale before.