American Express is introducing a new framework designed to keep AI shopping agents within the limits set by users, as the payments industry confronts the practical risks of autonomous purchasing tools moving from novelty to mainstream use. The company's approach is intended to ensure that AI systems can assist with shopping, but only execute transactions that have been clearly authorized by the cardholder. In a market where generative AI is increasingly being embedded into browsers, assistants, and retail platforms, the move signals that payments firms are no longer treating agentic commerce as a distant concept. They are preparing for it now.
Guardrails For Agents
At the center of the framework is a simple but consequential question: when an AI agent makes a purchase, how does the payment network know the user truly intended it? American Express is positioning its framework as a control layer for that problem, aiming to reduce the chance that an AI assistant buys the wrong item, exceeds a spending limit, or completes a transaction outside the customer's instructions. The company's effort is also meant to help merchants and issuers avoid a wave of disputes that could emerge if consumers begin relying on software agents to shop on their behalf.
The timing matters. AI-powered shopping tools are advancing quickly, and the industry is moving toward systems that can search, compare, and transact with minimal human intervention. That creates efficiency, but also a new category of operational risk. A mistaken click by a human shopper is one thing; an erroneous purchase made by an autonomous agent, potentially at scale and at machine speed, raises tougher questions about liability, consent, and chargebacks. American Express is effectively trying to build a permission structure before those problems become widespread.
Payments Meet Autonomy
The broader payments ecosystem has been preparing for this shift, but American Express's announcement underscores how urgent the issue has become. Traditional card rails were built around human decision-making, not software agents acting on behalf of consumers. As AI tools become more capable of completing end-to-end shopping tasks, card issuers and networks will need to distinguish between authorized delegation and unauthorized spending. That distinction is likely to become one of the defining technical and regulatory challenges of agentic commerce.
For American Express, the framework also serves a strategic purpose. The company has long marketed itself on trust, premium service, and fraud protection. By moving early on AI purchase controls, it is signaling to consumers and merchants that it intends to preserve that trust in a more automated commerce environment. The initiative could also help Amex maintain relevance as shopping behavior shifts toward AI-mediated discovery and checkout, where the payment layer may need to be more intelligent than a simple authorization prompt.
Industry analysts have warned that the rise of autonomous agents could create a new class of disputes if consumers later claim they did not approve a transaction, even when they had previously given broad permission to an assistant. That ambiguity is exactly what Amex's framework appears designed to reduce. In practice, the system may help define what counts as a valid instruction, how spending thresholds are enforced, and how transactions are logged for later review.
Trust Becomes Infrastructure
The announcement also highlights a larger truth about frontier AI: the next phase of adoption will depend less on model capability than on governance. Consumers may be willing to let AI recommend products, but they are far less likely to trust a system that can spend money without clear boundaries. That means the winners in agentic commerce may not be the companies with the most powerful models, but those that can make autonomy safe, auditable, and reversible.
American Express is not alone in facing this challenge, but its move is notable because payment networks sit at the final gate between recommendation and purchase. If they can establish reliable authorization standards, they may shape how AI commerce develops across retail, travel, and digital services. If they cannot, the result could be slower adoption, more consumer skepticism, and a rise in transaction friction just as AI shopping begins to scale.
For now, the framework is less about a single product than about setting rules for a new commercial era. As AI agents become more capable of acting on behalf of users, the payments industry will need to decide how much autonomy is acceptable, how consent is verified, and who bears responsibility when software makes a costly mistake. American Express is making a clear bet that those answers must be built into the system before the first wave of AI shopping becomes routine.