Why Amex wants AI shopping agents to prove user intent

The problem: payment credentials do not explain intent

Traditional card payments assume a human is present near checkout. A customer browses, chooses an item, confirms the cart and pays.

Agentic commerce changes that sequence. A customer may give an AI agent a broad instruction — such as buying a product when it is restocked, finding the best price under a cap, or reordering household staples — and the agent may act later, across merchants, under changing conditions.²

That creates a trust gap. A valid payment credential can prove that an account can pay, but it does not prove that an agent bought the right product, stayed within budget, used an approved merchant or acted before the customer’s instruction expired. TechCrunch reported that Amex is trying to address that gap by carrying the customer’s purchase instructions through the transaction lifecycle.³

The risk is not only fraud. Large language models and agent systems are probabilistic. They can misread instructions, fill carts incorrectly or act on stale context. Amex executives have pointed to a simple example: if a user asks for white Nike sneakers in size nine but the agent attempts to buy Adidas shoes, the payment system should be able to detect that mismatch before checkout is completed.³

How the proposed flow works

Amex’s proposed flow starts when an agent collects purchase instructions from the customer. Those instructions may be narrow, such as approval of a specific cart, or broader, such as permission to buy a product later if certain conditions are met.² The agent may ask clarifying questions until it has enough structured information to present back to the user.

The Create Intent API is the mechanism Amex proposes for sharing the confirmed purchase intent with the issuer or relevant payment participant. The payload can include data such as purchase summary, maximum amount, intent expiry, merchant information, item-level conditions, maximum number of transactions, order context and purchase criteria.² In product terms, it converts a natural-language shopping instruction into a structured authorization boundary.

After the customer selects a payment instrument, they authenticate and sign off on the intent. Amex’s technical explainer says sign-off can be performed through a passkey, or delegated by the issuer to another party, such as an agent or wallet provider, depending on the implementation.² Once authentication is complete, the system creates two artifacts: an Intent ID and a proof-of-intent token.²

The Intent ID is a unique identifier for the purchase-intent record. It acts as the primary key for later lifecycle actions, such as updating, canceling or validating the intent.² The proof-of-intent token is described by Amex as a JSON Web Token that represents the high-level constraints the customer approved.²

That token can then be used to obtain a scoped digital payment credential. The credential can be constrained by the user’s approved intent, including merchant, expiration date and transaction count, and may be configured for single use or multiple uses across one or more merchants.⁵ At checkout, the agent uses this scoped credential instead of the customer’s actual card number.³

What the token is meant to prove

The proof-of-intent token is not simply another payment token. It is meant to show that the user authenticated and approved a defined set of purchase constraints before the agent attempted to pay.² That distinction is important for agentic commerce architecture.

A payment credential answers one question: can this account be charged? A proof-of-intent token is designed to answer another: did the user authorize this agent-mediated purchase under these conditions? If the purchase request falls outside those conditions, the issuer or payment network can treat that as a risk signal or block the transaction, depending on the implementation.⁴

This gives issuers earlier visibility into transaction context. Amex’s technical explanation says issuers should receive purchase-intent information so they can assess risk, authorize delegated execution and scope payment credentials against approved intent.² Because issuers often manage cardholder disputes, that same record can later help evaluate whether a challenged transaction matched what the cardholder approved.²

Why lifecycle controls matter

Agentic purchases may not be single, immediate events. A user could ask an agent to buy concert tickets when prices fall, purchase several items from different merchants, or reorder supplies over time. Amex’s framework therefore treats intent as something that may need to be created, updated or canceled during its lifecycle.²

Digital Transactions reported that the framework is designed to keep the customer in control, including the ability to edit or cancel intent through Amex or through an agent, depending on the implementation.⁴ That matters for product builders because the user interface cannot end at initial consent. Users will need ways to review active intents, revoke outdated ones and understand which agents have authority to act.

The same lifecycle model could help merchants and fraud teams. Digital Transactions noted that the Intent ID and proof-of-intent token can be used later in the purchase journey to validate that a purchase request aligns with what the customer authorized, while also supporting authorization decisions, fraud prevention and dispute resolution.⁴

Why Amex is moving before agentic commerce scales

Amex says it has not yet seen a rise in disputes from agentic commerce, but it is trying to act before autonomous purchase volume grows.³ That timing is significant. The payments industry has learned from earlier waves of digital commerce that retrofitting trust, liability and dispute controls after adoption can be expensive and fragmented.

Consumer interest is already visible. TechCrunch reported Amex survey data showing that 66% of surveyed U.S. consumers were likely to use AI agents to research the best price for a product or service, while 61% were likely to use them for recommendations over the next 12 months.³ TokenPost reported additional survey results showing that consumers were much more comfortable letting agents buy everyday, low-cost items than high-value goods such as electronics and appliances.⁵

That pattern suggests agentic commerce may start with constrained, repeatable purchases before expanding to higher-risk categories. For payment networks, that creates a window to define primitives — intent records, scoped credentials, agent identity, cart context and dispute evidence — before autonomous checkout becomes widely embedded in AI assistants, wallets, merchant sites and marketplaces.

What builders should take from the framework

The main design lesson is that consent needs to be structured. A checkbox or conversational “yes” may not be enough when an agent can act later, across merchants or under conditional instructions. Builders should think in terms of intent objects: structured records that capture what the user approved, when they approved it, which agent is acting, what limits apply and when authority expires.⁶

Second, payment credentials should be scoped to the job. The Amex model uses the proof-of-intent token to obtain a digital credential that can be limited by the customer’s approved constraints.⁵ That approach reduces the need to expose durable card credentials to agents and gives networks and issuers more ways to contain errors.

Third, dispute readiness should be built into the transaction flow. If the system records the approved merchant, maximum amount, cart details, conditions and expiration before checkout, it creates evidence that can be evaluated later if the customer, merchant, issuer or agent platform disagrees about what happened.² The AI Insider described the framework as supporting fraud prevention and dispute resolution by generating an Intent ID and proof-of-intent token after customer approval.⁶

Finally, interoperability will be a key question. Amex wants the framework to help inform broader industry standards for AI-powered transactions.¹ DigitalToday reported that Amex aims to maintain purchase instructions from payment through post-processing stages, with the goal of making the framework an industry standard for AI-based transactions.⁷ Whether the market converges around Amex’s model, other network-led standards or wallet- and platform-led protocols, the direction is clear: autonomous agents need more than credentials. They need verifiable boundaries.

The broader direction for agentic payments

The Amex framework reflects a broader industry move from “can this payment be made?” to “was this autonomous payment authorized under the user’s stated conditions?” Fraud-prevention vendors are also moving into the category; Digital Transactions reported that Forter introduced an Agentic Commerce Engine to help merchants verify agentic shoppers and manage agentic transaction cycles.⁴

For AI product teams, the implication is that shopping agents will need to integrate with payment infrastructure more deeply than checkout autofill. Agents may need to capture structured intent, support user authentication, request scoped credentials, preserve logs, expose cancellation controls and pass context to issuers, wallets and merchants.

For payments companies, the challenge is to make those controls strong enough to prevent unauthorized purchases but flexible enough for natural agent behavior. An agent that can only execute a rigid cart is less autonomous. An agent that can spend broadly without verifiable limits is unlikely to be trusted.

Intent intelligence sits between those extremes: it gives the machine room to act, but only inside a boundary the user approved.