Amex Intent Intelligence puts agent purchase constraints on the issuer's clock

When the buyer is a model, the dangerous moment is not the witty product pitch. It is the second the agent commits money while the issuer still thinks a human is holding the card.

On 13 October 2026, American Express published its Intent Intelligence framework as a core service of the Amex Agentic Commerce Experiences (ACE) Developer Kit. The company wants the industry to treat authenticated purchase intent as a control that travels with the transaction. The useful question for builders is narrower: can the issuer see the customer's constraints before the agent spends?

Create Intent turns instructions into an Intent ID and a Proof of Intent Token

The Create Intent API is meant to catch what the customer actually authorised: merchant preferences, a maximum amount, an expiry, a maximum number of transactions, product preferences, and even dependencies between items so one purchase can wait on another. How those instructions are collected may vary by agent. Amex says the framework is meant to stay flexible across those models.

The customer then authenticates and signs off on the captured intent. That produces an Intent ID, a unique handle for the instruction set, and a Proof of Intent Token that represents the high-level constraints. In Amex's technical framing, that proof is a JSON Web Token. The token can be used to obtain a scoped digital payment credential for checkout instead of the customer's actual card number. Later steps in fraud checks, authorisation, and disputes can reuse the same Intent ID and token as evidence of what was approved.

Amex draws a practical line between delegated instructions and approval of a specific cart. Each pattern leaves different room for agent error. The Nike-size-9 example Hannah Bozian gave TechCrunch is the blunt version: if the customer authorised white Nike trainers and the agent submits Adidas, the issuer-side check is supposed to have something concrete to reject against.

October's publish details a service ACE already listed in April

Intent Intelligence was not invented for Disrupt week. Amex's ACE page already listed Intent Intelligence among specifications available as of 14 April 2026, alongside Account Enablement and Payment Credentials. The 13 October press release, amplified the same morning by TechCrunch, is the detailed public framework and the partner-briefing moment, with Bozian on Disrupt's Industry Stage.

Luke Gebb, Amex's Head of Global Innovation, framed the peg as trust and dispute reduction: authorised intent before the agent transacts should give issuers more confidence and cut costly chargebacks across the rail. Bozian told TechCrunch the point of pulling intent earlier is to mitigate risk before it occurs, including a future where agents reorder staples under price and size constraints without a human reviewing every cart.

Specs for Intent Intelligence are available; Agent Registration is not

On maturity, treat this as a published control surface with incomplete neighbours. Amex says Account Enablement, Intent Intelligence, and Payment Credentials specifications are available. Agent Registration and Cart Context are still under development. The company also says it is actively developing ACE Kit integrations and is in talks with partners across the payments ecosystem.

That is a usable grade for product and protocol teams reading the kit. It is not evidence that every merchant, wallet, or card member can already run the full loop in production. Amex also points to standards work with EMVCo for card-based agentic transactions and support for the x402 Foundation's internet-native payments effort. Those are commitments, not interchange proofs.

Prefer issuer-visible intent when the agent may reorder without a cart review

If your agent still stops for an explicit cart approval every time, many existing consent patterns may be enough for a pilot. The Amex model earns attention when the agent is allowed to act under standing constraints: household reorders, travel windows, corporate spend envelopes, or any flow where the issuer needs a rejectable object before authorisation.

In those cases, prefer a design that carries merchant, amount, expiry, and transaction-count limits into something the issuer can inspect. Keep cart-only approval for experiments where a human remains in the loop at the last step. Bozian's Adidas miss and the staples-reorder scenario are product illustrations, not independent incident data, but they show the class of failure Intent Intelligence is aimed at.

A published framework is not universal live coverage

Amex told TechCrunch it has not yet seen agentic-commerce disputes rise, and that the controls are proactive. That is the company's claim. Trendex figures Amex shared about consumers' willingness to use agents for price research and recommendations are survey colour, not proof the rail is ready.

ACE materials also preview Amex Agent Purchase Protection for eligible customers when a registered agent sends authenticated purchase intent and then errs. That protection is described as coming soon, with registered-agent lists still to follow. Subjective instructions remain a hard edge: intents that lean on words like "best" or "really nice" are difficult to verify cleanly. A happy-path demo that buys one item does not prove the reject paths work across agent, issuer, and merchant.

Run the reject paths before you trust the brochure

Agent commerce will be judged on who can stop a wrong purchase in time. Read the Intent Intelligence flow, then test a permitted buy followed by a wrong merchant, an over-budget total, an expired intent, and a cancellation while the agent is still running. Keep the brochure language for the press kit until those rejects leave a clear record.