Did the Agent Do What You Asked? What Amex's Intent Intelligence Framework Means for AI Commerce

When a person tells an AI agent to “buy me a pair of white size 9 sneakers under $120,” a surprising amount of trust is packed into that one sentence. The agent has to interpret it correctly, pick the right merchant, stay inside the budget, and pay without anyone double-checking. Large language models are nondeterministic, and TechCrunch notes that this behavior can lead agents to make unauthorized transactions.

On October 13, 2026, American Express put forward an answer: the Intent Intelligence Framework, a way to capture and authenticate what a customer actually asked for before an agent spends a cent. This post explains how it works, what it signals for agentic commerce, and what any business building or buying AI agents should take from it.

What American Express Announced

According to the Business Wire release carried by Yahoo Finance, the framework records a customer’s purchase instructions and limits, such as the merchant and maximum budget, before an AI agent buys anything. The customer confirms that intent through identity authentication. The framework then issues an Intent ID and a Proof of Intent Token.

The framework is publicly available and delivered through Amex’s Agentic Commerce Experiences (ACE) Developer Kit. TechCrunch reports that Amex introduced the ACE toolkit in April 2026, and that Intent Intelligence is now part of those services.

The Three Parties in Every Agentic Purchase

Hannah Bozian, Amex’s VP of Agentic Partnerships and Strategy, summed up the change this way: “These new transactions will have a consumer, a merchant, and an agent.” That third party is the problem. Traditional payment systems were built to verify a person and a merchant. They have no native way to ask whether the agent did what the person wanted.

How Intent Intelligence Works

Digital Transactions describes four elements: creating and confirming intents, establishing proof of intent, securing a payment credential, and applying the intent ID and proof to the purchase flow.

Step 1: Capture the Intent

A Create Intent API records the customer’s instructions and constraints. Amex says collection may vary by AI model, and it aims to support all agentic commerce models rather than a single agent design.

Step 2: Authenticate and Approve

The customer verifies their identity and approves the captured intent. This is the human checkpoint, and it happens up front rather than at the moment of payment. It is a different idea from a last-second confirmation dialog. For more on keeping humans in control, see our post on the “Approve” button.

Step 3: Issue a Token Instead of a Card Number

The approved intent produces an Intent ID and a Proof of Intent Token holding high-level constraints. The agent uses the token to obtain a digital payment credential, which it presents at checkout in place of the real card number. The agent never holds the card itself.

Step 4: Check Purchases Against the Intent

The Intent ID and token can support fraud prevention and dispute resolution, and an authorization log shows which agent acted and when. In TechCrunch’s example, if a customer asks for size 9 white Nike sneakers and the agent orders Adidas, Amex could block the transaction. Bozian says the approach “pulls intent earlier into the transaction flow” than other implementations do.

Why This Matters Beyond Payments

The Gap Between “Authorized” and “Intended”

A card transaction can be fully authorized and still be wrong. Classic fraud tools ask whether the payer is who they claim to be. Agentic commerce adds a harder question: did the agent do what was asked? Proof of intent gives issuers, merchants, and customers a shared record to answer it, which matters most when a purchase is disputed.

Control Without Friction

Amex expects agents to act more independently for low-risk, repeat purchases such as household staples where only price and size matter, with the framework checking each order, according to TechCrunch. Digital Transactions adds that customers can edit or cancel an intent through Amex or through their agent. The design goal is autonomy inside a boundary the customer set once.

Early Days, Real Volume

Amex told TechCrunch it has not yet seen a rise in disputes tied to agentic commerce, and says it wants to be proactive. Digital Transactions reports Amex sees “thousands” of agentic-commerce transactions weekly across its network integrations. Consumer appetite also looks real: in an Amex Trendex report, 66% of surveyed U.S. consumers said they were likely to use AI agents to research the best price, and 61% said they would use them for product recommendations over the next 12 months.

A Bid for a Standard

Amex hopes the framework becomes an industry standard. Digital Transactions says it is built for Amex’s model as issuer, acquirer, and network but adaptable to others, and the Business Wire release says Amex is engaging with EMVCo to support secure card-based agentic transactions. Expect competing approaches from other networks and wallets, so treat this as a strong signal of direction rather than a settled standard.

Lessons for Businesses Building or Buying AI Agents

The pattern applies well beyond shopping. Any agent that spends money, sends messages, or changes records needs the same three things: a recorded intent, a scoped credential, and an audit trail.

1. Separate Intent From Execution

Capture what the user wants as a structured, signed record before the agent acts. Compare every action against it. This is the principle behind the failures we covered in what loopjacking means for “human-in-the-loop” guardrails: approving one thing and getting another.

2. Never Hand Agents the Master Credential

Swapping the card number for a token that carries limits mirrors good non-human identity practice: short-lived, narrowly scoped credentials tied to a specific agent.

3. Log Which Agent Did What, and When

Amex’s authorization log reflects a broader need. When something goes wrong, you need an answer to the questions raised in what happens when your AI agent makes a mistake. Without an audit trail, liability is a guessing game.

4. Ask Your Payment and Platform Vendors About Intent

If you sell online, ask your processor how it will verify agent-originated orders. If you build agents, ask whether your platform supports intent records and scoped payment credentials.

Conclusion: Actionable Takeaways

Amex’s framework will not be the last word on agentic payments, but it frames the right question. The next trust problem is not “is this payer real?” but “did the agent do what the human wanted?” To prepare:

• Inventory every place your agents can spend money or commit your company to something.
• Require a recorded, user-approved intent for each of them.
• Issue scoped, revocable credentials instead of sharing master keys or card numbers.
• Keep an audit log tying every action to a specific agent and intent.
• Watch the standards work, including EMVCo, before locking into one vendor’s approach.

Frequently Asked Questions

What is the Amex Intent Intelligence Framework?

It is a framework published on October 13, 2026 that captures and authenticates a customer’s purchase instructions before an AI agent completes a purchase. It issues an Intent ID and a Proof of Intent Token that the agent uses to obtain a payment credential.

Does the AI agent get my actual card number?

No. According to Amex, the token lets the agent obtain a digital payment credential to use at checkout instead of the real card number.

Can a purchase be blocked if the agent gets it wrong?

Amex says it could. In its example, an agent that orders Adidas shoes when the customer asked for size 9 white Nike sneakers could have the transaction blocked.

Is the framework available to developers now?

Yes. The framework is publicly available through the Amex Agentic Commerce Experiences (ACE) Developer Kit, though Amex says it is still developing further ACE integrations.

Is this an industry standard yet?

No. Amex hopes it becomes one and says it is engaging with EMVCo, but other payment companies may take different approaches.

Does this only matter for online shopping?

No. Any agent that can spend money or make commitments benefits from recorded intent, scoped credentials, and audit logs.

Sources

• TechCrunch: Amex introduces a framework to prevent AI agents from making unauthorized purchases - main news report with quotes from Amex
• Business Wire via Yahoo Finance: American Express Publishes Intent Intelligence Framework - Amex’s press release (company-authored)
• Digital Transactions: AmEx’s Next Step for Agentic Commerce Is a Framework - trade coverage of the four framework elements
• The AI Insider: American Express Introduces Intent Intelligence Framework - additional coverage of the launch
• American Express: Agentic Commerce Experiences - Amex’s ACE developer page
• arXiv: Secure Autonomous Agent Payments - research paper on verifying authenticity and intent in agent payments