Shopping online is starting to lose its human hands. In agentic commerce, an AI agent searches, compares and pays for things on a person’s behalf. On Tuesday, American Express published a framework meant to keep those agents inside the lines.
New here? I publish one agentic AI governance post every weekday. Subscribe to the blog and it lands in your inbox the moment it goes live.
The framework is called Intent Intelligence. It makes an agent prove what the card member approved before any money moves. The news lands just as consultants forecast trillions of dollars in agent-led sales. It also arrives while banks, stores and lawyers still argue about one thing. Who pays when a machine buys the wrong item?
A shopper that acts
A chatbot suggests a product. An agent buys it. That is the whole difference, and it is a big one.
The person states a goal, like white sneakers in size 9 for under $150. Then the agent does the rest. It finds the shoes, checks the price, fills the cart and pays. McKinsey calls the shift a move from separate steps to one continuous flow driven by intent.
Meanwhile, the payment industry spent the past year building pipes for this. OpenAI and Stripe built a checkout protocol for agents. Google published one too. Visa and Mastercard each launched their own agent programs. So agentic commerce now has rails. What it still lacked was proof of what the shopper actually wanted.
The gap in the old rules
Networks wrote card rules for two parties, a buyer and a seller. Agents add a third. “You have a customer, a merchant, and now you’ve got an agent,” an Amex vice president, told Digital Transactions.
That third party breaks the dispute process. The buyer did let the agent shop. But the buyer may still reject what it picked. The law firm Goodwin warned in June that card disputes could fill with simple buyer’s remorse. And language models are not fully predictable, so a misread request becomes a real charge.
What Amex built
Intent Intelligence sits inside a developer kit that Amex launched in April. Amex built the kit for the companies that make shopping agents.
The design works as a chain of four checks, according to Amex’s technical post.
- First, the agent records the limits, such as the store, the top price and an end date.
- Next, the card member signs off, often with a passkey.
- Then Amex issues an intent number and a signed token.
- Last, that token unlocks a card credential that only works inside those limits.
So the agent never sees the real card number. Every call also runs over locked, two-way encrypted links, an earlier Amex engineering post says. The signed pieces expire fast, so a thief cannot reuse them.
How one purchase plays out
Take the sneaker case, as per TechCrunch, The shopper asks for Nike. The agent drifts and tries to buy Adidas. Because the approved intent reached the card issuer first, Amex can stop that sale before it clears.
That timing is the real news. Most systems check intent at checkout, or only later in a dispute. Amex checks it before the agent spends. The record then follows the purchase, so it can settle an argument months later.
Who pays when the agent is wrong
Amex pairs the framework with a promise it calls Agent Purchase Protection. Under it, Amex plans to cover charges where a registered agent strays from what the member approved, Amex’s program page says.
Still, the fine print matters. Amex must have registered the agent. The member must also try a return first. Vague wishes, like “the best one,” may not count. And Amex says the full terms will come later, so it is not live coverage yet.
More on agent intent and control
- Intent files: My plain guide to writing down what an agent may do before it acts.
- Govern the harness: Why I govern the harness, not the model, at every stage.
- The agent harness: Where the term came from and how each vendor uses it.
- Subagents: How one agent hands work to another and where control slips.
What it means for trade
The money at stake in agentic commerce is large. McKinsey estimates agents could steer up to $1 trillion in U.S. retail sales by 2030, and $3 trillion to $5 trillion worldwide. Those are forecasts, not results, but they explain the rush.
For stores, the buyer changes. McKinsey advises retailers to treat agents almost as a new class of customer. That means prices, stock and return rules a machine can read. But many are not ready. In an Amex survey of 502 business leaders, released on October 6, 83 percent saw an opportunity. Only 41 percent were testing or using it. For banks, the goal is trust at scale. If issuers can see intent, they can approve more good sales and block more bad ones.
For global trade, standards decide the pace. Amex chairs a card-industry working group on agent payments at EMVCo, Digital Transactions reports. Mastercard and Google published a rival spec in March. If the networks agree on one shape, an agent in Dallas can buy from a shop in Tokyo under the same rules.
My take
I think this is the most useful idea in agentic commerce this year. Not because of the tokens, but because of where the check sits.
Limits that live only in a prompt are easy to talk around. In Anthropic’s Project Vend, a Claude model ran a small office shop. Staff talked it into discounts and free goods. Nothing outside the model said no. Amex moves the no out of the model and into the payment rail.
That is the lesson I carry from my work in regulated loan origination, too. In short, write the intent down, sign it, and check it before the agent acts. Anthropic’s own guide to building agents warns that small errors compound. External checks stop the compounding.
Instructions in, results out was IT. Intent in, outcomes out is agentic AI. Amex just put a signature on the intent. I would copy the pattern now and wait for outside data before I believe the results.
Would you let an agent buy within a signed intent, or would you still want the final click? Tell me in the comments.
Was this post useful? One click tells me what to write next.
Subscribe to my LinkedIn Newsletter
- Agentic AI Architect: control design for enterprise agents.
- Agentic AI Case Studies: deployment evidence, one teardown at a time.
- Agentic AI P&L: cost, payback, and risk in a CFO’s voice.
- Agentic AI SDLC: the lifecycle that builds the agents.
- Subscribe to the Blog: I publish a new post every weekday on one agentic AI governance question. Enter your email and each post arrives in your inbox the moment it goes live.
© Dr. Harish Kotadia, Ph.D., All Rights Reserved, 2026
Dr. Harish Kotadia, Ph.D., is an Enterprise AI Architect with 20+ years of IT consulting experience serving Fortune 100 clients, specializing in agentic AI systems built on Anthropic Claude, AWS Bedrock, and Google Vertex AI.
Disclaimer: This blog post is based on publicly available academic publications, vendor documentation, open standards, and news items from reputed media sources linked above. This post is intended for educational purposes, to help the enterprise agentic AI community build a shared vocabulary from public, authoritative sources.
Views and opinions expressed here are my own and do not represent those of any employer or client, past or present. The analysis presented is my independent interpretation of the published sources linked above and does not constitute legal, financial, or consulting advice of any kind.