What Amex's AI agent purchases framework means for you

American Express rolled out a new framework this week that makes AI agents prove what they meant to buy before a card gets charged. The fix exists because its card network expects far more AI agent purchases without a human present to sanity check the receipt, and that's the same gap plenty of small businesses are opening up in their own n8n and Zapier flows right now.

The new piece is called the Intent Intelligence Framework. A Create Intent API logs what the agent is being asked to buy: the merchant, the budget, the product details. The cardholder confirms that intent once, and Amex issues an Intent ID and a Proof of Intent Token tied to a payment credential, so the agent never handles the actual card number. If an agent asked to buy running shoes from one store checks out at a different one instead, or blows past the dollar cap it was given, the system is built to catch the mismatch before the charge clears.

Why AI agent purchases are suddenly everyone's problem

Most small businesses don't have a card network building this for them yet, but plenty already have an agent with purchasing power. A reorder flow that hits a supplier's API when stock drops. An ad platform that reallocates budget on its own. A bill-pay automation that approves a vendor invoice without anyone opening it first. None of those are hypothetical. They're the same n8n, Zapier, and Make patterns already live in shops like this one, just one step short of touching money with no stop sign in place.

What Amex actually built, in plain terms

Three things make the framework worth copying even if you never touch Amex's API. First, the intent gets written down before the purchase happens, not reconstructed afterward from a receipt. Second, the agent gets a token scoped to that one purchase instead of a standing card number it can reuse. Third, a mismatch between the stated intent and the actual transaction triggers a block, not a note for later review. Dispute resolution and fraud review both get easier when the system already knows what was supposed to happen.

Four guardrails to add to your own agent workflows

1. 1

   Set a hard cap, not a soft one

   Give every purchasing-capable automation a dollar limit per transaction and a separate limit per month, enforced inside the workflow step itself, so it doesn't depend on anyone remembering the policy doc.

2. 2

   Log the intent before the action

   Write what the agent is about to buy and why to a sheet, a database, or a Slack channel before the API call fires, not after the charge lands.

3. 3

   Require a second approval above the cap

   Anything over the limit should route to a human for a yes or no, with the request and the reason attached, before the order goes out.

4. 4

   Review the log every week

   A five-minute weekly scan of what agents actually bought catches drift early, long before it shows up as a surprise on the statement.

If an agent in your shop already holds a card number, an API key with spend attached, or approval rights on a vendor invoice, that's worth a look before it's worth a headline. Happy to walk through what that workflow looks like today and where the cap belongs.