Amex Now Guards AI Agent Purchases: SMB Guide

American Express just answered a question most small businesses haven’t asked yet: what happens when an AI agent goes shopping with your card? This month Amex rolled out its Intent Intelligence Framework inside the Agentic Commerce Experiences (ACE) Developer Kit, and with it Amex Agent Purchase Protection — an industry-first commitment to back purchases made by registered AI agents. If you accept cards, spend on cards, or are wiring AI agents into either side of that equation, this is the plumbing of the next commerce cycle arriving early.

Here’s the plain-English version, plus what we’d do about it this quarter.

What Amex actually built

Amex’s framing is simple: agentic transactions have three parties — a consumer, a merchant, and an agent. Traditional card rails were designed for two. So Amex is extending its closed-loop network with a few new pieces:

• The Create Intent API. When a customer tells their AI assistant to buy something, the agent captures the purchase instructions — merchant name, maximum budget, product preferences, constraints — and sends that authenticated intent to Amex before the transaction happens. The card network now has a machine-readable record of what the human actually authorized.
• Agent Purchase Protection. If a registered agent makes a purchase and Amex received the authenticated intent, the purchase gets Amex’s backing — the same dispute and fraud posture card members expect from human-made purchases. An agent buying the wrong thing stops being a gray zone.
• Agentic tokens. Amex payment credentials wrapped for agent use, so the agent never holds a raw card number.
• End-to-end visibility. Every step of the commerce lifecycle — intent, authorization, purchase — is logged. When something goes sideways, there’s an audit trail of what was authorized, by which agent, and when.

Notably, Amex says it hasn’t seen a spike in agentic-commerce disputes yet. This is a proactive build — the company wants the framework to become an industry standard before the transaction volume arrives.

Why small businesses should care now

Two reasons, one for each side of the counter.

As a buyer: your business already runs software that acts semi-autonomously — auto-replenishment, bidding tools, subscription management. The next generation of those tools is agentic, like the always-on assistants we covered in OpenAI’s Dots. The moment an agent can comparison-shop and check out on its own, “who authorized this $900 charge” becomes a board-level question for a five-person shop. Intent frameworks are how that question gets a clean answer.

As a merchant: agent-driven purchases change your funnel. When an AI agent buys on behalf of a customer, it doesn’t see your beautiful hero image — it reads your product schema, your price, your return policy, and your checkout API. If your data is messy, you’re invisible to the agent. And as networks like Amex formalize authenticated-intent flows, merchants who support them cleanly will win the agent channel the same way early SEO adopters won search.

The failure mode this prevents

We’ve written before about what happens when AI agents go off script. Commerce is the highest-stakes version of that problem. A content agent that hallucinates a blog paragraph is embarrassing; a purchasing agent that hallucinates a bulk order is a cash-flow event. Without an intent layer, the dispute lands in a gray zone: the customer says “I didn’t buy that,” the merchant says “the transaction was authenticated,” and nobody can prove what the human actually told the machine.

Amex’s answer — capture intent with explicit constraints up front, then verify the purchase against it — is the same pattern we use when building supervised automations for clients: define the envelope (budget, merchant, quantity), log the authorization, and let the agent move fast inside hard limits. It’s good architecture, now backed by a card network’s fraud machinery.

What to actually do this quarter

1. Audit your own automations. Anything that spends money — ad platforms, auto-ordering, cloud budgets — should have explicit spend caps and a human-visible log. If it doesn’t, that’s your first project, before any agent touches your card.
2. Clean your product data. Structured pricing, availability, and return-policy data on your site is how purchasing agents will “see” you. Treat schema markup as agent-facing sales collateral.
3. Watch the standards race. Amex wants the Intent Intelligence Framework to be an industry standard, and it’s already in talks with partners across the payments ecosystem. Visa, Mastercard, and PayPal won’t sit this out. Expect a common shape to emerge within a few quarters — don’t hardcode to any single vendor’s API yet.
4. Test one narrow agent purchase flow. Pick one low-risk, high-frequency purchase (office supplies is the classic) and pilot it with explicit caps. You’ll learn more from one bounded pilot than a year of reading announcements.

The bottom line

Agentic commerce is arriving the way mobile payments did: a trickle of infrastructure announcements first, then a sudden expectation. Amex just moved the timeline up. Small businesses that get their spend controls and product data in order now will be sellers and buyers in the agent economy on their own terms — instead of discovering it through a disputed charge.

We build supervised AI agent setups for small businesses — with hard spend limits, audit logs, and guardrails designed for exactly this kind of change. See how at squidcircle.com/squidbot.