As autonomous software agents move from experimental labs into everyday shopping, the finance world is scrambling to plug a critical hole: who pays when an AI bot makes a mistake? American Express answered that question on October 6, 2026 by publishing its Amex Business Playbook for Agentic Commerce a roadmap that adds liability shields to the technical plumbing introduced earlier in the year with the ACE Developer Kit.
The playbook centers on two pillars. First, the Agent Purchase Protection initiative promises to reimburse eligible card members and merchants if a registered AI agent triggers an unintended charge. Although the program is still in the announcement stage, Amex stresses that the fine print will define what qualifies as a “registered” agent and under which circumstances the protection kicks in. The second pillar is a nascent Merchant AI Advisory Council designed to give retailers a voice in shaping network standards as they grapple with three divergent postures—welcome, browse-only, or outright block of agent-driven traffic.
Why the safeguard matters now
Industry surveys reveal a stark mismatch between merchant optimism and consumer confidence. An internal Trendex poll shows 83% of businesses are eager to tap AI agents for sales, yet only 15% of shoppers trust a bot with high-value purchases. By contrast, 66% will let an agent compare prices and 61% are open to receiving product recommendations within the next twelve months. Moreover, 93% of merchants believe the AI provider should foot the bill for erroneous transactions. Without a credible backstop, that optimism risks turning into spreadsheet-level speculation.
Inside the Intent Intelligence Framework
The technical centerpiece of Amex’s response is the Intent Intelligence Framework. It begins with a Create Intent API that captures a shopper’s exact purchase instructions—brand, price ceiling, size, and any ancillary constraints—before the agent proceeds. The user then authenticates themselves, generating a unique Intent ID and a Proof of Intent Token. This token supplies a secure, disposable digital credential that the agent uses at checkout, keeping the real card number hidden.
Because the intent travels with the transaction from start to finish, Amex can intervene early. For example, if a user asks an agent to buy a white Nike sneaker in size 9, but the agent mistakenly submits a request for an Adidas model, the system flags the mismatch and aborts the purchase. The same mechanism can feed fraud-prevention engines and streamline dispute resolution, creating a verifiable audit trail of which agent acted, when, and under what constraints.
Future-proofing autonomous buying
Hannah Bozian, Vice President of Agentic Partnerships and Strategy at Amex, notes that today most agents request explicit consent before each spend, but the company anticipates a shift toward more autonomous re-ordering of staple goods—think toilet paper or coffee beans—where only price and quantity matter. In those scenarios, the Intent Intelligence Framework acts as a safety net, ensuring the transaction aligns with the buyer’s pre-approved parameters without constant manual confirmation.
The rollout coincides with Amex’s upcoming appearance at TechCrunch Disrupt 2026 on October 13, where Bozian will detail how the framework can become an industry standard. Meanwhile, the firm is already in talks with payment-network partners to embed the new agentic tokens across the broader ecosystem. While disputes from agentic commerce have not yet surged, Amex’s pre-emptive approach aims to keep that metric flat as the market scales.
