美国运通推出新框架,确保AI代理仅执行用户明确授权的购买行为,防范越权交易风险。
Amex’s AI Agent Purchase Authorization Framework mandates three real-time checks for all agents accessing its network: biometric verification (e.g., fingerprint or face ID), dynamic per-transaction limits tied to user spending history, and an auditable, voice- or text-based intent confirmation (e.g., ‘Confirm purchase of AirPods Pro for $299?’). Twelve platforms — including Shopify and Walmart Labs — have already signed on, with coverage expected across >80% of major e-commerce AI shopping agents by Q4. For AI application founders, this is a hard signal: financial-grade trust isn’t optional — it’s table stakes. Skipping this framework risks two direct costs: chargebacks (37% loss rate per unauthorized transaction) and exclusion from bank-recommended merchant lists (a critical traffic source). Early teams should embed lightweight authorization middleware (e.g., open-source AuthAgent) into their MVP dialog flow — confirming intent *before* action — rather than retrofitting later. This satisfies compliance while organically building user trust through transparency.