Rise of Agentic AI Shopping Bots Reshapes Payments as Experts Warn Banks of Risks

Agentic commerce is moving from product recommendations to delegated checkout, turning AI agents into actors inside the payment chain. Authentication and authorization can establish that an agent is allowed to spend, but not that it selected the right item, merchant, quantity or price. Model errors, prompt manipulation and stale checkout data can therefore produce a validly authorized yet economically incorrect payment, complicating liability, refunds and dispute handling for consumers, merchants, issuers and payment networks.

Google introduced the Agent Payments Protocol, or AP2, on Sept. 16, 2025, followed by OpenAI and Stripe’s Agentic Commerce Protocol and ChatGPT Instant Checkout on Sept. 29. Visa has illustrated spending controls with limits of £30 per transaction and £100 per day. The architecture debate is now shifting beyond one-time permission toward transaction-level checks that bind the amount, merchant and cart to user intent, supported by replay protection, real-time revocation, audit trails and human review.

Payments fraud is shifting from stolen card numbers to the impersonation of legitimate users. Traditional systems often investigate only after a transaction settles or a dispute arises, when losses have already occurred. Agentic AI can continuously analyze payment behavior and risk signals, intervene autonomously during the payment process and block suspicious transactions, making it an important focus for financial institutions seeking stronger real-time fraud prevention.

Recent reporting indicates that agentic AI has already been deployed in payment workflows, where it can continually reassess risk before a transaction is completed. The technology is designed to combat rapidly accelerating scams that exploit genuine accounts or identities. However, the available event data does not identify the institutions using it or disclose the value of blocked transactions, fraud losses or the report’s publication date, making it impossible to confirm the scale or specific impact of its adoption.