Fed's Barr Warns Banks Can't Prove AI Agents Follow Orders

Barr put the question straight to an industry sprinting into AI: how do you know the agent did what the customer actually wanted? Speaking at the Singapore Fintech Festival on November 11, 2025, the Federal Reserve governor said banks will struggle to make sure agentic AI — autonomous systems that carry out multi-step tasks without checking back at every step — follows customer instructions. For a business built on authorization, audit trails and liability, that gap touches nearly everything.

Large banks are already exploring generative AI for core work including credit assessment, fraud detection and trading, Barr said. His bar for those deployments is steep: outputs need to be controlled, precise, explainable and replicable. A research summary of his remarks spelled out the downside of missing that standard — bias in lending decisions, and unintended effects such as trading algorithms that feed market volatility.

The failure mode is easy to picture. A customer tells an agent to pay their bills, and the agent has to decide which bills, from which account, in what order and by when. Every judgment call in that chain is a spot where intent and execution can drift apart.

Banks aren't waiting for regulators to catch up. Between October 13 and 20, 2026, JPMorgan Chase, Citigroup, Charles Schwab and Wells Fargo all moved to restrict third-party AI agents from logging into customer accounts, citing unauthorized transactions and murky liability for so-called rogue agents. Wells Fargo went further: terms effective November 24, 2026 shift liability for unauthorized actions taken by AI agents onto customers.

As of late 2026, dozens of banks are building their own internal agentic tools while locking outside agents out. US interagency guidance in April 2026 explicitly excluded generative and agentic AI from model risk management requirements, and the Financial Stability Board has only tried to frame what makes agentic AI distinctive — no direct rules exist. Banks are plugging the gap with internal governance and contract changes.

Next up: whether the Fed or other agencies bring agentic AI into formal risk frameworks, and whether more banks follow Wells Fargo in pushing liability onto customers. The Financial Stability Board's work could shape rules beyond the US. For fintechs building consumer AI agents, the October 2026 restrictions from four of American finance's biggest names cut off a direct route into customer accounts, at least for now.