OCC to bank examiners: AI is just another risk, no separate limit

The Office of the Comptroller of the Currency told its bank examiners Wednesday to stop treating artificial intelligence as its own kind of risk. Banks won't need a separate risk limit for AI either. Instead, examiners will judge AI based on how a bank uses it, how big the bank is, and how complex its operations are. The OCC calls the approach technology-neutral and says it won't play favorites with any one type of tech.

The regulator also rolled out what it calls its first framework for supervising AI at large national banks. Examiners should worry less about how fancy an AI tool is and more about the damage it could cause if it fails. A complex tool locked inside a bank might be less dangerous than a simple AI agent that can start payments on its own. When risk runs higher, the OCC says banks may need to cap what AI can do without a human, or keep a person closely in the loop.

The new framework closes a regulatory gap. Back in April, the OCC, the Federal Reserve and the FDIC refreshed their rules on managing model risk, but they left out generative and agentic AI, calling it new and fast-changing. The agencies had said they planned to ask the public for input on how banks use AI.

Several big U.S. banks didn't wait for the OCC's guidance. Citigroup, Bank of America, JPMorgan Chase and Charles Schwab have already blocked or limited outside AI agents from reaching customer accounts, fearing those agents could shift unused cash into higher-interest accounts. On JPMorgan's third-quarter earnings call, CEO Jamie Dimon said the bank will limit agent access to client accounts, calling rogue agents a danger and warning there could be thousands of them. Bank of America CEO Brian Moynihan has said he wants a human involved before money leaves an account.

Regulators are still wrestling with who gets blamed when an automated payment fails. Bank of England Deputy Governor Sarah Breeden, who wants tougher AI safeguards, spoke at the European Central Bank's forum in Sintra, Portugal. She suggested every autonomous AI trading tool should have a kill switch so people can stop it immediately if it sends prices diving.