Federal Reserve Governor Michael Barr raised a timing problem for deposit-taking institutions on November 16: a consumer’s AI agent could detect a better rate or a sign of trouble and move money before the institution’s usual monitoring and response can catch up. In his speech on AI and financial stability in Milan, Barr described a possible acceleration of deposit outflows, not a measured run at a credit union or a new Federal Reserve rule.
The distinction matters for a credit union that has no plans to deploy its own agent. The software that might initiate a transfer could be acting on a member’s behalf, outside the credit union. And several members acting through the same service could respond to the same rate or risk signal at the same time. Neither possibility establishes how many members will use such a tool or how much money would move.
Two questions for one transfer
Barr said major banks have reported AI agents trying to take actions in retail accounts without the banks being able to verify authenticity, consent and alignment with customer instructions. He did not name credit unions or identify a measured effect. One useful internal question is what evidence would establish that a particular transfer was authorized: who or what issued it, what scope the member approved, and what limits the institution can apply. A separate question is what treasury can see once authorized transfers leave. An authorized batch can affect liquidity without being a fraud case.
The NCUA examiner guide already discusses stress scenarios such as higher-than-expected deposit runoff and testing funding needs at different points in an event. Barr’s remarks offer a scenario to consider within that existing work. They do not require a particular alerting interval, transfer limit or agent policy.
Measure the time from movement to decision
For a tabletop exercise, run the same hypothetical withdrawal amount over two periods: spread over the day, then concentrated in a short window outside normal staffing hours. Use payment and share-account activity your institution can actually see. Measure how long it takes for the activity to appear, an alert to reach someone who can act, and treasury to decide whether it needs funds. Then test what information that person has about available balances and settlement deadlines.
Keep the exercise separate from member-level restrictions. A spike may reflect legitimate transfers, fraud, or a data problem; a scenario alone is not evidence to block transfers. Give fraud and treasury separate decision paths, with a clear handoff if an incident affects both. The useful result is not a prediction of an AI-driven run. It is evidence about whether the credit union can see a fast outflow, assign it and act while it still has options.