On November 16, Federal Reserve Governor Michael Barr gave a speech in Milan on the financial stability implications of artificial intelligence. One of his warnings belongs on every bank’s agenda: AI agents can turn a bank run into something that unfolds in minutes.

The reasoning is simple. An agent managing someone’s money can spot a rate difference, or a signal about a bank’s soundness, and move funds almost instantly. If thousands of agents react to the same signal, deposit outflows can outpace the bank’s ability to manage liquidity and regulators’ ability to step in.

It is already happening

The most important part of the speech is not the hypothesis but the fact: according to Barr, major banks already report AI agents trying to act in retail customer accounts without the bank being able to verify authenticity, consent, or alignment with what the customer actually wanted.

This deserves a careful reading. These are not sophisticated attacks. Often they are tools the customer authorized, loosely, to “take care of my finances.” The bank receives a transaction that is technically valid but ambiguous where it matters: who actually decided to make it.

Several US banks responded by blocking outside agents from customer accounts. It is an understandable reaction, but it does not scale. Customers will use agents anyway, and a bank that only blocks will be left out of that relationship.

This is an architecture problem

Barr points to three lines of control: strict authentication, explicit consent verification for specific actions, and technical controls that limit what an agent can do. Translated into design decisions:

1. The agent needs its own identity. Today most agents operate with the customer’s credentials, or with a token that does not distinguish between the person and the software acting for them. A bank ready for agents must be able to answer, for every transaction: was it a person or an agent? Which agent? Who authorized it, and when?

2. Consent must be per action, not a blank check. Authorizing an agent to “manage my savings” is not the same as authorizing it to move the entire balance to another institution on a Friday night. High-impact actions need specific, verifiable, time-stamped consent.

3. Limits of authority belong in the platform, not in the agent. You cannot rely on an agent to restrain itself. Maximum amounts, allowed destinations, time windows and how fast an action can be repeated are rules the bank must enforce server-side, just as it already limits transfers today.

I would add a fourth point, given the systemic risk Barr describes: observability of aggregate behavior. If many agents start moving funds at the same time, the bank must see that as one signal, not as thousands of unrelated transactions.

What a bank can do today

There is no need to wait for a rule. The first steps are:

  • Map where automated transactions enter today: open APIs, aggregators, access with customer credentials.
  • Design an identity model that separates customer and agent, with scoped, revocable permissions.
  • Define which actions require stronger consent, and enforce it in the platform.
  • Add agent behavior to liquidity and fraud monitoring.

AI agents will become one more channel, as online banking and wallets did. Banks that design that channel on purpose will be able to offer it safely. The rest will receive it anyway, without control.


I help banks and fintechs design the identity, payments and integration architecture they need to operate with AI agents. See services.

Source: Michael S. Barr, “Some Potential Financial Stability Implications of Artificial Intelligence”, Federal Reserve Board, November 16, 2026.