Could AI agents moving your money trigger a fast bank run, like the one that sank Silicon Valley Bank?

This explores whether AI agents that move money for people could set off a fast run on one weak or concentrated bank, the way Silicon Valley Bank collapsed, without the whole banking system failing. The corpus gives a fairly clear answer: yes, the narrow risk is real, and the systemic one is probably overstated. Economist Cullen Roche, who is skeptical of AI bank-run scenarios in general, still concedes that agents could turn withdrawals that once took days into withdrawals that take hours. That threatens specific banks with weak balance sheets, even if it doesn't threaten the economy as a whole Could AI agents speed up runs on individual weak banks?.

The reason the danger is local is that money moved by agents doesn't leave the banking system. It moves from one bank to another. Pompliano and Roche argue that 'cash-sweeping' agents, which shift your balance to wherever rates are best, change what deposits cost banks without draining the total Would AI cash-sweeping agents trigger a systemic bank run?. That's reassuring at the system level, but it's exactly what an individual bank sees as a run: its own deposits vanish, even though every dollar lands somewhere else. McKinsey's analysis explains why banks are exposed in the first place. Much of their deposit and credit-card profit depends on customers not bothering to move their money. If even 10 to 20 percent of consumers hand that job to agents, deposit margins could shrink by 30 to 50 basis points Will AI agents break the deposit and credit card inertia?. A bank that relies on sleepy depositors loses its cushion before any panic starts.

The part you might not expect to matter comes from research on AI pricing agents. Independent agents, with no communication between them, tend to settle on the same behavior. They reach above-market prices just by reacting to each other's visible moves Can independent AI pricing agents collude without communicating? Do LLM pricing agents autonomously collude in repeated auctions?. Pairing a larger model with a smaller one doesn't break this pattern. The smaller one follows the leader Does model-size difference break LLM pricing collusion?. None of these studies looks at banks. Still, the mechanism carries over in a worrying way. SVB's run spread through group chats among a closely connected set of depositors. Agents built on similar models and reading the same signals could act in sync with no group chat at all. That is an inference across domains, not a finding the corpus has tested.

What would slow such a run down is friction, and the commerce notes suggest it's being designed away. One 'allow always' setting can let an agent act without checking back with its owner Do broad consent settings pose bigger risks than prompt injection attacks?. Stablecoin setups let a master agent create sub-wallets with their own spending rules at machine speed, with no intermediary approving each one Can stablecoins let agents create spending authority without approval?. Fed Governor Waller's list of what still blocks agent finance (authentication, liability, fraud) is a list of trust problems. Run speed isn't on it What blocks autonomous agents from handling purchases at scale?. One experiment points to a possible fix: a public, enforced rulebook that runs while agents operate cut collusion among LLM agents far more than written rules did Can runtime governance reduce LLM market collusion more than written rules?. Something similar, like circuit breakers on agent withdrawals, is the obvious next question. The corpus doesn't yet have direct work on it, and it has no empirical studies of agent-driven deposit flight.