AI agents can empty a shaky bank's accounts in hours, not days — how much warning time does a bank actually need to survive that?

This explores how fast AI agents could pull money out of a fragile bank, and whether there's a known breaking point. The short answer: the collection has no number for it. No note estimates how many hours of withdrawals, or what share of deposits leaving, would push a weak bank into failure. What it does have is a useful reframing of where the danger actually sits.

The key move comes from economist Cullen Roche. He argues that agents threaten specific banks, not the banking system as a whole. His concern is speed: agents could compress a withdrawal wave that used to take days, the kind that sank Silicon Valley Bank, into a few hours, leaving a bank with a weak balance sheet no time to raise cash or reassure depositors Could AI agents speed up runs on individual weak banks?. So the useful question isn't really "how much velocity". It's "how much response time does a weak bank need, and do agents leave it any?"

Roche and Anthony Pompliano explain why this stays a single-bank problem. When an agent sweeps your cash from one bank to a higher-paying one, the money doesn't leave the system. It just changes address. In aggregate, they argue, the effect is to reprice deposits, so banks have to compete harder on interest rates. It doesn't drain the system, especially since people will adopt these agents gradually and banks will respond. But they flag one exception: banks with concentrated depositor bases Would AI cash-sweeping agents trigger a systemic bank run?. That's the weak-bank case your question points at.

McKinsey offers the slow-motion version of the same pressure. Banks earn a lot from customers who never move their money. If 10 to 20 percent of consumers let agents sweep their cash, McKinsey estimates deposit margins could tighten by 30 to 50 basis points Will AI agents break the deposit and credit card inertia?. Here's what that suggests: before any sudden run, agents may first wear down a weak bank's profits, leaving it with a thinner cushion when a fast run does arrive. There's also evidence that agent-controlled money really does move at machine speed now. On the x402 payment network, the median time from funding an agent wallet to its first spend fell from 73 days to 18 minutes Are AI agent wallets becoming genuinely operational instead of experimental?.

The answer, then, is that the corpus describes the mechanism but not the threshold. Agents shrink the time between the first doubt and a bank failure, while slowly squeezing margins in the background. Pinning down an actual breaking point would take liquidity-stress modeling the collection doesn't contain yet.