AI Agents Could Create a Modern Deposit Bank Run

By Jay Kemp, President of Provoke Management.

Bank runs used to be loud, with lines out the door and a name in the news by noon. That is the version every risk model was built to catch. In his latest Banking Insights Video, "The Agentic Bank Run," Jim Marous argues it is the wrong one to be watching for. His opening is almost quiet. "The next bank run will not look like a traditional run," he says. "There will be no crowd and no panic, because no one will be knocking on your branch's door."

Nobody has to run anymore. The money can simply leave.

Marous highlights the shift to who is making the decision. When a customer's AI agent can find a better rate and transfer the funds overnight, deposits drain gradually in small amounts that fail to trigger alarms. He is careful to emphasize that this is not a story about panic. "The real exposure is dynamic repricing," he says, "the money you thought was loyal becoming measurable, comparable, and movable before your model catches up." Loyal, measurable, movable. Those three words rarely share a balance sheet.

"For decades, your customer's inconvenience was your strongest retention tool." — Jim Marous, Banking Insights Video: "The Agentic Bank Run"

Stickiness Was Never Loyalty

Most of the video revolves around an assumption that almost no deposit strategy openly acknowledges: that customers won't bother to move their money. Marous has a name for that assumption. He also has a more direct term. "Stickiness has a simpler name," he says. "Friction."

Anyone who has switched banks knows the frustration he refers to: the direct deposit that must be redirected, the autopay that gets broken, the afternoon lost dealing with it. That inconvenience kept people with the same bank, and banks quietly counted the result as loyalty. "That friction was quietly doing the work we credited to loyalty," Marous says.

Here's the uncomfortable part. The friction is disappearing. An agent handles the paperwork, checks the market, and turns a switch into a tap. The customer might still press approve, but the effort that once protected the balance is gone. What is left with it is the share of loyalty that was only ever an inconvenience in disguise. And inconvenience does not survive automation.

A Number Most Banks Cannot See

So what replaces it? Marous introduces a metric he calls agent-exposed deposits, with a figure worth considering. Citing first-quarter FDIC data via Neil Stanley at CorePoint, he notes that roughly 85% of deposits have immediate availability. "The vast majority of your balances are not locked into anything," he says. "They can move the second something decides to move them."

Agent-exposed deposits are the portion of available money that software can detect, compare, and move before any human signals reach the bank. Traditional reports do not capture this. A seven-year-old checking account and a balance alert indicating one rate change look identical in a core system. They are not the same. When an agent is monitoring, they could be completely different.

The figures involved here are significant. McKinsey has estimated that shifting just 5% to 10% of checking balances into higher-yield accounts could eliminate 20% or more of the industry's deposit profits. Forrester predicts that machine-initiated traffic to bank websites will increase by 40% this year, while human visits decline by 20%. Additionally, this June, for the first time, automated traffic surpassed human traffic across the internet, now growing at roughly eight times the rate. The infrastructure is being built whether banks are aware of it or not.

The Machine Banks Already Built

The most powerful moment in the episode is also the most self-aware. Marous asks bankers to remember the sweep account, the tool the industry built decades ago to move idle cash overnight into the best available return, no human needed. "An agent is a sweep account that the customer controls," he says, "targeting the entire market and comparing every available rate every day."

Read that again. The same optimization logic banks developed to benefit themselves is now directed outward, working for the customer and against the bank's margins.

Marous calls this agentic attrition, the faster and louder cousin of the silent attrition he has written about before. The balance is lost the moment a bank stops offering the best rate, and the bank finds out later, if at all. His phrase for it sticks: "a smaller number on Monday than you saw on Friday," with no cancellation notice and no exit survey.

What Banks Can Actually Do

His prescriptions are unglamorous, and none of them involve the app.

Stop pricing as if people are lazy, and model deposits as though every dollar gets re-shopped daily, because a growing share of them will. Spend relationship capital where an agent cannot follow, on the hard human moments software has no business touching. Then measure exposure directly by sorting the deposit base into three buckets: relationship-protected, rate-sensitive, and agent-exposed. Treating them as interchangeable, Marous warns, is "a measurement error."

Then he widens the lens, and this part is worth not skipping. Deposits are only the first thing to go because they are the easiest. After that, it might be the card, then the loan, and eventually the advice itself. "The question is no longer whether your customer will keep an account with you," he says, "but whether they will allow you to make a single decision on their behalf."

Banking spent a decade making its digital experience irresistible to people. The uncomfortable truth in his argument is that people may not be making many decisions much longer. The competition now includes the software between a bank and its customer. And that software does not care who wins.

Watch "The Agentic Bank Run" on YouTube, or listen on Apple Podcasts.