Would AI cash-sweeping agents trigger a systemic bank run?

Responding to Apollo chief economist Torsten Slok's warning that agents like "Muse" could "sweep household cash automatically into accounts paying 3.3% to 5.0%, instead of the 0.1% national average on checking accounts" and trigger a system-wide bank run, newsletter author Anthony Pompliano argues "AI agents are not going to take down the financial system." He quotes economist Cullen Roche's rebuttal at length: "when cash moves from a checking account at Bank A to a savings account at Bank B, Bank A's deposits fall and Bank B's deposits rise by the same amount. Total deposits in the banking system don't change... The money never leaves the system. It just changes addresses."

Roche's mechanism rests on two claims: banks don't lend out deposits ("loans create deposits"), and deposits leave the banking system only through a few channels — "physical cash withdrawals, payments to the Treasury..., or money that ends up at the Fed's reverse repo facility." A bot reallocating a saver's cash to a higher rate isn't one of those channels, so the deposit base is repriced, not depleted. Pompliano adds that adoption itself is gradual — "consumers adopt new technology in waves," and banks facing a competitive threat will raise rates, acquire the fintechs that built the sweep products, or capture the flows themselves, since "Robinhood's cash management feature is powered by JPMorgan Chase, Goldman Sachs, and Wells Fargo."

This complicates the adoption-speed assumption behind How willing are people to let AI control their finances? and Does naming a payments brand increase trust in AI transactions?: both surveys document a trust gap that slows how fast people hand banking control to an agent, which is exactly the waves, not overnight adoption curve Pompliano leans on to defuse Slok's static-world scenario. Where those notes measure consumer hesitation directly, this excerpt argues the hesitation (plus bank repricing) is what prevents the hesitation from mattering at systemic scale even once agents do act.

The excerpt is a Substack rebuttal essay, not a modeled or empirical study — it offers no deposit-flow data, no estimate of how fast agent adoption will actually proceed, and no accounting for banks whose deposit base is geographically or sector-concentrated rather than diversified like the examples cited. It also only addresses the aggregate, system-wide case; see Could AI agents speed up runs on individual weak banks? for the narrower risk the same source concedes.