Could AI Assistants Trigger a Fast Bank Squeeze?

Apollo warned on September 27, 2026 that agentic AI could automate deposit flight, in a Daily Spark insight titled 'Is an Agentic Bank Run Coming?' Apollo Daily Spark

The note centers on Muse and similar agentic AI assistants. It describes a near future in which those assistants could automatically sweep household cash into accounts paying 3.3% to 5.0% instead of the 0.1% national average. The framing is prospective. It is a warning about capability and incentive, not a claim that such sweeping has occurred at scale.

The item was listed in Apollo's Insights & News with a date of September 27, 2026. Torsten Slok is Apollo's Chief Economist.

The broader context here is why banks count on inertia. Retail balances have low deposit beta, meaning banks lift the rate they pay much more slowly than market rates move. That stickiness supports net interest margin, the gap between what a bank earns on loans and pays on deposits, and deposit franchise value. It lets managers assign behavioral duration to non-maturing deposits, treating checking money as long-term funding. Compress search cost and execution lag to near zero and that assumption weakens. Money can migrate with no solvency worry. Rate arbitrage alone reprices the book.

Looking at liquidity management, the shift is from scattered attrition to correlated outflow. Manual rate shopping is dispersed across time, attention and hassle. Agent instructions are programmatic and can fire on identical triggers for many people at once. That changes intraday liquidity demands, contingent funding plans, and outflow assumptions under LCR, the Liquidity Coverage Ratio rule for holding ready assets, and internal stress tests. Wholesale backstops reprice quickly when demand is synchronized.

Looking through a pricing lens, the pressure falls on low-beta funding models. Institutions using noninterest-bearing and low-rate retail balances to fund longer-duration assets face faster pass-through. Deposit beta converges toward wholesale beta, the market cost of borrowing. Hedging, buffer sizing and term funding built for sluggish repricing would need recalibration. The adjustment runs through funding cost before credit.

In my view, the binding constraint is delegation, not model capability. Sweep authority, authentication, error liability, counterparty exposure limits and receiving-bank inflow controls still govern whether instructions settle. Households hold operating cash for immediacy, bill pay certainty and overdraft avoidance, not yield alone. Standing authority, exception handling and revocation rights will determine realized flow. Agents lower the cost of finding and executing the spread. They do not rewrite account agreements.

For risk teams watching for early stress, adoption curves and correlation matter more than rate gaps alone. Permissioned, always-on sweeping creates regime change in outflow velocity. Unpermissioned or narrowly scoped assistants leave inertia largely intact. Surveillance should focus on sweep disclosures, transfer frequency, failure rates and concentration at recipient institutions. Those mechanics will move before headline deposit totals do.

The bottom line here is friction was the moat. Automation removes it.