01
THE SPREAD IS ALREADY BIG ENOUGH TO NOTICE.
Apollo Chief Economist Torsten Slok flagged agentic assistants that could sweep household cash from checking accounts averaging about 0.1% into accounts paying roughly 3.3% to 5.0%. His warning: widespread optimization could strip banks of the cheap deposits they rely on. The agent doesn't need a crisis. Yield alone is a permanent reason to leave.
02
$23 TRILLION IS SITTING IN LOW-INTEREST CHECKING.
McKinsey estimates about $23 trillion of global deposits sit in checking accounts earning little interest. Its modeling says moving only 5% to 10% of checking balances into higher-yield accounts could reduce total banking profits from deposits by 20% or more. Not every household needs to automate. A minority does the damage.
03
$42 BILLION LEFT SVB IN ONE DAY.
The Federal Reserve's review of Silicon Valley Bank says depositors withdrew more than $40 billion on March 9, 2023, and management expected more than $100 billion of additional outflows the next day. The Fed called the withdrawal speed unprecedented — coordinated behavior across a concentrated network, amplified by social media. SVB was not an AI bank run. It is proof a run can compress into hours with no AI at all.
04
THE DALLAS FED IS ALREADY MODELING THE FRICTION DISAPPEARING.
A 2026 Dallas Fed analysis says deposit stickiness partly depends on frictions that slow reallocation — and that instant settlement and agentic AI could theoretically let yield-seeking depositors switch banks almost instantly without direct action from the account holder. This is now a central-bank white paper, not a sci-fi pitch.
05
AI CAN ALREADY DO REAL CASH-MANAGEMENT WORK IN A SIMULATION.
Bank of Canada researchers tested a general-purpose AI agent on high-value payment-system liquidity management. The agent maintained buffers, prioritized payments, and adjusted recommendations under liquidity shocks — without domain-specific training. That does not prove consumer agents will trigger runs. It shows cash-management judgment is already within reach of general-purpose models.