Somewhere in a Blockworks co-founder's timeline, a number is doing a lot of quiet work: 0.1%. That's the savings rate an Apollo economist cited as the floor American depositors are sitting on, against 3% to 5% available in money market funds. His thesis โ AI agents will close that gap automatically, and in doing so, trigger something like a bank run.
The headline writes itself. The mechanism doesn't. Because the gap between 0.1% and 3% has existed for over two years, and households have been walking across it the entire time without an agent in sight.
Numbers first, narrative second. The Fed's reverse repo facility, which peaked near $2.5 trillion in late 2022, has drained to near nothing. Over the same window, US money market fund assets climbed from roughly $5 trillion toward $7 trillion. That migration happened in the years before any consumer-grade agent could log into a bank portal, read a rate table, and submit a transfer.
So the driver was never AI. It was arithmetic. What agents change is friction โ the manual account opening, the manual transfer, the manual monitoring that kept the migration rate below its theoretical maximum. Drop that friction toward zero and you don't create a new phenomenon; you accelerate an old one and compress its timing. And the receiving end has been prepared for years: money market fund reform, tokenized Treasury products, and regulated stablecoin frameworks all exist precisely to absorb cash that no longer wants to sit in a checking account.
That distinction matters for anyone pricing bank equities or on-chain dollar supply right now. Acceleration is a different risk than ignition.
Here's where the popular version of this story starts drifting from the data. The 0.1% figure is real, but it isn't the average US savings rate. FDIC-reported averages sat closer to 0.4%โ0.6% at the peak of this hiking cycle. The 0.1% number maps much more cleanly onto the legacy base savings products at the largest money-center banks โ accounts that have paid 0.01% for most of their existence.
That's the widest spread in the system. It's also the stickiest pool in the system, and the reason is mundane: payroll direct deposit, branch relationships, bundled mortgages, business checking. The pool with the biggest headline gap is the hardest one for an agent to drain.
The pool that actually bleeds is elsewhere. Regional and community banks hold deposit bases weighted toward rate-insensitive retail savings, with narrower wholesale funding channels and net interest margins far more sensitive to funding cost. They are the absorbers of this shift. Money-center banks shrug. Money market funds, short-duration Treasuries, tokenized T-bills, and stablecoins are the recipients.
Then there's the part the original framing skips entirely: synchronized behavior. Traditional deposit flight is emotional โ a headline, a rumour, a line outside a branch. Agent-driven flight is rational and simultaneous. If a handful of agent platforms read the same rate tables and apply the same ranking logic, their outflow decisions correlate. That's the same failure mode as the 2010 Flash Crash, transplanted from the trading book to the liability side of the balance sheet โ and it's the only genuinely new variable AI introduces here.
I've spent enough time inside reentrancy audits to recognize the shape. The exploit was never the new line of code. It was the assumption that every actor would behave independently. Agent concentration breaks that assumption. If three platforms control the default routing of household cash, their comparison logic becomes, functionally, a price-coordination mechanism โ an antitrust question wearing a fintech costume. Modularity isn't the freedom to scale; it's the freedom to correlate.
Which brings me to the capability question nobody is auditing. I pulled apart contract logic for small ERC-20 projects in 2023 and found a reentrancy hole that would have drained $50,000 from fifteen lines of Solidity. Financial agents deserve the same scrutiny. Right now the best consumer agents can handle browser-level tasks: log in, read a rate page, compare, initiate a transfer. That's a proof-of-concept edging toward production, not production. Financial-grade reliability needs idempotency, error rollback, audit trails, and failure modes that don't move money twice. I have not seen a public evaluation suite that scores any of this.
The bottleneck isn't the model either. It's authorization. Cross-institution money movement requires open banking APIs and account aggregation, and the US still has no mandatory open banking framework. Banks can throttle, restrict, or simply not expose the endpoints. So the realistic near-term product is recommendation plus one-click confirmation, not autonomous execution. That single design choice deflates most of the instant-run narrative โ humans still sit in the loop, and humans are slow.
One more layer if the destination is stablecoins rather than money market funds: reserve transparency risk, redemption procyclicality, and a cross-border supervisory vacuum stack on top of the rate bet. That is no longer an AI risk at all. It is a digital-dollar risk, and the original framing never separates the two.
Then the rate cycle. The entire thesis rests on a spread that only exists when policy rates are high. If the Fed cuts, money market yields compress toward deposit rates and the arbitrage that powers the agent narrative evaporates. The biggest vulnerability in this story is that it needs a specific macro regime to stay alive. Anyone building a thesis on AI-driven deposit flight is implicitly long rates.
Track it properly instead of trading the headline. The Fed's H.8 release, the H.4.1, and ICI's weekly money market data give you the actual flow. RRP balances tell you whether cash is parked or circulating. Log those series for two quarters before you price a structural shift.
Call it what it is. This isn't a bank run. A run requires panic, and panic requires the asset side of the balance sheet to be impaired โ loans marked down, securities sold at a loss, capital consumed. None of that is happening here. What's moving is the funding structure. The right term is deposit disintermediation, or funding migration, and the difference isn't academic: a run is a solvency event, migration is a repricing event. Confusing the two leads to the wrong policy response and the wrong trades.
Follow the incentives behind the framing. Apollo runs an alternative asset manager with large insurance annuity liabilities and heavy fixed-income and credit allocations. A world where household cash abandons bank deposits in favour of higher-yielding vehicles is a world where credit provision migrates toward private lenders. That's not a conspiracy; it's a positioning. Meanwhile the Web3 outlet retelling the story has its own beam: if the destination is stablecoins and tokenized Treasuries, the on-chain dollar supply expands, and crypto infrastructure gets repriced upward. Nobody in that chain disclosed the interest.
There's also a regulator-shaped hole. Banks carry capital requirements, liquidity coverage ratios, and deposit insurance costs. Money market funds and stablecoins operate under a different, lighter regime. Migration of this kind is partly a transfer of activity from heavily supervised entities to lightly supervised ones โ regulatory arbitrage wearing the mask of consumer empowerment.
And the quietest risk of all: households moving from FDIC-insured deposits into uninsured funds, with no explicit risk conversation. The yield is visible; the insurance gap is not. Code is law, but vigilance is the price of entry โ and nobody is handing retail users the entry fee schedule.
If this thesis is right, the questions that matter are unanswered. Who carries liability when an agent misjudges and moves a household's cash in the wrong direction โ the user, the platform, or the bank? Is there a circuit breaker, the way equities have one? Does an agent provider holding the default routing for millions of accounts count as systemically important infrastructure? And what happens when every agent decides to leave at the same second? The next watch item isn't a headline โ it's the ICI weekly print, and a Federal Reserve that has to decide whether it is supervising models or money.