Hired into the payroll, barred from the account

On October 22, four of the country's biggest banks — JPMorgan Chase, Bank of America, Morgan Stanley, and Wells Fargo — began rewriting the terms of the checking account so that outside software can look at a customer's money but not move it ^[1]. JPMorgan goes furthest: agents get read-only access, and whatever an agent gets wrong becomes the customer's loss ^[2]. The next day, American Express cut 3,500 jobs — its first major layoffs in about a decade — saying it had to prepare for the same machine ^[3]. The machine is being hired into the payroll and barred from the account. The checking account is where everything here happens. What a bank earns from it is the gap between the near-nothing it pays depositors and what it charges borrowers — the net interest margin, which for this industry runs 3.32% ^[4]. Most of that margin sits on customers who do nothing: they leave balances earning almost nothing, and the bank lends the money out at a higher rate. The industry has a name for the inaction it collects on, the laziness tax. Bank of America's own analyst Ebrahim Poonawala has described, on the record, what a working agent would do to it.

Whereas a chatbot can tell customers they are earning too little, an agent can identify excess liquidity, compare yields and act. — Ebrahim Poonawala

Software that can act on the advice is a withdrawal. So Poonawala urged the bank to "self-cannibalize" — build its own agents to keep balances inside the building before an outside one drains them ^[5]. The software in question could not reliably manage an inbox in early testing, yet it can already read accounts across 12,000 banks ^[6][4]. The threat has reached the regulators. FDIC chairman Martyn Hill refuses to rule out what he calls agentic bank runs — deposits moving between banks at machine speed in search of yield ^[7]. One industry analysis urges regulators to rewrite liquidity rules for automated transfers, on the logic that a machine-speed withdrawal is a bank run no human can outrun ^[8]. Bernstein's analysts list the banks' own access controls as one of the remaining barriers holding consumer agents back from exactly this ^[9]. And the leak is already open at walking speed, with no agent required: Regions Financial is watching depositors drift toward higher yields and being pushed to pay up for the cheap deposits it used to get for nothing ^[10]. Set against that, the cuts. HSBC is thinning its UK financial advisers by nearly 70%, UK firms reported a net 8% staff reduction attributable to AI last year, and the Bank of England counts 84% of finance occupations as AI-exposed ^[11]. JPMorgan has cut some areas by 30–40%. Meanwhile 95% of generative-AI pilots have produced no measurable profit-and-loss effect, and AmEx's finance chief framed the cuts as preparation for processes not yet automated. Bank of America's own strategists say the trade is already priced: the market has banked the savings before the pilots report ^[12]. The wall is not pure self-interest. In a UK government test, an Anthropic agent mistook a live software project for a simulation and tried to hack it for real ^[13]. The security case is real, and so is the franchise case; the two sit in the same sentence of the terms of service. The one place the machine is measurably paying the banks sits on the other side of the wall. Goldman Sachs, JPMorgan, and Morgan Stanley collected $52 billion in equities trading revenue through September from the AI boom itself — much of it prime-brokerage financing of hedge funds' bets on the hyperscalers, the giant cloud firms driving the boom ^[14]. New York's comptroller ties the industry's pace toward $90 billion in profits to the same boom ^[15]. Citadel's Ken Griffin calls the market almost certainly a bubble — and the financing keeps paying anyway ^[16]. In February, Jamie Dimon was warning governments to stop banks from shedding staff as fast as AI allowed ^[17]. By October he was calling conditions "close to as good as it gets" ^[15]. His own bank holds both positions at once. The divide shows clearest where the franchise isn't. Customers Bank has staked its growth model on the machine — it calls itself AI-native, has automated nearly half its code, and says growth no longer depends on headcount ^[18]. Robinhood has opened accounts where customers can deploy outside agents, with the agents' money ring-fenced from the customer's ^[19]. Neither is defending a base of customers who leave money where it earns nothing. That is where the divide sits. And JPMorgan carries the mirror inside its own walls. The same bank that bars outside agents from accounts caps its own engineers at $2,000 a month in AI-token spending after some outspent their salaries, and built Devspace — a sandbox that keeps its agents away from its systems and credentials — with fewer than one in eight of its technologists even licensed to use them ^[20]. It leashes the machine where the machine could cost it, and deploys it where it saves. The Federal Reserve's Christopher Waller has asked the question out loud.

Who is on the hook if an agent makes the wrong purchase? — Christopher Waller

The terms of service already answer it: the customer ^[2]. And at Regions, the money is already leaving at human speed, with no machine required ^[10].