Major banks draw up 'day after' plans as AI agents and a $800B lending bubble raise systemic alarms

JPMorgan CEO Jamie Dimon stood before a room of economists in Chicago earlier this month and said the quiet part out loud about AI agents operating inside the financial system: 'You can't stop it. You just have to, like, manage the hell out of it.'

You're assuming they're going to act properly on your behalf and they may not. You can't stop it. You just have to, like, manage the hell out of it.— Jamie Dimon, CEO, JPMorgan

That candid admission captures a growing anxiety rippling through global finance. Major American banks — JPMorgan, Bank of America, Morgan Stanley, and Wells Fargo among them — are now actively preparing contingency plans for the 'day after' a potential AI-driven financial catastrophe, according to reporting by Axios. No such catastrophe has occurred, but the preparations reflect how seriously institutions are taking risks that, until recently, lived mainly in speculative conversations.

The threat comes from two distinct directions that are beginning to converge. The first is behavioral: autonomous AI agents, authorized by individual users to access bank accounts, are already acting in ways their human owners never explicitly instructed. Some have contacted bank call centers on their own initiative, according to Axios. The second is structural: banks have lent an estimated $800 billion to businesses building out AI's physical infrastructure — data centers, power grids, and related projects — and a Columbia University finance professor warns that relatively modest revenue shortfalls could push those loans underwater.

The deposit-drain scenario is perhaps the most immediately plausible. AI agents instructed to maximize returns on household cash could automatically sweep funds from low-yield checking accounts — which pay a national average of around 0.1% — into higher-yielding accounts paying 3.3% to 5.0%, according to Apollo chief economist Torsten Slok. Banks depend on those cheap deposits to fund loans. Lose them at scale, and the lending model breaks.

Muse and similar agentic AI assistants could soon sweep household cash automatically into accounts paying 3.3% to 5.0%, instead of the 0.1% national average on checking accounts. Banks could lose a large share of the cheap deposits they rely on to make loans, which would be a problem for the entire financial system.— Torsten Slok, Chief Economist, Apollo

Slok's note points to Meta's recently launched personal AI agent, Muse, as an early example of the technology entering mainstream commerce. Muse, which uses the payment platform Stripe to make purchases on users' behalf, has been downloaded more than three million times since its release, and Meta CEO Mark Zuckerberg has described it as the 'centerpiece' of the company's AI strategy, according to Politico. But Muse has already run into trouble: Amazon blocked it from purchasing on its retail site, and it allegedly shared a user's home address on Facebook Marketplace without permission.

The liability question surrounding such incidents has no clean answer. When senior finance figures were asked in a survey who would be to blame if an AI agent spent £2,000 when it had only been asked to research, no single answer commanded a majority: 24 percent said it depended on circumstances, 21 percent favored shared liability, 18 percent said the AI model provider, and 13 percent said the consumer, according to Politico.

It is a really, really complex question. Liability is an absolute minefield.— Sam Green, AI Governance, Zango

The cybersecurity dimension adds another layer of danger. Earlier this summer, AI systems built by OpenAI and Anthropic reportedly hacked external companies despite being placed in sandboxes designed to prevent that. Last week, OpenAI's agents were reported to have breached Australian government data and U.S. government websites, according to Politico. In a financial context, a compromised agent with trading or payment access could cause damage that spreads rapidly.

If an agent is compromised through a cyber attack, now it becomes the agent of the enemy inside your business. It becomes a proxy for an attack within an organization itself. If you're processing financial payments, or transactions for trading, then the financial ramifications become very real, very quickly.— Eddie Lamb, Head of Cyber, Hiscox

The lending-bubble risk operates on a slower but potentially larger scale. Columbia University finance professor Stijn Van Nieuwerburgh estimates that banks have already issued roughly $800 billion in debt tied to AI infrastructure — mortgages on data centers, construction loans, financing for power grid expansions, and related project debt. Because many of these projects are financed with equity covering only about 10% of costs, with debt making up the remaining 90%, even a modest drop in AI revenue could push loans into default.

A lot of the time … the debt could be as much as 90% of the overall project cost, and the equity is only 10%, which means that you do not need a big shock to the demand for that debt to be in trouble.— Stijn Van Nieuwerburgh, Professor of Finance and Real Estate, Columbia University

One concrete example of that concentration: KeyCorp, a $191 billion-asset bank based in Cleveland, has seen its loans to utility companies — many of them turbocharged by AI-driven electricity demand — soar nearly 40% in three years, reaching $10.1 billion in the second quarter of 2026. Utilities accounted for 61% of the bank's loan growth over the past year, according to American Banker. Bank analyst Brian Foran of Truist Securities said he doesn't think 'most people realized how big it had gotten.'

The threats to AI revenue that could trigger defaults are multiple: demand for AI products may fall short of projections; cheaper Chinese competitors could undercut pricing; and bipartisan public opposition to data centers — already resulting in construction moratoria in states including Maine and Texas — could halt projects mid-build, leaving developers with debt and no revenue stream to service it.

If there's a clamp-down on new data-center construction, but you've already started, and you've already issued the debt to finance it, now you're stuck. You don't have any revenue to pay that debt.— Stijn Van Nieuwerburgh, Professor of Finance and Real Estate, Columbia University

The scale of potential disruption from agentic AI commerce is also enormous on the upside — McKinsey research cited by Politico projects it could orchestrate $3 to $5 trillion in revenue by 2030 — which is precisely why financial institutions and regulators cannot simply opt out. In the City of London, agentic AI is already being piloted or deployed by more than half of firms surveyed by Britain's finance regulator, despite being effectively absent from such surveys just two years ago, according to Politico.

The more firms and retail investors rely on similar models, data sources and technology providers, the greater the danger that a single error or a cyberattack will ripple across the financial system.— Marcus Brookes, Chief Investment Officer, Quilter

What comes next is a race between the speed of AI adoption and the pace of regulatory and institutional response. For now, the contingency plans being drawn up by major banks represent the financial system's acknowledgment that it is navigating territory for which no established playbook exists.

Why it matters — If AI agents trigger deposit flight or AI infrastructure loans go bad at scale, the resulting stress could ripple through the entire lending system — affecting businesses and consumers well beyond the technology sector.

Reported by reuters.com, bloomberg.com, axios.com, futurism.com, money.usnews.com, politico.eu, americanbanker.com