European Sovereign Infrastructure
Editorial illustration of a neoclassical bank facade drawn in fine gold line art on a deep navy background, with glowing gold particle streams rushing through its columns and fanning out into a synchronised swarm of teal arrow-like shapes, suggesting AI agents moving money at machine speed. Faint circuit traces run behind the building.
COVER FEATURE

Could AI agents start the next bank run?

In March 2023, Silicon Valley Bank lost around $42 billion in deposits in a single day. That was the first smartphone bank run. The question now is what the next one looks like when the people moving the money are no longer people at all.

By John Wroath, Editor

Published 29 September 2026

In March 2023, Silicon Valley Bank lost around $42 billion in deposits in a single day. Nobody queued outside a branch. The run happened on phones, in group chats and on social media, and it was faster than anything regulators had planned for.

That was the first smartphone bank run. The question I keep coming back to is what the next one looks like when the people moving the money are no longer people at all.

AI agents are moving from answering questions to taking actions. They book travel, reconcile invoices and, increasingly, manage cash. Corporate treasury teams are piloting agents that sweep balances, rebalance across banks and react to market signals without waiting for a human. That is genuinely useful. It also changes one of the oldest dynamics in finance: the bank run.

So, could AI agents start a run on the banks? My honest answer is yes, but probably not in the way the headline suggests.

What agents change about a run

A bank run is a coordination problem. Each depositor asks one question: will others pull their money before I do? The bank fails not because everyone acts irrationally, but because everyone acts rationally at the same time. Agents change three things about that.

Speed. SVB showed that a run can move from rumour to collapse in about 36 hours. An agent does not need to wake up, read the news, call a colleague and log in to online banking. It can detect a signal and move funds in seconds. The window a bank or central bank has to respond shrinks from hours to minutes.

Herding. This is the one that worries me most. If thousands of treasury agents run on a handful of foundation models, trained on similar data and reading the same feeds, they will tend to reach the same conclusion at the same moment. Human panic is messy and staggered. Machine caution can be perfectly synchronised. Correlated decisions are exactly what breaks a bank.

Human panic is messy and staggered. Machine caution can be perfectly synchronised.

John Wroath

Bad inputs. Agents act on what they read. A misread earnings note, a fake press release or a deepfake video of a CEO could trigger withdrawals before any human has checked whether it is true. Social media already spreads rumours fast. Agents could act on them faster.

This is not just my speculation. The Bank of England's Financial Policy Committee has repeatedly flagged AI-driven herding as a financial stability concern, and its July 2026 Financial Stability Report warned of increased risk linked to rapid progress in frontier AI.

Why it is not inevitable

It would be easy to stop there and write a scary post. But there are real counterweights, and it is worth being honest about them.

Deposit insurance still works. In the EU, deposits are guaranteed up to €100,000. In the UK the limit is now £120,000. For most households and small businesses, there is simply no reason to flee. A sensible agent acting for them should know that, and should not move insured money in a panic.

Agents do not yet hold the keys. Very few people or companies give an agent unrestricted authority to move large sums. Payment limits, dual approvals and bank-side fraud controls all sit in the way. For now, most agents recommend and humans approve.

Banks can automate their defences too. Central banks now run standing liquidity facilities that can be tapped quickly. Banks can apply circuit breakers, throttle unusual withdrawal patterns and communicate in real time. If the attack is automated, the defence can be as well.

Agents might be calmer than we are. An agent that properly weighs deposit insurance, a bank's capital position and the source of a rumour could actually hold steady where a frightened human would not. Well-designed agents could dampen panic rather than amplify it.

Where the real risk sits

Here is where I think the popular framing gets it wrong. The danger is not your personal finance app emptying your current account because it read a scary tweet. It is uninsured corporate money.

Most of SVB's deposits were above the insurance limit. They belonged to startups, venture funds and tech companies, many of them connected through the same investors and the same WhatsApp groups. That is what made the run so fast. When one fund told its portfolio to move cash, dozens followed.

Now picture the same concentration, but with treasury agents instead of CFOs. Corporate treasury is exactly where agent adoption is likely to arrive first. The money is large, the tasks are repetitive and the business case for automation is strong. These are also the balances with the most reason to run, because nobody is guaranteeing them.

So my honest view is this. AI agents probably will not invent a bank run out of nothing. What they could do is take a genuine weakness, the kind that exists at some bank somewhere most years, and turn a manageable wobble into a collapse before anyone has time to intervene.

What should happen now

None of this is an argument against agents. It is an argument for designing them, and the system around them, with runs in mind.

For companies deploying treasury agents: build in a human pause for large or unusual moves. Make agents check the source of a signal before acting on it. Ask your vendor how their model behaves in a stress scenario, and whether every other customer's agent will do the same thing at the same moment.

For banks: assume the next run will be machine-speed. Stress tests built around a few days of outflows need a scenario measured in minutes. Monitoring should spot correlated agent behaviour, not just large individual withdrawals.

For regulators: herding risk from shared models is a concentration risk, much like reliance on a few cloud providers. It deserves the same scrutiny. There may also be a case for standards on how agents handle deposit decisions, such as mandatory cooling-off periods for moves triggered by news rather than by business need.

Trust at machine speed

Banking has always run on trust, and trust has always had a speed limit set by how fast people could talk to each other. The branch queue became the phone call, then the group chat. Each step made runs faster.

Agents remove the speed limit entirely. That does not make a crisis inevitable. It does mean the institutions, the rules and the agents themselves need to be built for a world where confidence can drain in minutes.

The question is no longer whether AI agents could start a bank run. It is whether we will design for that possibility before the first one happens, or after.

What do you think? Would you let an agent move your company's cash without a human sign-off?

End of article

Related reading

COVER FEATURE

European Sovereign Cloud in 2026: From Data Residency to Real Control

Europe's sovereign-cloud market has moved beyond political rhetoric. Sovereignty is now influencing procurement, infrastructure investment, cybersecurity policy and the architecture of cloud services across the continent. This article examines what cloud sovereignty means in practice, how the leading provider models compare, and what organisations should verify before accepting any claim that a service is sovereign.

John Wroath

ANALYSIS

AI needs a sovereign layer too. Can federated infrastructure deliver it?

AI raises sovereignty stakes beyond cloud alone: training data provenance, model weight custody, and GPU supply chains all carry jurisdictional exposure that cloud sovereignty frameworks were not built to answer on their own. Federated infrastructure is one credible response, and Europe is already building it institutionally.

John Wroath

ANALYSIS

The Yen Beneath the AI Boom

How Japan's interest-rate normalisation could expose the leverage financing the world's largest technology investment cycle. A stronger yen would not switch off AI, but it would quickly reveal which projects were built to generate returns and which were built because money was cheap.

John Wroath

Primary sources

Disclosure: Could AI agents start the next bank run? is published as part of Edition 01 of European Sovereign Infrastructure. The publication is editorially independent. No source cited in this article had sight of the copy before publication.