AI Agents Could Soon Move Your Money Between Banks Automatically — And Banks Have A Problem

Written By James Taylor
AI Could Take Over The Search For Better Savings — Who Controls The Transfers?

The Fight For Your Idle Cash

Could AI Trigger A Bank Run? The Warning Behind Automatic Money Switching

Apollo’s chief economist has warned that automated cash switching could threaten cheap bank funding, while payment authorities face unresolved questions over consent and control.

An AI assistant could eventually do something many savers put off: check what their cash earns, find a better account and arrange the transfer. For the customer, that could mean more interest with less work. For a bank accustomed to holding low-cost deposits, it could mean paying more to keep them.

On 27 September 2026, Apollo chief economist Torsten Slok raised that possibility in a note titled “Is an Agentic Bank Run Coming?” His warning concerns a future in which households delegate cash management to software. It does not establish that AI assistants already provide unrestricted, reliable switching between banks at mass-market scale.

The distinction matters. There is a credible commercial threat here, but the leap from a useful comparison tool to a system authorised to move savings involves permissions, product checks and financial safeguards.

Why Idle Cash Matters To Banks

Slok’s comparison was striking. His note contrasted a US national average checking-account rate of 0.1 per cent with selected alternatives offering between 3.3 and 5 per cent. Those figures describe his US comparison, rather than a universal rate available to every saver or a like-for-like assessment of every product’s conditions.

He argued that widespread use of agents to improve household cash returns could deprive banks of a substantial share of the cheap deposits supporting their lending.

The attraction for a saver is easy to illustrate. In a hypothetical example, 10,000 pounds held for a full year at 0.1 per cent earns 10 pounds. At 4 per cent, it earns 400 pounds. The difference is 390 pounds before tax, assuming unchanged rates, no fees and simple annual interest. These are illustrative rates, not current product recommendations.

Today, someone has to notice that gap, compare conditions and complete the move. A capable agent could reduce the time and effort involved. That would make customer inertia a less dependable source of bank income.

What An Agent Would Need To Do

An AI agent uses tools and permissions to carry out tasks, rather than only generating an answer. In banking, those tools would need to connect reliable account information with an authorised payment process.

Consider a hypothetical instruction: keep enough money available for bills and move the rest into suitable, protected savings. Before acting, the system would need to distinguish spare cash from money already committed to rent, a mortgage payment or an upcoming purchase.

It would also need to read the destination account’s terms. A headline rate could depend on a temporary bonus, a balance cap, restricted withdrawals or eligibility requirements. An account paying more interest may be unsuitable if the customer needs immediate access.

Opening a new account adds another boundary. Identity checks, customer acceptance and the institution’s own controls still apply. Finding a product is different from being able to open it and fund it.

There is also a meaningful middle ground: software prepares the comparison and transfer, while the customer approves the final action. That could remove much of the administrative burden without handing over continuing discretion.

A Profit Squeeze Is Different From A Bank Run

Banks face two related risks in this scenario. One is a gradual increase in funding costs as customers demand better returns. The other is a sudden withdrawal of deposits that creates a liquidity problem.

In April 2026, McKinsey described agents becoming an important interface between customers and financial institutions over a possible three-to-five-year horizon. Its analysis also estimated that shifting 5 to 10 per cent of checking balances into higher-yield accounts could reduce banking profits from deposits by 20 per cent or more. That is a scenario estimate, not an observed loss across the industry or a forecast of every bank’s total earnings.

Banks could respond to tougher competition by increasing savings rates, accepting lower margins or changing their funding mix. More expensive funding could influence the price or availability of new lending, although competition and each bank’s balance sheet would shape the outcome.

A run presents a more immediate difficulty. Banks hold assets with different maturities and cannot necessarily turn them into cash as quickly as customers can request withdrawals. The Bank of England’s liquidity framework recognises that mismatch and the importance of accessible liquid assets.

Moving money from one bank to another does not, by itself, remove deposits from the banking system. It changes their distribution. The institution losing the money can still face pressure even when another bank gains it.

The Risk When Many Agents Act Together

The more serious scenario is synchronised behaviour. If large numbers of agents use similar data and decision rules, they could react to the same rate change or warning at roughly the same time. That is an analytical possibility, rather than evidence of an AI-driven banking crisis already occurring.

Digital withdrawals can already move quickly. In a March 2026 speech, Bank of England official Phil Evans described how round-the-clock banking, mobile transfers and social media had accelerated deposit movements. He stressed the need for banks to prepare to access liquidity at speed.

Automation could add another source of rapid action. A transfer rule that looks sensible for one household may produce a difficult funding shock when repeated across many accounts.

The outcome would depend on how agents are designed, how customers use them and how banks manage their exposures. Liquidity buffers and central bank facilities provide defences, but operational readiness determines whether an institution can use those defences promptly.

Who Is Responsible When The Transfer Is Wrong?

The Retail Payments Infrastructure Board’s June 2026 consultation, published by the Bank of England, explicitly considered agentic payments. It described agents initiating transactions within defined parameters and identified unresolved issues around consent, authentication and accountability.

That consultation is evidence that payment authorities are considering the technology. It is not approval for every proposed autonomous financial service.

A customer instruction must translate into enforceable limits. An agent might be permitted to transfer only between verified accounts belonging to the same person, keep a specified cash buffer and seek fresh approval before adding a destination. Those are examples of possible controls, not a claim that a universal standard has already been adopted.

The system would also need to resist misleading information. A fraudulent message or manipulated product page could try to influence its next action. The practical question behind autonomous-agent control is what the connected payment system permits if the model makes a mistake.

Protection must remain part of the comparison. The US Federal Deposit Insurance Corporation distinguishes insured bank deposits from non-deposit investments, including money market mutual funds. A higher displayed return does not make those products interchangeable. Rules and coverage also depend on jurisdiction and the specific arrangement.

The Assistant Could Become The Gatekeeper

There is a further commercial question beyond interest rates: which organisation controls the customer’s choices?

If an assistant becomes the place where someone compares accounts and arranges transactions, banks may have fewer opportunities to explain products directly. An agent provider could influence which institutions appear, which terms receive attention and which offers reach the customer.

That makes its incentives relevant. Does it compare the whole eligible market, a limited panel or only commercial partners? Does a commission affect recommendations? A service presented as a personal assistant still needs a clear answer to whose interests it serves.

Banks could build their own agents, offer clearer product data or compete through better service and pricing. A bank-owned assistant may be convenient, but convenience alone does not establish that it searches widely for the customer’s best option.

The next meaningful evidence will be in the permissions and records of deployed services: which accounts they can access, which transfers they can execute, when approval is required and who bears responsibility for an error. Those details will determine whether automatic switching delivers a better savings service or simply places a new intermediary in charge of the decision.

Sources

Next Reads

James Taylor
Next

OpenAI Is Building An AI That Designs Computer Chips — And The Feedback Loop Could Be Enormous