The agent rotates faster than the core deposit models can learn

On Sunday 27 September 2026, Torsten Sløk, chief economist at Apollo Global Management, put a question in his daily note “The Daily Spark”: “Is an Agentic Bank Run Coming?” The note runs to two sentences and a chart. The first says that personal AI agents such as Meta’s Muse could soon sweep household cash automatically into accounts paying 3.3% to 5.0%, rather than leave it at the US national average of 0.1% on checking accounts. The second adds that if every household used such agents, banks could lose a large share of the cheap deposits they rely on to make loans. By Monday the warning of an “agentic bank run” was in the headlines, at Bloomberg Law as well as CoinDesk.

The headlines turned a question into a warning and a funding question into a liquidity question. Neither captures the point as precisely as Torsten Sløk’s own second sentence. An agent that moves balances to the best available offer every few months does not produce the kind of 30-day withdrawal that the Liquidity Coverage Ratio (LCR) is built to absorb. It undermines the assumptions underneath. The rules by which banks in the EU judge the stability of their deposits rest on the length of a customer relationship, the type of account and ten years of volume history, and all three presume that customers rarely compare rates. Issue 16 of the newsletter “the agentic banker” dealt with the rate gap itself, issue 17 with the objections from the large US banks. This piece looks at the models a European bank uses to steer its deposits. What follows is a derivation from rules, rate data and research, not a forecast. According to Meta and Plaid, Muse can so far only view accounts. Neither announcement mentions a transfer function.

Torsten Sløk’s second sentence is about funding

The chart behind the note compares eleven offers from US fintechs and online banks, from 3.3% at AlumniFi to 5.0% at Adelfi, with two national averages published by the Federal Deposit Insurance Corporation (FDIC): 0.4% for savings and 0.1% for checking accounts. In its overview dated 19 October 2026 the FDIC shows 0.07% for interest-bearing checking accounts and 0.38% for savings. Expressed in dollars, which the note itself does not do, and using the 0.1% from the chart, the author’s own calculation puts the gap on a $10,000 balance at between $320 and $490 a year, depending on whether one takes the weakest or the strongest offer in the chart. The note gives no estimate of how much money might move, or how quickly. “Bank run” appears only in the headline, and there as a question. The phrase is provocative, yet Torsten Sløk describes the mechanism soberly in his second sentence: cheap deposits on which banks rely to fund lending.

The agent Torsten Sløk names is real and widely used. Meta launched Muse on 8 September 2026, a personal agent that, according to the company, can open a browser, fill out forms and negotiate. Accounts can be linked through the data provider Plaid, which says users can “share balances, transactions, investment holdings, mortgage information and more”. On 21 October Meta reported more than ten million users. JPMorgan Chase and the broker Charles Schwab, meanwhile, are drawing a line. Jamie Dimon, chief executive of JPMorgan Chase, told analysts on the third-quarter call on 13 October, as reported by CNBC: “We're not going to allow just any … outside agent to come in inside the company, because it's just too risky for clients and for JPMorgan.” Rick Wurster, chief executive of Charles Schwab, said the firm blocks agents that try to log in on behalf of clients, adding that this rarely happens so far. Charlie Scharf, chief executive of Wells Fargo, called the expectation of a mass shift in deposits “completely overdone, especially for the time period we can see”.

These blocks are unlikely to be permanent, and the banks themselves do not present them that way. A JPMorgan Chase spokeswoman told CNBC that the bank is “committed to enabling agentic connections securely”. On 9 October Sierra published a draft Personal Agent Protocol, known as Poppy, which is meant to set out the rights a customer can grant an agent. Meta is among the first partners, and Bank of America, BBVA and Wells Fargo have joined as design partners. The standard is not yet in use. So the question is less whether agents will reach accounts than on what terms. For a bank in the EU the more useful question sits a level deeper: on what grounds do its rules treat a deposit as stable, and how much of that survives once software takes over the rate comparison that customers have so far not bothered to make?

The LCR flags a rate-chasing depositor only once a second feature appears

Under Delegated Regulation (EU) 2015/61, the Liquidity Coverage Ratio determines how much liquidity a bank must hold for a 30-day stress period and assigns an outflow rate to each type of deposit. Article 24 treats a retail deposit covered by deposit insurance as stable, with a rate of 5%, if it forms part of an established relationship or sits in a transactional account, which the regulation describes as an account where “salaries, income or transactions are regularly credited and debited”. A relationship counts as established if, among other things, the contract with the bank has run for at least twelve months, the customer holds a residential or other long-term loan there, or uses another active product. All other retail deposits carry a rate of 10% under Article 25(1).

Article 25(2) lists the features that call for a higher rate, and two of them describe an agent’s target rather closely: an “internet access-only account” and an interest rate that “significantly exceeds the average rate for similar retail products”. The list also covers total balances above €500,000, short maturities or notice periods below 30 days, and depositors from third countries or deposits in foreign currencies. The real teeth sit in Article 25(3), however. It takes a total balance above €500,000 or two of the other features, such as an online-only account paying an above-average rate, to move a deposit into a band of 10% to 15%, within which the bank sets the rate itself. An above-average rate on its own is not enough. A bank that skips the assessment has to apply 15% to 20%. The idea is not new. The European Banking Authority (EBA) listed rate-driven products and internet-only banks as risk factors in its 2013 guidelines, repealed those guidelines in 2017 and carried the idea forward in Article 25.

This creates an asymmetry that the debate usually misses. The higher outflow rates apply to the bank that receives the money, not to the one that loses it. At the customer’s main bank the salary account remains a transactional account with an outflow rate of 5%. Only the balance on it shrinks. The Claude Money article published on 7 October noted that a bank loses neither deposits nor payment flows to an agent sitting on the account, because the account stays where it is. That holds for the account, not for the balance. The benchmark moves too: if banks align their rates because agents exploit every gap, the average rises and the test of a rate that “significantly exceeds the average rate” bites less often. The LCR measures the risk of a withdrawal within 30 days. What a rotating agent changes plays out over quarters, and the LCR was not built for that. Pedro Machado of ECB Banking Supervision questioned the calibration of the outflow rates himself in a speech in October 2025, with uninsured, digitally active deposits in mind.

A saver who switches every three months never becomes a regular customer

The Bundesbank’s interest rate statistics for monetary financial institutions (MFIs) show whether there is anything to gain in Germany. Household overnight deposits earned an average of 0.51% in new business in August 2026, deposits with an agreed maturity 2.33%. The overnight figure is a blend, though. By the Bundesbank’s own description it includes “Konten mit und ohne Zahlungsverkehrsfunktion”, accounts with and without a payment function, so current accounts and instant-access savings together. The market looks different. The Finanztip comparison published on 21 October lists 4.1% as the best teaser rate on instant-access savings, limited to three months, followed by two offers at 4.0% for four months. The best ongoing rate is 2.45%. The ECB’s deposit facility rate has been 2.50% since 16 September. A teaser rate above four per cent therefore lies well above the rate at which a bank can place surplus liquidity with the central bank. It is a customer-acquisition cost and pays off only if the customer stays once it expires.

Earning teaser rates continuously means moving on to the next offer after three to four months. People rarely do that, because every switch costs attention, while software does it at no extra effort. For the LCR this has a consequence that follows from the wording: at each new bank the twelve-month period for an established relationship under Article 24(2)(a) starts again. A balance that moves every quarter never reaches it, unless the customer holds a transactional account or another product at the bank concerned. The features the regulation ties to stability only emerge over time, and an agent never leaves a balance at one bank long enough for them to emerge. The platform Raisin, long known in Germany under the WeltSparen brand, says in an undated statement on its US corporate website that it powers over $90 billion in deposits worldwide. It does not disclose how much of that moves with the cycle of teaser rates.

There is a good deal to say against the scenario, and the strongest version of the objection rests on the evidence. The study “Sticky Deposits, not Depositors” analysed ten million accounts at US credit unions and found that 10% of depositors hold 70% of balances. Customers with small balances respond to rate changes, those with large balances hardly at all. The study “Dynamic Competition for Sleepy Deposits”, published in March 2026, finds that only about 15% of account closures stem from rate shopping, while more are prompted by moving house or a bereavement. Charlie Scharf’s argument points the same way: anyone who wanted a better rate could already switch online. That is true, but it describes the cost of switching. What an agent takes off the customer’s hands is something else: the need to think about the rate at all. The same study on sleepy deposits says as much: “The advent of FinTech services that automate shopping for deposits could erode this source of franchise value.”

The turn in rates showed the same movement in slow motion

Germany has already been through a shift driven by rate gaps, without any agents. Between June 2022 and June 2024, overnight deposits held by euro area households at German banks fell by €129.9 billion according to ECB time series, and savings deposits redeemable at notice by €133.5 billion. Deposits with an agreed maturity rose by €423.5 billion over the same period, from €260.8 billion to €684.3 billion. Total household deposits grew by €160 billion. The share of overnight deposits fell from 69.5% to 60.9% and had recovered to 64.2% by August 2026. The money changed product and stayed within the banking system, and it did so over two years. It is plausible that the rate gap caused it, but this remains a derivation: the sources searched contain no Bundesbank study on how readily German depositors move for a better rate.

For banks’ liquidity calculations the movement was benign. The money went into fixed-term deposits, and under Article 25(4) of the LCR regulation a bank may exclude from its outflows deposits that cannot be withdrawn within 30 days or carry a material penalty for early withdrawal, provided the product in question meets the conditions. Banks thus appear to have tied in their customers for longer, at a higher price. An agent chasing teaser rates on instant-access savings goes the other way: money available every day, which can move precisely because it is available. The second historical case is American and illustrates speed rather than mechanism. According to the Federal Reserve’s review, Silicon Valley Bank “lost over $40 billion in deposits” on 9 March 2023. These were uninsured, concentrated deposits from technology firms and venture capital investors who fled out of concern for the bank’s solvency. A rate-chasing agent acting for households moves insured balances in search of yield. The resulting risk is one of cost, and it builds over months.

The European evidence fits this distinction. An ECB working paper (“Mind the App”, Working Paper No 3092, August 2025) analysed data from 110 directly supervised banks between 2016 and 2024. In normal times the authors find no systematic effect of digitalisation on deposit volatility. In the months with the most extreme outflows, one percentage point more online banking raises outflows by about 0.28 percentage points, and for insured retail deposits they find no effect, under stress or in normal times. In October 2023 the Basel Committee observed that digitalisation is removing “many of the frictions that may have previously slowed down the magnitude of liquidity outflows”, referring to the uninsured deposits of the failed US banks. In the ECB supervisory data, insured household deposits show no digitalisation effect even under stress. They start to move where someone compares the rate on the customer’s behalf.

The core deposit model learns from a world without agents

The bigger bill lands in interest rate risk in the banking book (IRRBB). Overnight deposits have no contractual maturity, so banks model a core, the portion that stays for years despite being payable on demand, and use it to fund long-term loans. For the standardised approach, Article 8 of Delegated Regulation (EU) 2024/857 requires banks to identify the stable part “using observed changes of the volume of the deposits due to upward and downward movements of the risk-free interest rate for a period of at least the preceding 10 years”. Among the factors that shape the pass-through rate, the same provision names “competition from other firms”. It caps the core component of the stable part at 90% for transactional and 70% for other retail deposits, and the average maturity at five years and 4.5 years respectively. Internal models are not bound by these figures. They too, however, can learn only from the behaviour their data contain.

Ten years of volume history contain no agent that moves balances every quarter. The model estimates the stability of a world in which customers were inert, and it registers a structural break only once that break has shown up in the data for several years. Anyone who has run a derivatives book will recognise the position behind it. From the bank’s side, an overnight deposit is a written option: the customer may withdraw at any time or demand a higher rate, and the bank earns a premium because the right is almost never exercised. An agent is a holder who exercises every option the moment it is in the money. No deposit disappears as a result, but the premium does, and with it part of the value the balance sheet assigns to the deposit franchise. The sleepy deposits study puts a number on it: depositor inertia accounts for 57% of average deposit franchise value.

In rate planning the matching measure is the deposit beta, the share of a policy rate change that a bank passes on to depositors. An FDIC working paper from May 2024 measures a cumulative deposit beta of 0.27 for US banks over the 2022 and 2023 rate cycle, 0.25 at the smallest institutions and 0.44 at the largest. An agent that compares rates continuously pushes these values up for the customers it serves. The same logic carries into the Net Stable Funding Ratio (NSFR), which gives stable retail deposits 95% and other retail deposits 90% available stable funding under Articles 428n and 428m of the CRR. The counter-argument deserves to be put at its strongest. The same inertia that protects margins stabilised banks in 2023, and the sleepy deposits study notes that it “significantly reduced” default probabilities during that year’s banking turmoil. A bank that loses margin therefore also loses a buffer. For thinly capitalised institutions the cost question can thus become a solvency question, and only then a run question. Quite apart from agents, the EBA urged banks in June 2026 to check whether their funding plans adequately allow for “the risk of abrupt repricing dynamics in deposit markets”.

An in-house agent eats into the bank’s own margin first

The obvious answer for a bank is an agent of its own. According to chief financial officer Jeremy Barnum, JPMorgan Chase is piloting a tool called Smart Cash that predicts a customer’s cash flow and moves funds “to maximize yield”. In Europe such an offer has a regulatory head start. Under Article 15 of Delegated Regulation (EU) 2018/389, a bank may waive strong customer authentication (SCA) when a customer moves money between their own accounts at the same institution. An outside agent that wants to move money to another bank, by contrast, needs a licensed payment initiation service under the Payment Services Directive (PSD2), the customer’s explicit consent and, as a rule, SCA for each payment it initiates. The overhaul of this framework into a third Payment Services Directive (PSD3) and a Payment Services Regulation (PSR) has been politically agreed, but according to the sources reviewed formal adoption was still pending.

The head start comes at a price. An in-house sweep moves balances from a current account paying an average of 0.51% to a better-paying account at the same bank, and every euro moved lowers the margin the deposit business earns today. With overnight deposits making up 64.2% of household deposits at German banks, that is not a marginal effect. The bank keeps the deposit and gives up part of the income it has so far earned thanks to its customers’ inertia. The alternative, admitting outside agents through a controlled interface as the draft Personal Agent Protocol envisages, at least preserves the customer relationship from which Article 24 of the LCR Delegated Regulation derives stability. Which route pays depends on how much of a bank’s deposit margin rests on customers who do not compare rates today. That figure appears in no reporting template, yet it determines how seriously the question in Apollo’s note should be taken.

Recommendations for treasury, asset-liability management and sales

Glossary

Christian Schablitzki – the agentic banker

Strategy & Management Consultant · Agentic AI expert for financial institutions

More than 20 years in investment banking and derivatives trading, followed by over 10 years advising financial institutions. Currently a Partner at Infosys Consulting in Germany. Certified as Claude Partner – Claude Code (Anthropic), as well as in Google AI, Generative AI Leader (Google Cloud) and IBM RAG and Agentic AI.

LinkedIn profile → · agenticbanker.ai →