Financial systems have historically received a stabilising subsidy from behaviour: people delay switching, withdrawing, redeeming and prepaying. That delay has economic value, and the prudential rulebook already models it — banks assign behavioural assumptions to deposits that are legally callable on demand, insurers hold capital against lapse behaviour, and money market funds carry explicit liquidity mechanisms for periods when redemption pressure outruns asset liquidity. Behavioural persistence is not a curiosity. It is an input into financial architecture — and it has never been a price.
Agentic finance could compress that persistence without removing the liquidity mismatch underneath it. When software continuously optimises what households once reconsidered occasionally, the behavioural duration embedded in balance sheets can shorten faster than contractual terms, liquidity buffers and prudential assumptions adjust. The question is not whether automation is good or bad. It is who supplies — and who pays for — the duration and liquidity that behaviour used to supply implicitly.
"Sticky deposits" sounds like a personality trait. It is an economic observation: a class of money that is legally callable at any moment and yet behaves, year after year, as long-term funding. The current account paying 0.1% while money market instruments yield several percentage points more is not evidence that customers are loyal. It is evidence that they are slow — and that their slowness is worth something to the institution on the other side of it.
The observation has acquired a news hook. On 27 September 2026, Apollo's chief economist Torsten Slok published a note asking whether an "agentic bank run" is coming: AI assistants could sweep household cash automatically out of transaction accounts paying a 0.1% national average and into alternatives yielding roughly 3.3% to 5% — balances leaving not in panic but by optimisation (Torsten Slok, "Is an Agentic Bank Run Coming?", Apollo, The Daily Spark, 27 September 2026). The scenario deserves to be taken seriously and precisely: as a scenario. There is no evidence yet that agents are draining deposit bases at scale. What the warning actually identifies is older and better documented than the technology it names — the financial system has been quietly banking a behaviour nobody contracted for, and something is now in a position to stop supplying it.
The industry hears "bank run" and reaches for the familiar crisis playbook. But the mechanism described here does not require solvency fear. A run can begin with rational withdrawal if each depositor knows that others may do the same; the distinctive feature here is that the initial trigger is not panic but optimisation. A funding input that was behavioural rather than contractual can be repriced automatically — and once enough holders act at once, the distinction between yield shopping and run dynamics disappears. The essay is about what that change in starting point reveals.
A deposit can be legally callable and economically long. That single sentence contains most of modern bank funding. The contract says the customer may leave at any moment; observation says most of them, most of the time, do not. The difference between those two statements is duration that appears on no term sheet — behavioural duration — and it has balance-sheet value: it is what allows an institution funded overnight, in legal form, to hold assets for years. Stickiness is not a contractual property. It is a statistical property of behaviour — and for decades the distinction was easy to ignore, because behaviour was slow enough to look like infrastructure. Behavioural duration, throughout this essay, means the expected persistence of a legally short or callable position, including the delay with which its holder exercises an economically valuable exit option.
Part of the deposit franchise is therefore the value of a behaviour the customer is not explicitly paid to provide: staying. Not the whole of it — a franchise also contains payment services, deposit insurance, relationship convenience and the bank's own liquidity and credit work, and it would overstate the case to call the franchise nothing but collected friction. But the persistence component is real, separable in principle, and priced to exactly nobody. The customer does not invoice for it. The bank does not disclose it. It appears only indirectly, as net interest margin earned on liabilities that stay longer than they legally must.
The same invisible asset recurs across the balance sheet of household finance. Borrowers who could refinance a fixed-rate mortgage the day rates fall wait months or years. Policyholders hold insurance contracts through periods when surrendering and replacing them would pay. Fund investors leave redemption rights unexercised through drawdowns that would justify exit. In each case, an option that is legally exercisable is economically exercised with delay — and someone on the other side of the contract has built that delay into the price.
A deposit can be legally callable and economically long. The balance sheet contains duration that the contract does not.
None of this is a secret the industry keeps from its regulators. It is the reverse: the prudential rulebook is where behavioural duration is most explicitly acknowledged, measured and converted into balance-sheet inputs.
The Basel framework for interest rate risk in the banking book treats non-maturity deposits as behavioural positions outright — the depositor's right to withdraw in search of yield is its textbook example of behavioural option risk. Banks are required to distinguish the "core" portion of such deposits — the share expected to remain stable even through significant rate changes — and to assign it a behavioural maturity; the framework recognises that these deposits have historically been relatively stable as market rates moved, and allows that observed behaviour to enter the modelling of interest-rate risk and funding characteristics (BCBS, Interest rate risk in the banking book, Principle 5 and application guidance on non-maturity deposits, as integrated in the consolidated Basel Framework). Read that back slowly: a supervisory standard instructs banks to model a legal demand liability as multi-year funding, because that is how it observably behaves. Expected prepayment on fixed-rate loans is handled through analogous behavioural assumptions — documented, monitored and regularly updated estimates of how customers actually exercise the options their contracts give them.
Insurance regulation goes further and converts the behaviour directly into capital. The Solvency II standard formula contains a lapse-risk sub-module that stresses policyholder behaviour in three directions — a permanent increase in lapse rates, a permanent decrease, and a mass-lapse event in which a large share of policies surrenders at once (Commission Delegated Regulation (EU) 2015/35, Art. 142). An insurer's solvency capital requirement thus depends, in part, on the modelled possibility that customers stop being slow. Behaviour is not soft information anywhere in this architecture. Once it changes the cash-flow profile of a financial contract, regulation already treats it as a balance-sheet variable.
That is the essay's factual spine, and it matters that it predates the technology entirely. The balance sheet is partly modelled on what customers do, not on what contracts allow them to do. The question agentic finance poses is what happens when the doing changes faster than the modelling.
Liquidity regulation encodes the same observation from the other side. The EU's liquidity coverage ratio does not treat retail funding as one thing: stable retail deposits — covered by a guarantee scheme and either held in a transactional account or part of an established relationship — receive a 5% assumed outflow in the 30-day stress window, specified characteristics move deposits into higher-outflow treatment, and contractual restrictions that make withdrawal within 30 days genuinely unlikely can earn preferential treatment (Commission Delegated Regulation (EU) 2015/61, Arts. 24–25). It is tempting to summarise this as "the regulator assumes stickiness", and the summary would be wrong in an instructive way. The framework does not mandate the friction. It differentiates liabilities according to the expected probability and speed of their departure — it encodes assumptions about how quickly different money can leave. The calibration is conditional, reviewed, and revisable. It is also, necessarily, calibrated to the observed behaviour of humans operating at human speed — and it already treats access characteristics as outflow-relevant: among the features that push a retail deposit into higher-outflow treatment, the regulation lists the account being accessible via internet only. The rulebook has, in effect, already conceded that money that is easier to move, moves.
One corner of the system has already confronted the same problem. Money market fund regulation does not simply assume redemption inertia: it recognises explicitly that redemption behaviour can become inconsistent with available asset liquidity, and it provides designed responses — where a public-debt CNAV or LVNAV fund's weekly maturing assets fall below threshold and daily net redemptions on a single day exceed 10%, the manager's board must assess liquidity fees, redemption gates or suspension, each with specified caps and durations (Regulation (EU) 2017/1131, Art. 34). The stabiliser, in other words, is already contractual there: where behavioural liquidity could not safely be relied upon, architecture was added around the promise of redemption. It solves one product's version of the problem, not the system's — but it shows the move.
The compression of behavioural delay did not begin with AI agents, and the clearest evidence predates them. On 9 March 2023, roughly $42 billion in deposits left Silicon Valley Bank in a single day, with a further $100 billion scheduled to leave the following day had the bank remained open (FDIC and Federal Reserve post-mortem material, 2023). Official reviews of the 2023 turmoil are explicit that technology mattered: mobile banking and instant transfer infrastructure made deposits faster to move, and networked communication made the decision to move them arrive simultaneously across a depositor base (FSB, depositor behaviour and the 2023 bank turmoil). The physical and informational friction surrounding a bank run had collapsed before a single agent existed.
What remained, until now, was decision friction — and the distinction matters. An interface removes transaction friction; an agent removes decision friction. Online banking, instant rails and comparison platforms made moving money cheap, but the move still required a human to notice, decide and act — intermittently, heterogeneously, with attention that wanders and preferences that differ. A human may know the alternative pays more and still do nothing; the effort of checking terms, weighing the difference and remembering to act preserves inertia even at zero transaction cost. Agents attack precisely that residual. A discretionary, episodic human response becomes a continuously running optimisation process that never sleeps, never forgets to check, and never decides the spread is too small to bother with. Nothing about the underlying option changes. The deposit was always callable, and the legal right stays with the customer. What changes is that its exercise can now be delegated to something that exercises options — continuously, on authority granted once — and what follows need not look like a crisis at all. Millions of customers do not have to panic. They merely have to stop being passive.
The honest version of the risk requires a qualification the alarmed version skips. Automation by itself does not make behaviour homogeneous. Different agents can carry different objectives, constraints, providers and risk tolerances; in principle, a population of well-differentiated agents could make aggregate behaviour more heterogeneous than the herd it replaces.
The dangerous variable is not automation but a conjunction: automation, plus common signals, plus common model architectures, plus insufficiently differentiated constraints. Work at the Bank for International Settlements identifies exactly this channel — automaticity and the use of shared algorithms and data as potential sources of correlated behaviour, from herding and liquidity hoarding through to runs and fire sales — and warns that decision processes compressed into software can compress crises that previously unfolded over days into hours (BIS analysis of AI and financial stability). If retail agents were to consume the same market data through a small number of widely used foundation models under similar objectives, an ambiguous signal would not get a distribution of interpretations arriving over a week. It could get one response, at scale, at once.
This remains scenario analysis, and the essay flags it as such. The mechanism is strong; the empirical record is one manual preview and a set of central-bank warnings. But the direction of the mechanism is not symmetric: each element of it would shorten the window in which a treasurer, a board or a supervisor can intervene — and the prudential calibrations of Sections III and IV were all set inside that window.
The danger is not that customers become irrational. It is that rational responses become continuous, simultaneous and cheap.
Ranking the exposure is an exercise in residual friction — how much delay still stands between an automatable decision and its execution. The ranking that follows is an analytical hierarchy, not a forecast: it orders instruments by the behavioural and institutional friction standing between an economically rational decision and executable liquidity.
First, instant-redeeming cash-like instruments, and in particular tokenised money market funds marketed on continuous redemption. Where a new cash-like product promises continuous or near-instant liquidity against assets whose liquidation remains slower, it recreates the mismatch that money market regulation learned to manage through liquidity buffers and redemption tools — without necessarily carrying those tools. Where the promise is instant and the stabiliser is neither behavioural (agents hold the tokens) nor contractual (nothing gates them), the distance between the liquidity promise and the underlying liquidation process is smallest. This is the corner of the system where the essay's question is not hypothetical but a product specification.
Second, non-maturity deposit assumptions. Rate-shopping that becomes continuous and automatic attacks the "core deposit" construct directly: the share of demand money that stays through rate cycles shrinks when something is watching the rate cycle full-time. It does not go to zero — deposit insurance, payments utility and the convenience of the banking relationship all preserve genuine stickiness that has nothing to do with inattention — but the behavioural maturity assigned to what remains is a number with a direction, and the direction is down.
Third, and slower, the embedded options: mortgage prepayment and insurance surrender. Agents can identify the economically optimal exercise continuously, but execution still runs through underwriting, eligibility, notaries, tax consequences and legal process. The friction there is institutional rather than behavioural, and institutions automate more slowly than decisions do. The exposure is real; the timescale is longer.
It would be the wrong conclusion to defend the friction. Some of it was an economically valuable stabiliser bundled with an opaque transfer of value; an essay mourning its disappearance would confuse the intermediary's private benefit with the system's social function. But it would be equally wrong to celebrate the deletion as pure efficiency. Some delay is economically productive precisely because it prevents coordination failure — notice periods, redemption fees, gates and settlement windows allocate liquidity risk and stop one holder's immediate exit from taxing everyone who stays.
The problem, properly stated, is not friction. It is friction whose economic function is real but whose price and incidence are invisible. The behavioural version bundled a genuine stabiliser with a spread whose allocation was largely implicit, and disclosed neither. The design answer is to unbundle: keep the stabiliser, surface the price. The instruments already exist — notice deposits that pay explicitly for committed duration; swing pricing that charges the redeeming investor the liquidity cost of the redemption; throttles and gates that convert a run into a queue; intraday liquidity priced as the service it is. A notice deposit is not a degraded current account; it is a different product. One sells liquidity, the other sells stability — characteristics that today are bundled together in ways that depend on behaviour, a bundling agentic finance makes hard to sustain. Money market regulation, again, has already made the move once: where behavioural liquidity could not be assumed, the stabiliser was designed in. Designed friction has a failure mode of its own, and the record shows it: visible triggers invite front-running, and the Commission's own review notes that threshold-linked gates may have encouraged pre-emptive redemptions in March 2020 by investors moving to escape them (European Commission, COM(2026) 350 final, 11 May 2026). The lesson is not that explicit stabilisers fail; it is that their calibration and trigger design are part of the product, not an afterthought.
What an instant financial system cannot do is skip the decision. Someone must determine who receives immediacy, who provides the duration behind it, and who pays for the difference. The behavioural regime answered those questions by accident and disclosed the answer to no one. Its successor has to answer them in writing.
Once immediacy becomes the default, the candidate answers to "who supplies the missing duration" can be listed cold. The depositor accepts a lower yield on an explicitly stable product — the old bargain, now priced. The bank pays up for funding and passes the cost through in lending rates and fees. The customer accepts a contractual notice period in exchange for the spread the behavioural version quietly kept. The fund imposes fees or gates under stress, allocating the liquidity cost to those who demand liquidity when it is scarce. The bank holds larger buffers and earns a thinner spread. The system leans harder on central-bank facilities, moving the cost onto a public balance sheet. Or the asset side adjusts — shorter, more liquid, less transformed, reducing the quantity or maturity of credit supplied at a given price — and the economy's supply of long finance is repriced accordingly.
The essay does not need to pick among them; markets and rulebooks will, unevenly and jurisdiction by jurisdiction. The point is narrower and harder: behavioural persistence used to make the question invisible, and reduced the frequency and intensity of the liquidity stress that formal safety nets were built to absorb. Automation removes the ability to pretend the answer is "nothing". The duration that behaviour supplied does not disappear when behaviour speeds up. It has to be supplied somewhere else, by someone, at a price that will — for the first time — be observable.
Behavioural delay was an economically productive but implicit component of financial architecture — a load-bearing wall nobody drew on the plans. Customers never collectively agreed to supply banks with a stable funding base; they behaved in ways that made it stable. Insurers did not negotiate a surrender profile with their policyholders; they observed one. Banks did not contract for multi-year maturity on current accounts; they modelled it. Regulation did not mandate stickiness; it repeatedly had to model what happens if customers stop being sticky, which is the clearest admission available that the stickiness was doing structural work all along.
Digital finance removed part of the delay and 2023 showed the timescale that remains; tokenisation is separately shortening the money itself. Agentic finance could remove much of what remains. What it cannot remove is the mismatch the delay was hiding: liabilities that promise immediacy funding assets that do not have it. Automation does not remove liquidity risk. It removes the behavioural delay that used to hide it — and in doing so it converts an assumption into an invoice.
If immediacy becomes the default, stability has to become a product: contractual, operational, priced, and signed by someone who knows they are buying it. The old system ran on a design nobody signed. The new one will not get that option.
Written in a personal capacity. Analytical views only — not legal or investment advice.
Julian Gretzinger — Investor and writer on monetary history, real wealth mechanics, and financial markets. substack.com/@juliangretzinger