When the deposit beta goes agentic

Deposit franchises are priced on a depositor who has to notice and money that takes a day to move. Software is coming for both.

Aug 02, 2026

If the Fed raises rates by five hundred basis points, you can pretty much rely on US banks to pass roughly forty cents of every dollar through to depositors. Every asset liability committee in the western world knows this figure as the deposit beta. Now if you’re in an ALCO meeting, then the chances are you’ll read a low beta the same way as everyone else. You’ll be looking around the room nodding along because a low beta means funding is cheap, the franchise is strong with deep relationships and loyal customers.

A working paper from the ECB published this year suggests that reading may be subject to some challenge in the coming years. And I think it’s really a consequential piece of banking research for anyone watching what tokenization does to bank funding. What it finds is that deposit stickiness is not actually a behavior at all, rather it’s more like a census of who hasn’t left yet. The deposit franchise it describes turns out to be priced on two frictions at once. Those frictions are the depositor’s attention, and the speed of the money. If you’ve got this far, and you’ve read any of my previous stuff, then you know I’m very focused on what atomic settlement does to the second. The ECB has now measured what happens to the first, and now I can’t stop thinking about how the two are about to meet on the same balance sheet.

A census, not a behavior

ECB Working Paper 3255, by Ugo Albertazzi, Finn Faber, Alessandro Gavazza, Oana-Maria Georgescu and Ernest Lecomte, builds a structural model of euro area deposit markets from 2007 to 2024 and asks a simple question. Why is the pass through of policy rates to deposit rates so low, and why does it keep falling? Deposit betas declined from roughly 0.30 in the 2007 hiking cycle to roughly 0.10 in 2022, and the standard explanations reach for the bank side. Bankers will be familiar with impacts of things like excess liquidity, capital positions and concentration.

The paper rejects all of that, I must say almost brutally. The authors test the explanations by re-running history inside their model. First they made every bank identical, same liquidity, same capital, same competitive position, and deposit rates barely move. Then they made every depositor identical instead, and lo and behold rates jump! In that version of 2024, households get paid nearly half a percentage point more on their everyday accounts. So actually the low rates were never about what banks were doing, they were about which customers were left holding the accounts.

Let me pull the thread and try and break this down a bit more. So when rates rise, the most rate sensitive, highest balance depositors leave first, into term deposits, money market funds, e-money etc. Then what remains is a pool that is, almost by design, less rate sensitive than the one the bank started with, so the measured beta falls, the bank’s pricing power rises, and the margin expands. In the authors words, rate sensitive depositors switching out

“decreased the average rate sensitivity of the remaining pool of sight deposits. In turn, banks market power over sight deposits increased.”

The franchise didn’t infact get stronger. The customers who would have tested it left, and the beta was measured on the ones who stayed.

Its worth looking at the data here, because the top 10% of depositors hold 28% of household deposits, and the markdown on demand deposits, the gap between what the money earns and what the bank pays for it, equals 92% of the deposit business’s gross revenue.

The deposit franchise isn’t a lending business with a funding side. Its infact what most of us who have worked in this industry already know, its an inertia business, and so nearly all of the revenue is actually the measured inattention of whoever hasn’t left yet.

The planes that came back

It reminds me of the famous piece of statistical history. In the Second World War, the US military studied the bullet holes on bombers returning from missions and proposed adding armor where the holes clustered. Abraham Wald, the statistician they consulted, told them to armor where the holes weren’t. The returning planes were the survivors. The holes they carried marked the places a plane could be shot and still come home, and actually the fatal bullet holes were on the aircraft nobody got to examine.

Deposit analytics has been armoring the bullet holes for forty years. Every stickiness estimate, every core deposit intangible valuation, every funds transfer pricing assumption is fitted to the depositors who stayed, and the model ends up promoting their inertia into a law of nature. Thats why this ECB paper is so fascinating, its as far as I know, the first study to instrument the survivors properly and show that the sticky pool is a residue, and not a population. Betas are state dependent. The same rate rise produces a different answer depending on who happens to be left when it arrives. Which I think means the number every ALCO treats as a parameter is actually a snapshot of a queue, taken mid exit.

And we already know, empirically, what the queue looks like when it moves. The FDIC’s forensic study of the 2023 failures, built from the actual core and wire records seized at SVB, Signature and First Republic, found that two of the banks lost roughly half their deposits in three business days, and the depositors who ran emptied their business operations accounts completely, the exact category every model treats as stickiest. At Signature, escrow style balances fell 83% in a couple of business days. The survivors behavior told us nothing about the planes that didn’t come back, right up until the shooting started.

Two frictions, one franchise

Now let’s separate the two things that 92% markdown is actually charging for, because I think they fail a bit differently.

The first is attention. The depositor has gotta notice that their money earns nothing, find a better rate, and then decide the difference is worth the bother. The second is settlement. Even a depositor who has noticed then faces cutoff times, banking hours, transfer forms, and a day or three of being out of the market, and for an operating balance that might be needed tomorrow morning, that delay is not just an inconvenience, it’s basically the reason not to move at all. I’d posit the deposit franchise earns its markdown because both frictions bind at once.

History shows what happens when only one of them erodes. Money market funds have been attacking the attention friction since the 1970s, and they now hold $7.6 trillion, but the erosion took fifty years, because every dollar still had to be noticed by a human and moved over rails that closed at five. Non interest bearing deposits at US banks fell from $5.5 trillion at the March 2022 peak to $3.85 trillion three years later, a 30% decline through that same slow, human channel. Painful, survivable, visible in the data as it happened. The franchise repriced and carried on, which is probably why every bankers instinct says this time is like last time.

Tokenization is what removes the second friction, and its why I’m so interested in this nerdy deposit pricing paper. A tokenized money market fund settles atomically, around the clock, with no settlement risk on the round trip. Sandy Kaul at Franklin Templeton has already named the end state,

the stablecoin is the checking account, the tokenized fund is the savings account, and you rotate between them at the speed of a swap.

I made the corporate version of this argument in The End of Idle Money, if $100 million can move into a tokenized fund and back inside an hour with no settlement risk, it will. The ECB paper supplies the retail half, the residual pools stickiness was never loyalty, it was the price of noticing plus the price of moving. Tokenized rails are gonna take the price of moving to zero. Which leaves the whole franchise standing on a single friction, attention.

The agent always notices

Attention is exactly what an agent deletes I think, and this is where it all gets super interesting.

The ECB authors gesture at the risk, flagging “digital platforms that aggregate and compare deposit rates” as a force that raises effective rate sensitivity. But a comparison platform still needs the human to look at it. An agent doesn’t. It is attention running continuously, at practically zero marginal cost, on behalf of exactly the low balance, inertial depositors who anchor the residual pool, and the FCA’s Mills Review, published this month, is the first regulatory document I’ve seen to take it seriously. It gives the risk a name, hyper switching, and its language is worth quoting because I haven’t really seen regulators write like this

“in a world of empowered AI agents, able to shift customer savings between banks and building societies instantly and effortlessly, the banking market could be more fragile, leading potentially to financial stability implications.”

You can’t help but note the compound in that sentence! Instantly is the rail. Effortlessly is the agent. Neither of the words alone produces fragility, an agent on legacy rails spends its life waiting for ACH windows, and an instant rail without the agent still needs a human to care. The Review flags mutuals and smaller deposit funded institutions as most exposed.

The demand side is further along than most bankers assume. The Review’s survey of just over 5,000 UK adults found 26% already trust general purpose tools like ChatGPT for financial advice, 20% say they would use a fully autonomous AI to manage their finances, and only 40% know there is currently no recourse if it goes wrong. It notes £300 billion sitting in low interest UK accounts, which is the pool an optimizing agent gets pointed at, and cites Alipay processing 120 million agent initiated transactions in a single week.

A McKinsey 2026 banking review makes the same point from the industry side. It forecasts that agentic AI sweeping idle cash to higher yield in real time threatens the roughly 60% of retail banking revenue that is net interest income, and generative AI reached 45% of the US working age population in two years, against fifteen years for digital banking. I’ll tell anyone that’ll listen that the platform shifts are moving faster and faster.

If we put it all together, then I think it changes the shape of the risk. The ECB shows betas are compositional, they fall because the attentive leave and the inattentive remain, and they are state dependent, the same rate rise produces a different answer depending on who is left in the pool. An agent flips the composition. The beta doesn’t drift up as delegation spreads. It’ll snap, discontinuously, the day the residual pool stops being mostly inattentive and starts being mostly delegated, and because the historical series was generated by humans switching slowly over slow rails, I think we’ll see that every backtest will look fine until the day it doesn’t.

Let me try and be clear about whats evidence here and whats my thinking. The ECB paper never mentions AI agents or tokenized money; its outside options are money funds and e-money, but the agentic reading is mine. But the paper’s core result doesn’t have to be agentic, it just needs something to lower the cost of noticing, and something else to lower the cost of moving. To me, the first is a product description of an AI financial agent. The second is the product description of tokenized settlement.

Repricing the inertia business

Euro area banks paid 23 bps on overnight deposits while earning 295 bps on the money. I know that American depositors get a better deal and American markdowns are thinner, but you can’t reprice that the way you’d reprice a product! It ain’t a product, its actually the business. Record margins in the industry are the reward for holding customers who had no easy way to leave.

The ECB authors got there way before I did. They close their paper by saying banks now depend on a core of inertial depositors, and that nobody knows how those depositors behave under stress, mainly because they’ve never really been tested. They also note that platforms comparing deposit rates across banks could make retail depositors far more rate sensitive. I think they stop one step short of what’s about to do that comparing.

So what does an ALCO do with this on a Tuesday morning? I can think of three things.

We should think about how we segment the book by who’s holding the phone, not by product or balance size. The mix of the pool is really the name of the game here, and a big determinant of that future mix is whether an agent sits between the customer and the account. The paper does show why thats new. The top tenth of households holds 28% of deposits, and above the 90th percentile the odds of leaving money in an overnight account fall off a cliff. Small balances have stayed put because paying attention costs more than its worth on $2000. But you have to assume that an agent makes that attention nearly free. A $2000 balance with an agent attached is likely going to behave like a $2 million balance with a treasurer attached.

Ok, so we then need to price for the agents ranking and not the branch across the street. The FCA reckons 11 million UK adults, about one in five, are open to letting AI act for them inside goals they set once. When the marginal deposit is allocated by a ranking, a bank that a machine can’t read isn’t in the game at all.

I think you need then to take a position on the rails. The money that leaves goes to a tokenized instrument over a tokenized rail, and whether that rail is your tokenized deposit or someone elses stablecoin decides whether the balance leaves the bank or just moves around inside it. Same receive before issue argument I made for the long tail, but now on the funding side.

I don’t want to be all doom and gloom here. Agents need mandates. Mandates need rules that don’t exist yet, no agent registry, no liability regime, no redress when it goes wrong. In the US those rules exist to lesser extent than they did two years ago. Section 1033 was supposed to settle who can act on a customer's behalf, what they get to see, and who eats the loss when it goes wrong. The CFPB then told the court its own rule was unlawful and should be vacated, and is of course now rewriting it. One of the questions formally reopened is whether a consumers representative means a fiduciary, or any third party the customer authorizes?

Banks won that case, and all the trade groups called it a win for accountability. What they actually removed was the rulebook the agent would have had to follow, but not the agent. Still, we all know adoption runs slower than the technology. It always has, and I’ve spent enough years inside banks to know how long the last mile can take.

But if you ask me whats different this time? Its the brakes are coming off together, not one after the other. Attention is becoming software while settlement is becoming atomic. Sequence would have bought a decade. Parallel probably buys a lot less than we’d like to think.

Banks survived money market funds, survived deregulation, survived the comparison sites, and repriced each time without dying. The base case here is adaptation. What has changed though is the instrument you use to see it coming. Every previous repricing showed up in the deposit beta while it was happening, because people switch in crowds, slowly, through branches and call centers that close. I really believe most of what looks like loyalty is just the loyal ones being all that’s left.

We already know what happens when that base stops being inertial. Silicon Valley Bank lost $42 billion in a day, and that was humans, on a Thursday, using rails that closed at five. Nobody’s beta caught it.

Deposit betas were never actually behavior. They were a census taken on rails that closed at five, and its unlikely either half of that survives over the coming years… Agents switch in code, over rails that don’t close. And the beta that’s supposed to warn you was designed and fitted to humans.

References

The measured mechanism

The empirical runs and the aggregates

The agentic layer and the rails

Earlier pieces in this arc: The End of Idle Money, Friction as the Franchise, and Every Bank Needs a Wallet.

Atomic Settlement

Discussion about this post

Ready for more?