Over the past two quarters, a single line item has been quietly hemorrhaging inside the quarterly filings of America's largest banks: noninterest-bearing deposits. That is the cheapest funding a bank can hold โ money that sits still, pays zero, and quietly subsidizes everything from mortgage books to trading desks. It is also the exact balance an autonomous agent is engineered to attack. I saw the wire tap before the wallet drained: the mechanism here is not a hack, not a run, not a scandal. It is a default setting. The moment a consumer delegates cash management to an agent that re-scans yield every ninety seconds, the "sleepy deposit" stops sleeping. The Crypto Briefing piece "AI agents could pry loose the sleepy deposits US banks depend on" names the symptom correctly and then stops. My job is to follow the wire to its termination point, because the entity that captures that flow is almost certainly not the one the headline implies.
Let me be precise about what that source actually contains, because the forensic discipline matters. It is a single-paragraph brief. Three claims, no data, no company names, no timestamps, no quoted sources. The thesis is directional: banks fund themselves on low-rate "sleepy" deposits, AI agents can mobilize those deposits toward higher-yielding fintech products, and the resulting rate competition compresses bank net interest margins. Everything else in the original is inference. That is fine โ I trade on inference before it becomes consensus. But you should know exactly how thin the evidentiary floor is before you reprice an entire sector on it.
Three actors matter, and the brief conflates them. First, the US bank as the incumbent and the casualty โ the implicit protagonist of the headline. Second, the high-yield fintech as the nominal beneficiary โ the entity the brief points at. Third, the AI agent itself as the catalyst โ a mechanism, not a business. The error the original makes is treating the second actor as the winner. In my read of the plumbing, the agent platform is the winner, the fintech is the next victim, and the bank is the first casualty. That sequencing is the entire trade.
Understand why now. Three external layers had to mature simultaneously, and none of them belongs to any bank. The first is agentic capability โ LLM-driven computer use that can navigate a banking portal, compare a money-market yield, and initiate a transfer without a human keystroke. The second is the real-time payment rail: FedNow and RTP. Without instantaneous settlement, an agent's "optimal reallocation" is throttled by ACH's T+1 or T+2 lag, and the arbitrage becomes stale before it clears. The third is the open-banking interface layer โ the CFPB's Section 1033 personal financial data rights regime and the aggregators like Plaid that sit between the consumer and the account. When all three are live, deposit mobility stops being a behavior and becomes an infrastructure. Governance isn't a feature you bolt on after the money moves; it is the money.
Here is the part the brief never says out loud: the real destination for reallocated deposits is not a bank-adjacent fintech savings account. It is a tokenized money-market fund or a stablecoin yield product, and that sink already exists. BlackRock's BUIDL, Franklin Templeton's on-chain government fund, the wave of tokenized treasury products, the yield-bearing stablecoins โ these are not speculative crypto instruments anymore. They are cash-management vehicles with 24/7 settlement, no banking hours, and no branch. The sleepy deposit was always going to wake up somewhere. The crypto rail simply built the bed first.
Now the core analysis, because this is where the money actually is.
The first thing to price is not margin compression. It is the impairment of an intangible asset that never appears on a bank's balance sheet: deposit franchise value. A bank's equity is valued, in large part, on its ability to hold cheap, sticky, noninterest-bearing deposits and lend them out at a spread. That franchise is worth hundreds of billions of dollars across the US banking system, and it is carried at zero on the books. When deposit beta โ the share of a market rate move that gets passed through to depositors โ rises from its historical ~0.4 toward 1.0, you are not just compressing this quarter's net interest margin. You are marking down the present value of every future spread the franchise was supposed to earn. The brief talks about "profit margins." The correct unit of account is the revaluation of an off-balance-sheet intangible, and that is a far larger and far faster repricing than the income statement alone implies.
Deposit beta is the variable to watch, not the deposit count. For two decades, US deposit beta was structurally suppressed by inertia. People did not move money because moving money was annoying. That friction was the moat. Now run the counterfactual: an agent with account access does not experience annoyance. Its switching cost is a function call. When the marginal cost of moving a dollar approaches zero, deposit beta stops being a behavioral constant and becomes a control variable that the agent tunes. Once beta is a control variable, the entire asset-liability management framework that banks built around "sticky core deposits" is modeling a world that no longer exists.
This is where my own audit experience sharpens the read. In late 2025 I pulled apart a proprietary AI-agent trading bot that was manipulating low-liquidity altcoin pairs. I compiled the wash-trading evidence, published the pattern, and the exchange delisted the token. The technical lesson I carried out of that investigation was not about fraud. It was about agency: once you give a machine discretionary execution authority over capital, its behavior stops being "customer behavior" and starts being "market behavior." A depositor who reallocates is a person. Ten thousand depositors whose agents run the same optimization are a single algorithmic actor wearing ten thousand faces. The bank's risk model sees ten thousand customers. The market sees one.
Which brings us to the architecture gap, and this is the structural fuel nobody is pricing. The core systems of most large US banks are still mainframe batch processors โ COBOL-ledgered, end-of-day reconciled. They were designed for a world where money moved on a schedule. An agent demands real-time, event-driven, sub-second reallocation. The bank's core cannot answer that question in the time the agent asks it. Technical debt inside a bank's core is not an IT line item anymore. It is a cash-flow risk, because the system's latency is the constraint on how fast the bank can even respond to a deposit request. The core modernization programs that banks have deferred for a decade just became defensive capital expenditure. They did not choose this timeline. The agent chose it for them.
The interface layer decides who holds the valve. There are three ways an agent can reach a bank account: a direct API through open banking, an aggregator like Plaid that sits in the middle, or legacy screen-scraping. Ranked by speed and reliability, it is direct API, then aggregator, then scraping. Ranked by who controls the relationship, it is the reverse. Whichever layer the agent standardizes on becomes the choke point for the entire deposit flow. If aggregators become the standard, they become the de facto clearinghouse for retail cash, and their economics โ per-call fees, data rights, exclusivity โ become a tax on every dollar that moves. The interface is leverage waiting to be wielded.
Here is the part that should terrify a bank treasurer: the same concentration that makes agents efficient makes them dangerous. If a meaningful share of depositors delegate to a handful of dominant assistants โ say a default assistant shipped inside an operating system โ those agents will converge on similar "optimal" allocations. Homogeneous strategies produce correlated exits. On a bad-news morning, thousands of agents could independently decide that the same bank's risk-adjusted yield no longer clears their threshold, and pull simultaneously. That is not a bank run in the 1930 sense of a queue outside a branch. It is a bank run in the algorithmic sense: a synchronized, machine-timed withdrawal that no human depositor consciously chose. SVB in 2023 was the preview โ social media compressed a multi-day run into a single afternoon. Agentic withdrawal compresses it into a single block. The crash wasn't a market event; it was a plumbing event. The next one is plumbing with a scheduler.
Now the responsibility vacuum, because this is where the legal risk hides and where the brief is silent. When an agent moves money and the money is lost โ to a mis-click, a manipulated prompt, an impersonated destination โ who is liable? The consumer authorized the agent. The agent acted within its instructions. The bank executed a valid-looking transfer. Regulation E governs electronic transfer errors, but it was written for human-initiated mistakes, not machine-initiated discretion. Authorized push payment fraud โ where the victim themselves authorizes the transfer โ sits in a similar grey zone. There is no statute that cleanly assigns liability across the user-agent-bank-fintech chain, and that gap is not a bug for everyone. For the well-advised, an undefined liability boundary is an arbitrage window.
I have watched this movie before, in a different theater. In 2021 I audited a Yearn Finance governance proposal, found the centralization risk the yield narrative was hiding, and helped mobilize more than a thousand holders to vote it down. The lesson was not that governance is broken. It was that governance is often theater staged around a decision that a small group has already made. The agent platform is the new sequencer: a nominally neutral execution layer that is, in practice, a single point of control over where the money goes. And here is the uncomfortable part โ we spent two years calling Layer 2 sequencers "decentralized" when they were single nodes in a costume. The same theatrical decentralization will be performed by the agent platforms. They will be marketed as neutral utilities. They will be choke points.
Which leads to the contrarian read, and this is the angle the brief misses entirely.
The consensus narrative โ banks lose, high-yield fintech wins โ is the wrong map. Run the disintermediation one layer deeper. The fintech that today captures the yield-seeking deposit is itself standing where the bank stood. It is a destination, and destinations get disintermediated. If the agent controls the default configuration, then the fintech is no longer a brand competing for a customer. It becomes a backend yield supplier competing to be included in the agent's default list โ a commodity input, selected by an algorithm, priced down to its spread. The fintech is not the winner of this trend. The fintech is the next layer of the trend, and it will be repriced the same way the bank is being repriced now, just one cycle later.
The entity that actually captures the value is the one that owns the entry point: the assistant that lives on your phone, your OS, your browser. That actor collects the customer relationship, the default authority, and the data network effect โ more users produce better allocation strategies, which attract more users. Banks do not have that network effect. Neither do most fintechs. A platform does. Which is why the most credible cross-over threats are not banks-versus-fintechs at all; they are the owners of the assistant. An OS vendor that already ships a savings product has demonstrated it can hold deposits. A default-assistant vendor already owns the moment of intent. When the interface and the intelligence are the same company, the bank and the fintech both become interchangeable plumbing underneath it.
And the second contrarian point, the one that reframes the whole trade: this is not an independent trend. It is a derivative of the rate cycle. The entire mechanism requires a positive spread between what a bank pays on deposits and what the market pays on cash. In a high-rate regime, that spread is wide and the agent has room to work. Begin a credible easing cycle, and the spread narrows, the arbitrage shrinks, and the urgency to move money decays. The brief describes a one-directional force. It is actually a rate-cycle amplifier โ powerful on the way up, self-defeating on the way down. Any position built on "AI agents will drain deposits" is implicitly a position on the Fed staying higher for longer. That is a macro bet wearing a technology costume, and it is the single most important caveat the original article omits.
There is also a precedent the brief ignores, and it is sitting in plain sight in the Chinese market. In 2013, a high-yield, frictionless savings product pulled enormous sums out of Chinese bank deposits in a matter of months, and was eventually folded into money-market-fund regulation. The path was not "innovation defeats incumbents." It was "innovation scales, incumbents complain, regulators absorb." Expect the same arc in the US: scale, then supervision. The regulatory question is not whether agents get to move deposits. It is who gets to authorize the agent, and under what liability regime, once the flow is large enough to matter for financial stability. That question is the real gate on this trend, and it has not been answered.
Let me also correct a small but costly misread. The brief frames high-yield fintech as the beneficiary. But the true, scalable sink for reallocated retail cash is not a bank-adjacent app โ it is the money-market complex and its tokenized equivalent. Money-market funds hold trillions and can absorb flow at a scale no fintech balance sheet can match. The on-chain tokenized versions of those funds give the agent something no traditional account can: programmable, 24/7, composable settlement. When your agent can move cash into a tokenized treasury position at 2 a.m. on a Sunday and back out before Monday's open, the banking calendar itself becomes a competitive disadvantage. The crypto rail did not create this demand. It simply removed the last friction the incumbent depended on.
None of this is a prediction that banks disappear. It is a prediction about where the margin migrates. Banks will likely be forced to wholesale-fund themselves more heavily โ leaning on brokered deposits and market funding โ which raises their cost of capital and degrades their funding quality. The institutions most exposed are the mid-sized and regional banks that built their franchises on cheap core deposits. The ones most insulated are those with irreplaceable transactional relationships โ the accounts where payroll, payments, and credit are entangled in ways an agent cannot trivially unwind. The moat was never the interest rate. It was the entanglement.
For the reader who wants the tradable signals rather than the narrative, here is what I am actually watching. Noninterest-bearing deposit ratios across the large banks โ two consecutive quarters of meaningful decline is the cleanest leading indicator that inertia is breaking. The pace of the CFPB's 1033 open-banking rule โ its survival or repeal determines whether agents get legal account access at scale, which is the on/off switch for the whole trend. The Fed's rate path โ because the spread is the fuel. FedNow and RTP penetration โ because real-time rails are the prerequisite. And the first marquee incident of an agent-driven consumer harm or a synchronized agent withdrawal โ because that event, not a white paper, will set the regulatory clock. Trust no one, verify the chain, strike first.
I have spent a decade watching capital migrate toward whoever removes the most friction. Speed is the only currency that doesn't inflate. The bank's friction was inertia; the fintech's friction is brand loyalty; the agent's friction is authorization. Whoever solves authorization โ legally, securely, and at scale โ owns the flow. That is the whole game, and it is being played right now, in default settings most depositors will never notice they enabled. While you read the news, I traded the rumor. The rumor is that the fintech wins. The trade is that the entry point does. The question worth sitting with is not whether AI agents will pry loose the sleepy deposits. It is who, exactly, will be standing there to catch them when they fall โ and whether that party answers to a regulator, a shareholder, or no one at all.