I keep coming back to a simple, slightly uncomfortable thought. If software that once took years and a small army of specialists can be rebuilt in a handful of days, who actually pockets the difference? Not in a press release. In the real books of a bank that still posts most of its truth overnight.
The Quiet Problem Inside Everyday Banking
Most people meet a bank through an app. The balance looks live. Transfers feel instant. Fees appear with a polite little notification. Behind that glass, a surprising share of the industry still runs on ideas drawn up when two new states were joining the union and business software was supposed to be readable by people who were not mathematicians.
That older language was built so clerks and managers could follow the logic. It worked. It kept working. It is still working, which is both impressive and a little absurd. Daily commerce on a massive scale still touches programs written in that style. When a state unemployment system buckled under a wave of claims a few years back, officials had to ask the public for people who still knew the old code. That should have been a wake-up call. For a lot of institutions it became another reason to wrap the core in more layers and hope the wrapping holds.
I do not think the savings from rewriting that stack are imaginary. I do think the industry is about to argue, quietly and then loudly, over who gets to keep them.
What The Overnight Ledger Still Does
Inside a large lender the core ledger is often a night person. The number you see at lunch can be an estimate. Bankers have a dry phrase for it. The real posting happens after the lobby is dark. A mainframe works through a queue in a fixed order. Transactions land. Interest accrues. Fees hit. Files go out so every other system can start the next morning with a shared version of reality.
Those programs talk through record layouts where a field is known by its place in a line, not by a friendly name. One taxpayer identifier can show up under dozens of labels across thousands of programs. Acquisitions made it worse. Each purchased bank arrived with its own heart. Management was usually too nervous to switch the old heart off. So the old hearts kept beating in a stack.
Replacing all of that has gone badly more often than well. One large Australian lender spent years and more than a billion in local currency and the industry still treats that project as a relative success. A British bank moved customers onto a new platform and watched the platform fail in public. People were locked out. Some saw other people’s accounts. Service took months to feel normal again. Fines and redress followed. Industry estimates put the failure rate of these modernizations somewhere around two thirds. With odds like that, most shops built scaffolding and left the foundation alone.
The app can look modern while the books still close in batch. That gap is where both the cost and the risk live.
Technology budgets at the biggest names are already enormous. One money-center firm has talked about spending close to twenty billion in a single year, with more weight now on modernizing the underlying application code and data. A large slice of that money, if I am reading the industry honestly, still goes to keeping the scaffolding upright.
Where Coding Agents Actually Help Today
The new coding agents are not magic wands for every screen a customer touches. They are strong at documentation, refactoring, and batch work. Those are the parts of a migration where you can hand the agent yesterday’s inputs and outputs and let it keep trying until the new code matches the old results. Batch can be a third to half of a typical move. That is not a small slice.
Real-time pieces are another story. Card authorizations, instant fraud checks, anything that cannot wait for a night job still sit in a harder box. I would not bet a weekend that those go first. Banks also have a security reason to hurry. Models that can hunt software weaknesses have already made supervisors hold urgent conversations. Older bank systems sit high on that worry list. If you run one of those cores, waiting starts to look expensive in a different way.
In my experience, the cleanest savings show up where the work is repetitive and testable. You prove the new job against the old files. You do it again. You do not put the agent in charge of a live authorization path on day one. That sounds cautious. It is also how you avoid becoming a case study.
Money Is A Commodity. Deposits Are Not Quite.
Here is the harder point, and it is the one I care about as an investor. Loans often behave like commodities. A company that wants a five-year term loan will collect a pile of term sheets and take the cheapest clean offer. If a bank’s costs just fell, that bank will give up some spread to win the deal. That part of the argument feels right.
Deposits have never really worked the same way. A line of research from academic economists has argued that banks hold real market power over depositors. When policy rates rise, banks lift what they pay slowly and only partway. Keeping that power costs money for branches, people, and systems, but almost all of that cost is fixed. Deposits then act like long-term fixed-rate funding. That is how a balance sheet can hold long mortgages without getting flattened every time rates jump. It is also why industry net interest margins have barely wandered across decades of cycles. Recent official readings still sit a little above three percent.
Bankers measure the leak with deposit beta, the share of a rate increase that actually reaches the customer. Checking accounts tend to have low betas. Online-only shops have high ones because those customers are hunting yield. During the last hiking cycle, super-regionals passed through more than many small banks. The very largest names passed through less than either group. After a famous regional failure sent tens of billions out the door in a single day, money moved toward size. The biggest balance sheets did not have to pay up as hard to keep it.
That is the piece a simple cost-cut story misses. Cheaper code does not automatically become thicker margin if the funding side starts to reprice.
When A Personal Agent Sees Every Account
A new class of personal agents can already read balances across thousands of banks and finance apps through existing data connections. They cannot always move the money yet. Some links are still read-only, and that detail matters. I still would not want to sit on a treasury desk pretending nothing changed.
Most households leave idle cash at a tenth of a percent because switching is a hassle. Opening a new account eats an afternoon. Nobody wants to be the person who breaks payroll or rent. If an agent already sees every balance and can fill the forms, most of that afternoon disappears. A well-known market note recently asked whether an agentic run could arrive. The gap it pointed to is blunt. Average checking still pays almost nothing. A cluster of digital lenders and cash products pay several percent. Cheap deposits are the raw material for loans. If they leave, the machine changes.
Markets already flinched one session when this idea hit the tape. Wealth platforms and large lenders sold off together. A broad financials fund dropped as well. People treated relationship banks and rate-sensitive banks as if they were the same animal. Sometimes that is lazy. Sometimes it is a warning that the first wave of agents will not pause to read a branch history.
- Operating cash tied to payroll and payables tends to stay put.
- Idle savings that nobody has checked in years is the first pile to move.
- Price-sensitive loans will absorb part of any cost win on day one.
- Vendor contracts can trap community banks even when the tools get cheaper.
A Simple Hypothetical That Stings
Imagine Bank A holds ten billion of balances that are really savings. The customers do not need the money next month. They have not looked in a while. Policy rates sit near four percent. Agents lift the beta on those balances by ten points. That is about forty basis points on ten billion, or forty million a year.
Now suppose a coding overhaul saves fifteen million a year in maintenance. That number is a guess, not a filing. Bank A is behind by twenty-five million before it gives up any spread to keep borrowers. My figures could be wrong in either direction. For a bank funded mostly by sleepy savings, I do not see the code win covering the deposit cost.
Corporate treasurers solved a version of this years ago. A company keeps enough in the operating account for payroll and suppliers. The rest sweeps into money funds or bills. Banks price those deposits knowing someone is watching. An agent gives an ordinary household a junior treasurer. Next month’s bills stay in checking. The surplus that has been sitting since the last shock probably will not stay at a tenth of a percent forever. I expect the reprice to start slow and get obvious in the next rate cycle.
Which Banks Keep The Windfall
Some franchises are safer than others. Deposits that run a business’s payroll, sit next to a credit line, or belong to an owner whose banker still picks up the phone will not jump for half a point. Banks that hold those should keep more of what cheaper code produces.
Banks that fund themselves with rate-shopping savings and win loans on price sit in a worse spot. They will likely pass savings to borrowers and pay more to depositors at the same time. That is not a morality play. It is how commodities work once the friction falls.
This is where ownership of the code matters. Big institutions often own their stack. Most American banks rent theirs. A regional reserve survey found that three processors together served more than seventy percent of banks, and a majority had used the same core provider for over a decade. When code gets cheaper for a community bank, it gets cheaper for the vendor first. Whether any of it reaches the bank depends on a contract that may still have years left and on how hard those vendors compete at renewal.
Agents do not wait for renewal dates. They reach a community bank’s depositors the same week they reach a money-center app. That asymmetry is easy to miss if you only watch efficiency ratios.
| Bank type | Likely code savings | Deposit risk |
| Large owner of core systems | High if batch work converts | Lower if operating balances dominate |
| Price-led lender | Real, then competed away | High on savings balances |
| Community bank on a rented core | Vendor captures first | Agents arrive on the same clock |
| Relationship commercial shop | Moderate but sticky | Lower if credit and service bind |
What I Would Actually Watch In A Stock
If I were looking at a bank name this month I would not spend much time celebrating a prettier efficiency ratio. Nearly everyone’s will improve if the tools work. I would look at what deposits did from 2022 to 2024. That stretch is the closest live test of stickiness the industry has had in a generation.
I would want a clean split between operating money and savings nobody has checked in years. I would want the renewal date on the core processing contract. I would want to know whether commercial customers still get a human who can solve a wire at 4:50 p.m. Those details sound dull. They decide who keeps the margin.
At smaller and mid-sized companies there is often a wide gap between what new tools can do and what the firm has actually put to work. With banks that gap only matters if the customers are still there when it closes. Cheaper code is coming. I am just not convinced most lenders get to keep much of it.
Why Legacy Code Lasted This Long
People outside the industry treat old code as sloth. That is too neat. The language survived because it did the unglamorous work of money: posting, accruing, reconciling, producing files other systems could trust. It was designed to be read. That made audits and handoffs possible. It also made change expensive, because every field position became a quiet contract with every downstream job.
Copybooks are a good example. A layout is not a modern API. It is a map of columns. Move a field and you break a chain you cannot see from the app team’s stand-up. That is why so many “modernizations” became extra layers. The core stayed. The layers multiplied. Cost rose. Risk hid in the joints.
I’ve found that institutions delay not because they love mainframes. They delay because a failed cutover is a franchise event. Customers remember the week they could not pay rent. Regulators remember it longer. Boards remember the headlines. Agents change the cost of trying. They do not erase the cost of being wrong in public.
Security Pressure Changes The Calendar
There is a second clock now. Tools that find and exploit weaknesses make old, poorly mapped systems look like open windows. Supervisors on both sides of the Atlantic have already treated that as urgent. You do not need a cinematic breach to feel the heat. You need a model that can read a forgotten interface faster than your change board meets.
That pressure cuts two ways. It argues for speed. It also argues against sloppy speed. Matching batch outputs is a controlled game. Opening a real-time path before the map is complete is how you mint operational risk. The winners will be the shops that use agents for the testable middle and keep humans on the edges that still scare me.
Fixed Costs, Market Power, And Why Margins Look Calm
The deposit-power story is easy to forget when rates are falling and everyone is hunting yield products. It matters more when you ask who keeps an AI dividend. Branches, relationship managers, and core systems are mostly fixed. Once those are paid for, an extra dollar of sticky deposits is extremely valuable. That is why margins can look oddly stable across cycles even as headlines scream about rate shocks.
If agents punch a hole in that stickiness, the fixed-cost machine still sits there. You do not instantly shrink the branch or the vendor invoice. You just pay more for funding while you wait for the next renewal or the next efficiency program. That lag is where equity stories get sloppy. A slide can show code savings this year. The deposit bill can arrive on a different calendar.
Cheap deposits are not a rounding error. They are the quiet subsidy that lets long assets live on short-looking liabilities.
Perhaps the most interesting aspect is how uneven the damage would be. A bank with payroll operating accounts and unused credit lines has a moat that does not show up in a beta average. A bank that gathered pandemic leftovers through a national ad campaign has a pile that looks stable until it is not.
Vendors, Contracts, And The Community Bank Bind
Community banks love to talk about local knowledge. I believe a lot of that talk. I also believe a ten-year core contract can turn local knowledge into a leased utility. If three processors dominate the survey data, pricing power sits upstream. Agents that rewrite batch jobs help the party that owns the code and the change tickets.
At renewal the conversation changes. A bank can demand a share of the productivity. Vendors can point to security, uptime, and conversion risk. Some of the savings will leak to banks. Some will stay in processor margins. Some will be competed away when the next processor tries to steal a logo. None of that is automatic, and none of it is timed to the week an agent finishes a refactor.
Meanwhile the household agent does not need a conversion weekend. It needs a login and a form. That is the mismatch I cannot shake. The cost curve and the funding curve do not move on the same clock.
Borrowers Will Ask For Their Cut
Even if deposits hold, loans may not. Credit is shopped. Spreads compress when a lender’s cost of doing business falls and rivals can see it. A five-year term loan is not a checking account. The treasurer on the other side of the table has a mandate. Give that person a reason and they will re-trade you.
Consumer loans have more friction, but comparison tools already exist. If agents start filling applications the way they fill account openings, the remaining friction drops again. I do not expect overnight chaos. I expect a grind. Basis points leave through the loan door while other basis points leave through the deposit door. The code line on the expense budget looks prettier either way.
What “Memo Posted” Teaches About Trust
The noon balance is a story the bank tells you so you can spend. The night job is the story the bank tells itself so it can survive an audit. Those two stories have been allowed to drift because customers rarely notice until something breaks. Agents that read every account make the drift more visible. If one product still pays almost nothing and another pays several percent for cash that is economically the same, the agent does not need a manifesto. It needs a rule.
That is why I keep saying the first move is not a classic run. It is a tidy sweep of surplus. Payroll stays. Rent stays. The leftover from the last boom does not. Banks that confuse those piles will tell themselves a soothing story about loyalty until the monthly funding report turns.
A Practical Checklist Without The Slideware
- Split deposits into operating cash and idle savings with honest definitions.
- Track beta by product, not by a single blended number that hides checking.
- Map how much of the tech budget is scaffolding versus core rewrite.
- Write down the core vendor renewal date and the exit cost.
- Test whether commercial relationships still include a reachable human.
- Ask what share of loans is won on price alone.
- Assume agents reach your depositors before your conversion finishes.
None of that is glamorous. It is how you avoid buying a narrative that treats every efficiency point as equity value.
The Human Texture That Models Skip
Switching still has a face. Direct deposit. Autopay. The fear of a missed mortgage draft. Older customers who want a person at a desk. Small firms that keep a lender because the last covenant waiver arrived on time. Those frictions are real. They are also shrinking. An afternoon of paperwork used to be a moat. Forms that fill themselves are a bridge.
I do not buy the idea that every household becomes a hedge fund overnight. People are busy. People are loyal in small ways. People also hate leaving free money on the table once someone points at it every week. The agent does not need to shout. It needs to be right twice, then it becomes furniture.
That is the cultural shift hiding under the tech story. Corporate treasurers already live this way. Families are being offered a lighter version of the same discipline. If that spreads, deposit market power does not vanish. It narrows. The banks that keep it will be the ones whose products are tangled with daily life, not the ones whose rates simply lagged.
Where The Argument Still Feels Thin
I should say what I do not know. I do not know how fast agents will be allowed to move cash. Read-only is a real constraint. Compliance, fraud, and bank partners can slow the last mile. I do not know how regulators will treat automated hopping if it starts to look like coordinated flight. I do not know whether vendors will cut prices or wrap the same invoice in a new “AI modernization” line.
Those unknowns cut both ways. They can protect incumbents longer than a viral clip implies. They can also hide a delayed break. Markets often sell the idea first and the cash later. That session where financials dropped together felt like the idea trade. The cash trade will be quieter and meaner.
A Longer View Of Who Gets Paid
If you zoom out, four parties stand in line for the money. Equity holders want a lower expense ratio. Depositors want a fairer yield. Borrowers want a tighter spread. Vendors want to keep the rent on the rented heart. AI coding strength does not pick a winner by itself. Market structure does.
Own the code, own sticky operating balances, and own a loan book that is not a pure auction, and you have a chance to keep a slice. Rent the code, fund with sleepy national savings, and win credits on price, and you become a pass-through. That is not a forecast of collapse. It is a forecast of thinner leftovers.
I keep circling the same sentence. The rewrite can be real and still not belong to the bank. That is the part the victory lap leaves out.
Closing The Loop Without A Slogan
The industry spent decades treating the core as too dangerous to touch and too valuable to replace. Agents change the first half of that sentence. They do not automatically change the second. Customers, contracts, and loan auctions still sit between a cleaner codebase and a fatter return on equity.
If you work inside a bank, the useful question is not whether the overnight jobs can be rewritten. Plenty of evidence says the testable parts can. The useful question is which balances will still be there at the old price when the rewrite is done. If you look at stocks, skip the victory slide on efficiency. Read the deposit mix. Read the vendor date. Read who still answers the phone.
Cheaper code is coming either way. The leftover cash will not land in a single pocket. It will be bargained over, slowly at first, then in the next rate cycle when nobody can pretend the surplus was invisible.
This discussion is for research and general information only. It is not investment advice and it is not an offer to buy or sell any security. Bank examples used as illustrations are hypothetical unless described as industry history in broad terms.
