Who Keeps The Money When AI Rewrites Bank Code?

Authored by Patrick Feeley via Substack,

The code most American banks run on was designed in 1959, the year Alaska and Hawaii became states. A committee of government and industry people wrote COBOL so that business programs could be read by people who were not mathematicians, and a good part of it was modeled on FLOW-MATIC, an earlier language from Grace Hopper, a Navy officer. I doubt anyone on that committee thought it would still be running banks in 2026. In 2017 Reuters estimated that about $3 trillion of daily commerce still ran through COBOL. In April 2020, when unemployment claims in New Jersey overwhelmed the state’s forty-year-old system, Governor Phil Murphy went on television and asked for volunteers who knew COBOL. The state had to go on TV to find programmers for its own unemployment system.

I bring this up because of a clip I posted last night of Bill Ackman talking with Shane Parrish on The Knowledge Project. Ackman said Cognition, the company behind the coding agent Devin, can rewrite a bank’s COBOL “in a matter of days as opposed to many months.” Someone replied asking me what I meant when I said commoditized lenders would compete the savings away. It is a fair question and I could not answer it in a tweet.

I do believe Ackman that the savings are real. Inside a large bank the core ledger still runs in batch. The balance a customer sees on the app at noon is an estimate (bankers call it memo-posted). The actual accounting happens overnight, when a mainframe works through a queue of jobs in a set order, posting transactions, accruing interest, charging fees, and producing files that every other system reads the following morning. The programs share data through copybooks, which are record layouts where a field is known only by its position. Cognition described one client where a single taxpayer ID showed up under dozens of different names across thousands of programs. Most big banks were put together through acquisitions, and each acquired bank came with its own core system that management was usually too nervous to shut off, so the old systems just piled up.

.@BillAckman tells @shaneparrish that AI will sharply cut what big banks spend on tech. Rewriting legacy COBOL (decades-old bank code) with @cognition now takes days instead of months. Ackman is right that the cost savings are real. The open question is who keeps them. Banks with …

— Patrick F. Feeley (@PFFeeley) September 30, 2026

Replacing all of that has gone badly more often than well. Commonwealth Bank of Australia spent five years and more than A$1 billion replacing its core, and people in the industry consider that one a success. TSB in the UK moved customers onto a new platform in April 2018 and the platform did not work. Customers were locked out, some could see other people’s accounts, and service was not back to normal until December. TSB ended up paying £32.7 million in redress and £48.65 million in fines. Cognition’s own figure is that roughly two-thirds of COBOL modernization projects fail. With odds like that most banks built layers around the old core and left it alone. JPMorgan expects to spend about $19.8 billion on technology in 2026, and its CFO told investors in February that the priority had moved to “modernizing the underlying application code and data.” I would guess a large share of that budget still goes to keeping the layers standing.

Cognition is fairly careful about what its agents can do today. Devin is good at documentation, refactoring, and batch jobs, which are the parts of a migration where you can give the agent yesterday’s inputs and outputs and let it keep trying until the new code matches the old results. Cognition estimates batch is 30 to 50 percent of a typical migration. The real-time systems (card authorizations, for example) are still out of reach. Banks also have a security reason to hurry. Anthropic’s Mythos model, which can find and exploit software vulnerabilities, had bank regulators in the U.S. and Europe holding urgent calls this spring, and Reuters quoted security experts who named legacy bank systems as especially exposed.

Ackman’s harder point came a little later in the conversation. “The problem with money generally is it’s a commodity,” he said. For loans I agree with him. A company that wants a five-year term loan will collect six or seven term sheets and take the cheapest one, and a bank whose costs just went down will give up some spread to win it. Deposits have never really worked like a commodity, and I think that is where his argument is missing a piece.

The best explanation I have read is from three NYU economists, Itamar Drechsler, Alexi Savov and Philipp Schnabl. Their paper argues that banks have real market power over deposits. When the Fed raises rates, banks raise what they pay depositors slowly and only partway. Keeping that power costs money for branches, bankers and technology, but almost all of the cost is fixed. So deposits end up behaving like long-term fixed-rate funding, which is how a bank can hold thirty-year mortgages without being wiped out every time rates go up. It is also why the industry’s net interest margin has barely moved over several decades of rate cycles. The FDIC has it at 3.32 percent today.

Bankers measure this with the deposit beta (the share of a rate increase that gets passed along to depositors). Checking accounts have low betas. Online banks have high ones because, as the St. Louis Fed put it, their customers are looking for yield. During the 2022 hiking cycle the New York Fed found that super-regional banks passed through more than small banks did, while the very largest banks passed through less than either. After Silicon Valley Bank lost $42 billion in deposits in one day, money moved toward size, and the biggest banks did not have to pay more to get it.

Meta’s Muse goes right at this. It launched September 8, the same day Cognition announced it had raised more than $2 billion at a $48 billion valuation with run-rate revenue near $900 million. Muse is a personal agent that reads accounts at more than 12,000 U.S. banks and financial apps through Plaid. On Tuesday, September 22, Schwab fell 6 percent, LPL fell 7 percent, JPMorgan and Wells Fargo each fell more than 3 percent, and XLF, the largest financials ETF, was down 2 percent. On Sunday Torsten Slok at Apollo put out a note titled “Is an Agentic bank run coming?” He pointed out that the average checking account pays about 0.1 percent while Revolut, SoFi, Wealthfront and others pay between 3.3 and 5 percent, and he warned that banks “could lose a large share of the cheap deposits they rely on to make loans.”

Muse cannot move money yet. The Plaid connection is read-only, and Meta deserves to have that said. I still would not want to be running a bank’s treasury desk this month. Most people leave savings at a tenth of a percent because switching is a hassle. Opening a new account takes an afternoon, and nobody wants to be the person who breaks their own direct deposit. If an agent already sees every balance and can fill out the forms, most of that afternoon goes away.

A hypothetical helps here. Bank A has $10 billion of deposits that are really savings, money the customers do not need next month and have not looked at in a while. Rates are at 4 percent, and agents push Bank A’s beta on those balances up by ten points. That costs Bank A about 40 basis points on $10 billion, or $40 million a year. Say Cognition saves Bank A $15 million a year on code maintenance, which is my guess and not a published number. Bank A is behind by $25 million, and that is before counting any spread it gives up to keep borrowers. My numbers could easily be off in either direction, but for a bank funded mostly by savings I do not see the code savings covering the deposit cost.

Corporate treasurers dealt with this a long time ago. A company keeps enough in its operating account for payroll and suppliers and sweeps the rest into money market funds or Treasury bills, and banks price corporate deposits knowing somebody is watching. An agent gives an ordinary family something like a corporate treasurer. Next month’s bill money will stay in checking. The surplus that has been sitting there since the pandemic probably will not stay at 0.1 percent, and I expect it to reprice slowly at first and much more visibly in the next rate cycle.

Some banks are safer than others. Deposits that run a business’s payroll and payables, carry a line of credit, or belong to an owner whose banker actually picks up the phone will not move for half a point, and banks holding those should keep most of what Cognition saves them. Banks that fund themselves with rate-shopping savings and win loans on price are in a worse spot, since they will likely pass the savings on to borrowers and pay more to depositors at the same time.

This part relates most to my own work. Ackman was talking about big institutions that own their code. Most American banks rent theirs. The Kansas City Fed found that Fiserv, Jack Henry and FIS together served more than 70 percent of banks in its 2022 survey, and 61 percent of banks had used the same core provider for over ten years. 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 run for years and on how hard those three companies compete at renewal. Agents do not have to wait for any contract, and they will reach a community bank’s depositors the same week they reach JPMorgan’s. On September 22 the market sold banks with deep local relationships about as hard as it sold the ones without them, and in some cases I think that was a mistake.

If I were looking at a bank stock this month I would not spend much time on efficiency ratios, since nearly everyone’s will improve. I would look at what the bank’s deposits did from 2022 to 2024, which is the closest thing to a live test of stickiness the industry has had. I would want to know how much of the deposit base is operating money and how much is savings nobody has checked in years, and I would want the renewal date on the core processing contract. At Sargasso Capital Management we spend most of our time on small and mid caps where there is a wide gap between what AI can do and what the company has actually put to use. With banks that gap only matters if the customers are still there when it closes. Bill Ackman is right that the code is about to get a lot cheaper. I am just not convinced most banks get to keep much of it.

A note on sources. Bill Ackman’s comments are from his September 2026 appearance on The Knowledge Project with Shane Parrish. Cognition figures come from its September 8 funding announcement as reported by Reuters, TechCrunch and SiliconANGLE, and from its April 2026 post on COBOL modernization. Muse details come from Meta’s launch materials and coverage by TokenPost, CNBC and Zacks. The Torsten Slok note was reported by CoinDesk on September 28. COBOL’s origins and New Jersey’s April 2020 call for COBOL programmers are widely documented. Banking data comes from the FDIC Quarterly Banking Profile for the second quarter of 2026, the Federal Reserve Banks of New York, St. Louis and Kansas City, the Federal Reserve’s review of Silicon Valley Bank, JPMorganChase’s February 2026 Company Update, the FCA’s TSB enforcement notice, Reuters reporting from April 2017 and April 2026, and Drechsler, Savov and Schnabl, “Banking on Deposits” (Journal of Finance, 2021). Bank A is a hypothetical example.

This post is for informational and research purposes only and does not constitute investment advice or an offer to buy or sell any security.