A language born in 1959 still sits behind many of the payments you make today. Yet every few years a bank promises to leave it behind. So why is that so hard, and what does the bill actually look like?
Why COBOL Refuses to Retire
It Works, and It Scales
First, COBOL and the mainframes beneath it were built for one job: moving huge volumes of records without errors. Banks have tuned that job for decades. Moreover, the hardware keeps evolving, since IBM launched its z17 mainframe in April 2025 with an on-chip AI accelerator. In other words, the platform is not standing still, even if the language is old.
The Code Is the Business
Next, consider what the code contains. Over the years, it absorbed every interest rule, fee exception and regulatory fix. Much of that logic is undocumented, and the people who wrote it have often retired. As a result, “rewriting” really means rediscovering what the bank does before you can rebuild it.
A Shrinking Talent Pool
Meanwhile, the workforce is aging. A widely cited 2017 Reuters report (republished on Sott.net) said COBOL programmers are most likely to be between 45 and 55 years old. That figure is dated, but the trend it describes has not reversed.
The Footprint in Numbers
The statistics below appear in almost every COBOL article. They trace back to that 2017 Reuters coverage, and they were repeated in industry write-ups and vendor blogs. Therefore, treat them as estimates, not audited facts.
The Real Migration Math
Commonwealth Bank: The Slow Success
Australia’s Commonwealth Bank (CBA) is the best-known case of a big bank actually finishing. In April 2008 it announced a plan to replace systems over 45 years old with SAP, with Accenture as partner, at a forecast cost of around A$580 million over four years. It also chose to phase the move and keep its existing customer and internet systems, to reduce risk.
However, the budget did not hold. Finextra reported a A$150 million increase in 2009 to include Bankwest and ASB, then a further one-year delay and spending above A$1 billion. Later, iTnews quoted the project at A$1.1 billion. A consultancy case study puts the final budget near A$1.5 billion over 2008 to 2013.
Put simply, A$580 million became roughly A$1.5 billion, about 2.6 times the original plan. Still, this is the good outcome, because the bank finished.
TSB and Postbank: When Go-Live Goes Wrong
Two further cases show the downside. Neither was strictly a “COBOL to Java” rewrite. Nevertheless, both were core banking migrations, so they illustrate the same risks.
| Case | What happened | Reported cost |
|---|---|---|
| TSB (UK), 2018 | All 5.2 million customers were moved in one weekend. Around 1.9 million were locked out, and problems lasted months. | £330 million reported by TSB in 2019, plus £48.65 million in fines and £32.7 million in compensation |
| Deutsche Bank / Postbank (Germany) | The first project, Magellan, was stopped in 2015 after about €1 billion. A second try, Unity, moved 12 million customers by 2023 but caused “unacceptable” disruption. | Over €1 billion on Magellan; disruption costs above €100 million, per Banking Dive |
The TSB fine and compensation figures come from regulators, and the £330 million figure comes from TSB itself, as relayed by a secondary summary. Some of the compensation may overlap with that total, so do not simply add them up.
Notably, an independent analysis of the Slaughter and May review found that the single-weekend approach was never seriously challenged at board level. In short, the plan failed before go-live.
Can AI Change the Math?
On February 23, 2026, Anthropic published a blog post arguing that AI tools can automate the exploration and analysis phases that consume most of a COBOL project. It claimed teams could modernize “in quarters instead of years”. Investors reacted at once, and IBM shares closed nearly 13% lower that day. (Full disclosure: Anthropic makes the Claude assistant, so weigh that claim as a vendor’s own.)
However, analysts urged caution. Futurum’s review noted that AI speeds up code analysis, but regulated banks still need audit trails and proof that the new system behaves exactly like the old one. Consequently, AI may shrink the discovery bill, while testing and cutover risk remain.
What We Have Learned
COBOL survives in core banking because it is reliable, deeply tangled with business rules, and backed by mainframes that keep improving. The often-quoted footprint numbers (95% of ATM swipes, 43% of banking systems) are 2017 estimates, so use them with care. Migration math is harsh: CBA’s project grew from A$580 million to roughly A$1.5 billion, while TSB and Postbank show how a rushed cutover can cost hundreds of millions and years of trust. AI promises cheaper analysis, yet it has not removed testing and regulatory risk. In practice, phased migration and honest budgeting beat a big-bang rewrite.
