Core banking systems rarely get replaced. Banks patch and wrap them instead, keeping them alive for decades after the people who built them have left the building.
Mitch Siegel, Principal and Partner for Financial Services Consulting at EY, spent 30 years advising banks on legacy and payments technology. He told me in our conversation for the Banking Reinvented podcast that the oldest core system he's come across is 57 years old.
"There's not a lot of people out there today that know COBOL. So it creates a pretty complex environment for change," he told me.
Roughly four in ten banking systems are still running on COBOL today, a share that's stayed roughly flat for years. Every bank running a system like that eventually has to act on it. What determines the outcome is how that action gets carried out.
Big bang vs. phased: two ways to fix it
Banks call these two paths big bang and phased. Big bang replaces everything in a single cutover. Phased migration replaces it piece by piece, proving each part before moving to the next.
Many imagine the solution to be a fundamental overhaul of the core or a swift cutover. This can mistakenly lead decision-makers to assume they need to wait for the perfect moment, or the perfect technology, before making the change, ignoring the cost of waiting that piles up in the meantime.
The AI shortcut fantasy
Anthropic's announcement in February fed exactly that fantasy: a single tool compressing years of work into a moment. The company said its Claude Code tool could speed up COBOL modernization work. IBM's stock dropped 13% in a day, its worst single-day fall since 2000. Wall Street read it as the beginning of the end for the mainframe.
IBM's response came fast. "Translating COBOL is the easy part," the company said. "The real work is data architecture redesign, runtime replacement, transaction processing integrity, and hardware-accelerated performance built over decades of tight software and hardware coupling." Most of what a bank runs is the decades of core banking integration wrapped around the COBOL: connections to every channel and system that depends on it.
Mitch said the same thing during our conversation. Bringing data scattered across an organization into any real source of truth isn't a quick fix. "That's not a six month journey. That's a multi-year journey."
TSB vs. Commonwealth Bank: same problem, two outcomes
In 2018, TSB, a UK retail bank, moved five million customer accounts to a new platform in a single weekend cutover, rather than phasing the migration. It failed. Customers were locked out. Fraud spiked, and the disruption wasn't fully resolved until December 2018. TSB's total cost, once fines and compensation were counted, landed around £400 million.
CBA took the opposite path on the same kind of problem. Starting in 2008, it began replacing the legacy core it had run since the 1960s, moving onto the new platform product by product instead of overnight: term deposits and retail accounts first, business accounts next, lending accounts last.
The program was budgeted at $580 million over four years. By 2011 it had grown to $1.1 billion, and some program retrospectives put the full scope closer to $1.4-1.5 billion.
In 2012, CBA's CEO stood in front of shareholders and called it done: five years, over $1 billion, every retail and business customer moved. The bank proved each piece before moving to the next, without the outages or fraud spike that hit TSB six years later.
What TSB and Commonwealth Bank actually prove
For years, the industry treated full replacement as the real fix and staged modernization as the fallback for banks that couldn't afford to do it properly. The data doesn't support that hierarchy anymore. Bain's own piece on de-risking bank transformations makes the case for the alternative directly: phased, staged migration lets a bank "decommission its old platform without the make-or-break risk of a big-bang migration," while replacing core systems the traditional way "is costly, complex, and time-consuming." It's the same logic behind progressive modernization: change the highest-value piece first and prove it works before moving to the next.
Here's my take: modernizing a bank means proving one piece at a time. That's the whole game, whether or not the COBOL debate ever gets settled.
The oldest systems Mitch spoke about will likely still be running in some form years from now, and that's fine. The hope here is smaller than a final fix: every bank proves one more piece works before it moves to the next, modernizing in phases rather than waiting for one perfect leap, the same way Commonwealth Bank did.
What this means for banks running COBOL today
This is the bet behind a Banking OS: connect to the core a bank already runs, and prove value one journey at a time. Digital Banking and Agentic Banking both work this way, landing on what a bank has today instead of waiting for a rewrite that may never come.
Whether AI compresses a COBOL migration from years to quarters is still an open question, worth watching closely.
Frequently asked questions
What's the difference between big bang and phased core banking migration?
Big bang replaces the entire core in a single cutover. Phased migration replaces it piece by piece, testing and proving each part before moving to the next.
Can AI actually modernize COBOL systems?
AI tools like Claude Code and IBM's Watsonx Code Assistant speed up the analysis and translation phases of COBOL modernization. They don't remove the harder work: data architecture redesign and the runtime behavior a bank's core has to guarantee.
Why did TSB's core banking migration fail?
TSB moved five million customer accounts to a new platform in a single weekend cutover in 2018. The new system failed on launch. Customers were locked out for weeks and fraud spiked before the disruption was fully resolved in December 2018.
How did Commonwealth Bank succeed where TSB failed?
Commonwealth Bank of Australia replaced its legacy core starting in 2008, migrating product by product over five years instead of in one event. The bank declared the program complete in 2012, without the outages or fraud spike that hit TSB.
How long does core banking modernization usually take?
Enterprise-scale programs on core banking or COBOL-based systems typically run two to five years when done in phases. Big bang attempts promise a faster timeline but carry far higher risk of failure.
