(Bloomberg) — Kresimir Mudrovcic and his team of programmers spend months on end trawling through computer code that can be three times as old as the crew’s youngest members.
Mudrovcic specializes in mainframe technology, involving computers tracing their roots to the dawn of the digital age and the ancient software that sometimes runs on them. Upgrading such systems is painstaking work, often entailing sifting through millions of lines of code to understand how specific functions operate. Mudrovcic, an IT consultant, compares it to archeology.
But the work is getting easier, thanks to the widening use of generative artificial intelligence to do some of the heavy lifting. “AI systems work like that smart, very experienced, very wise old colleague who knows everything,” said Mudrovcic, whose team recently deployed such tools to help speed up modernizing the pension system of a European government agency.
Similar efforts are underway at companies and governments around the world, as the urgency to address aging computer code increases. The US Social Security Administration plans to use AI to help upgrade its legacy Common Business-Oriented Language code base, and expects the project to take three years and cost about $1 billion, a person with knowledge of the matter said.
The SSA didn’t respond to requests for comment.
From online banking applications and airline ticketing services to pensions disbursements, critical systems are often undergirded by decades-old code, raising costs as well as the risk of failures and cyberattacks. US Treasury Secretary Scott Bessent has repeatedly stressed the need to overhaul government systems running on computer language like COBOL, which was invented in the late 1950s.
“When I started in college in 1980, I learned to program in COBOL,” he told Bloomberg’s Big Take DC podcast in February.
As much as 70% of software used by Fortune 500 companies was developed at least two decades ago, according to a December report from McKinsey. Global financial institutions alone are expected to spend some $57 billion maintaining legacy payments systems in 2028, research firm IDC estimates. That’s almost equivalent to last year’s net income at JPMorgan & Chase Co., the biggest US bank.
“You’d be surprised how many firms are still on COBOL, even banks globally,” said Gokhan Sari, a senior partner at McKinsey. The consultancy has developed a dedicated AI tool called LegacyX to help clients including banks remove obsolete code and revamp their systems.
Until Sam Altman’s OpenAI ushered in the generative AI frenzy with ChatGPT in late 2022, managing systems based on dated languages like COBOL and PL/1 meant tapping into a dwindling pool of talent as experienced programmers versed in such code retired. It wasn’t unheard of for companies to call former employees back from retirement when a system needed maintenance.
This is a particularly thorny problem for banks, whose consumer-facing applications can run on top of decades-old systems, putting them at a disadvantage to upstart fintech competitors like Revolut Ltd. So tangled are the various layers of code, they’ve often been referred to as “spaghetti systems.” And replacing them is delicate work.
“There are often a few people who know the code, it’s thousands of lines and you need to migrate to a new language,” said Görkem Köseoğlu, the former chief technology officer at Dutch banking group ING Groep NV. “These are risky and complex projects because while migrating to new code base, you also need to keep the lights on at the banks.”
Engineers have started using AI tools like ChatGPT, Microsoft Corp.’s Github CoPilot and IBM’s watsonX to make the process of maintaining and upgrading old systems easier and quicker.
Instead of going through it line by line, a programmer can upload or copy and paste large blocks of code into an AI tool, along with natural-language prompts like “What does this COBOL program do?” The AI can then explain how the code works and how different parts fit together — and even translate old computer languages into more modern ones like Java.
Since legacy systems often lack documentation, engineers are also using AI to create instruction manuals describing how they operate. That can help organizations avoid the time-consuming effort of reverse-engineering old code when they need to change or replace it.
The time and cost savings can be dramatic. The cost of modernizing a transaction processing system at a large financial institution, which three years ago would have run to more than $100 million, is now “well less than half” of that using generative AI, according to the McKinsey report.
It’s “considered a game changer comparable to the advent of cloud computing 20 years ago,” said Michal Paprocki, chief information officer at Brussels-based Euroclear SA.
Social Security Controversy
That doesn’t mean AI is poised to eliminate the need for human intervention. One potential risk is “hallucinations,” where the AI models invent fake answers — something that could have steep consequences when dealing with critical systems. For the foreseeable future, experienced engineers will need to oversee the work.
Efforts by the Elon Musk-led Department of Government Efficiency to revamp the Social Security Administration’s IT systems have already become a source of controversy, amid fears that moving too fast might create havoc.
On April 17, Democrats on the House Oversight Committee asked SSA’s assistant Inspector General, Michelle Anderson, to investigate whether plans to quickly overhaul the agency’s code could disrupt payments to more than 70 million beneficiaries. Representative Gerald Connolly of Virginia said he had information from an agency whistleblower that SSA planned to replace COBOL code in a matter of weeks or months and “rapidly rip out and replace critical IT systems without adequate planning and preparation.”
To a large extent, the aging software infrastructure that underpins global commerce has held up well. But the layering of successive iterations of increasingly advanced code to run more demanding functions is starting to strain computer networks, making them a financial drag for companies and governments alike.
A 2022 report from the Consortium for Information & Software Quality estimated the “technology debt” accumulated from poor-quality software — essentially the spending that would be required to replace it — at $1.5 trillion.
The potential benefits of upgrading or retiring legacy systems go beyond cost, in the form of freeing up resources for developing new tools and applications and making it easier to attract the best engineering talent, executives said. Programmers, in general, prefer to work with the more modern computer languages they’ve been trained in.
At Euroclear, which provides post-trade services, legacy systems are affectionately referred to as “legendary systems” since they perform the vital function of keeping record of the €41 trillion ($47 trillion) of assets it holds in custody. Now Euroclear is testing AI to help automate creating a record of how its existing code works — an attempt to future-proof its systems as engineers who know the older programs retire.
“It’s like having a smart colleague,” said Jaques Theys, head of advanced analytics and Business Intelligence at Euroclear. “Two years ago, [AI tools] were like the new intern. But the more they go, the smarter they get.”
–With assistance from Gregory Korte.
More stories like this are available on bloomberg.com
