How AI Helps With Mainframe Modernization’s Hardest Problem
In a computing environment as longstanding and complex as the mainframe, the challenge facing modernization projects is understanding the code, not transforming it—and AI is here to help
By Andrew Wig
How AI Helps With Mainframe Modernization’s Hardest Problem
In a computing environment as longstanding and complex as the mainframe, the challenge facing modernization projects is understanding the code, not transforming it—and AI is here to help
By Andrew Wig
“It was always, roll up the bus full of consultants and let them have at it for many, many months,” says Mark Sigler, a senior BMC product director who has seen his share of modernization projects in his 45 years as a mainframer.
“And oftentimes, it would fail.”
Can AI do any better? The mainframe industry is trying to find out.
A New Hope
In BMC’s 2025 State of the Mainframe Survey, 65% of 1,100 respondents said they were already using generative AI on the mainframe, while 74% considered it critical to their mainframe strategy over the next two years. Meanwhile, the mainframe modernization market continues to grow. Valued at about $18 billion in 2025, it’s projected to reach nearly $21 billion in 2026.
Enterprises that were once humbled by failed modernization projects in the pre-AI era may now be enticed to give it another go, now that AI exists as a tantalizing—albeit nascent—option. The promise of the technology is top-of-mind for IT leaders, but part of the modernization consultant’s job is to bring those decision makers back down to earth.
“I think there is an expectation that AI is the pixie dust that fixes everything,” Sigler says.
Enterprises may feel like that’s what it takes to modernize their tangled, monolith applications whose documentation only lives inside the heads of retired COBOL programmers.
AI and Modernization by the Numbers
$18B
→$21B
Sources:
1 BMC’s 2025 State of the Mainframe Survey.
2 Research and Markets’ Mainframe Modernization Market Report 2026
AI has accelerated code conversion.
Choosing what to convert, and when, remains the harder problem.
- Analyze
- Build
- Convert
- Measure
Explore the BMC approach to code modernization →
It’s Not About the COBOL
When IBM released watsonx Code Assistant for Z in 2023, it wasn’t its code generation capabilities that customers were excited about; it was the tool’s ability to understand code, a key aspect of modernizing monolith applications, notes Pete McCaffrey, principal product management lead, IBM Z.
“This code translation bit—it’s like 10% of the total job and decisioning as it relates to mainframe modernization,” McCaffrey says.
Wise modernization choices require an understanding of deeply opaque code bases. “The hard bit is the fact that you might have business logic that nobody’s touched or commented in 15, 20 years, and you’ve not been servicing and maintaining your code,” says Matt Whitbourne, VP of product management and design for BMC’s Automated Mainframe Intelligence (AMI) portfolio.
In that work, Whitbourne is more focused on explanation than conversion, because the former informs the latter—understanding what the code does helps a shop determine whether it needs to be refactored, converted or left alone.
It’s not really the coding language that needs to be modernized anyway, Whitbourne says, arguing that code in any modern language will be challenging to work with if left alone for an extended period.
“Go pick whatever’s the flavor of the month right now, don’t touch it for 15 years and don’t put the investment in to maintain it, and then you’ll have exactly the same problem in 15 years’ time,” he says. “It’s not the language; it’s that the business logic hasn’t really been modernized and maintained.”
It’s Still Early
- ‹ prev
- next ›
- 1
- 2
- 3
- 4
- 5
- 6
- 7
New tools entering the mainframe ecosystem are addressing the trickier parts of modernization. In addition to refactoring and translation, these tools can help developers understand legacy code, document it, decompose monolith code bases, generate test cases, assess risk through complexity analysis and automate CI/CD pipelines.
So far, the results of modernization projects suggest that the AI component is still in the early stages of implementation. Sigler says he is “yet to see a really large application go all the way through a modernization process” via AI. That includes “some of these are really well-heeled financial services companies that have good talent, folks that really understand AI and what it can and can’t do.”
Sure, there are stories, he adds. “I’ve heard of Bigfoot too, but I haven’t seen it yet.”
One gnarly beast he has seen is “Jobal,” the unmaintainable code that results from naive COBOL-to-Java conversion. “Jobal” doesn’t make a lot of sense for developers; plus, Sigler says, it suffers from architectural shortcomings that stem from translating a procedural language like COBOL into an object-oriented language like Java.
That challenge may be manageable in limited scope, he says. But when you’re trying to move a 6,000-program accounts-receivable application to the cloud, “it’s a lot of code.”
Not all COBOL needs modernizing anyway. Whitbourne recalls one customer who was experimenting with COBOL-to-Java conversion and ended up with 10x more code volume in the Java equivalent. “They’re like, ‘Well, I’ve got to service all of this now, and I know the performance that I used to get was predictable and it was bullet-proof,’” he illustrates.
The Mainframe’s Unique, Baffling Architecture
When Anthropic highlighted Claude’s ability to convert COBOL to Java in a highly publicized blog in February, the mainframe community responded by noting that mere code conversion is insufficient to modernize applications that have lived on a platform as architecturally complex as the mainframe. Mainframe environments tend to have decades of business logic layered across programs full of non-linear control flows and undocumented edge cases.
“There’s so much that’s built into the operating system and middleware that doesn’t exist, really, on other platforms, or it certainly doesn’t exist in the same kind of way,” Whitbourne says. “ … If you go and do the conversion from one language to another, all your hard work is still ahead of you, really. Because all those API calls that I used to depend on for using things like CICS and IMS, which are just fundamental transaction processing software that you get on the mainframe today—how are you reproducing that? How are you making sure I can still get to seven, eight ‘9s’ of availability, which is like 0.3 seconds of downtime a year?”
And how are you overcoming the scattered nature of your data? That’s another real challenge, says Dave Dalton, director of product management, BMC AMI Data. “It’s not legacy; it’s not old tech. It’s silos that are the enemy of success here,” he contends. “And Agentic AI with technologies like MCP starts to chip away at those silos in a way that we’ve probably not seen before.”
The Modernizer’s Checklist: Questions Code Conversion Alone Can’t Answer
New Technologists, New Technology
Mainframers charged with maintaining decades of legacy code tend to be more risk-averse than those on distributed platforms, especially those starting from scratch on greenfield coding projects.
“These applications are so important that there’s a nervousness in making changes, because you don’t want to be the person that screws it up and takes down the whole business,” McCaffrey says.
Their caution may extend to AI tools, but this group is also retiring, leaving behind younger technologists who have never used a green screen. BMC’s 2025 State of the Mainframe Survey suggests the generational handover is already complete.
In 2018, the share of mainframers aged 18–49 was relatively even with the 50-and-up demographic. By 2025, the younger group comprised 80% of the mainframe workforce.
The Generational Handoff
2018
2025
Mainframe boosters hope AI tools, along with initiatives to develop young mainframe talent, can help bridge the generational skills gap. “I think as an industry, we anticipated that a little too late, but soon enough to still do something about it,” Dalton says.
‘They’ve Got the World at Their Feet’
Among those taking over the platform, the 30-somethings who got their start before the proliferation of generative AI tools are in an enviable position, Dalton believes. That group, he notes, is expert at Python, is knowledgeable about mainframe technologies and modern approaches, and is using AI tools to build next-gen DevOps pipelines.
“But even before that, they were the first people to really lean in on Jenkins and those kinds of technologies on the mainframe,” he says. “So they’ve got the world at their feet, quite frankly. They’ve got all of these terabytes of critical business data and business processes that are running the planet.”
As they tackle long-overdue modernization projects, they are positioned to leverage those resources while blending their own mainframe knowledge with new AI tools—a potent combination.
