Skip to the stories
Windows & PowerShell

Agentic AI Could Shrink COBOL Migrations From Months to Weeks

  • Myles Suer
  • 948 views

OutSystems says agents can cut COBOL-to-low-code migrations from six-plus months to eight weeks. Mainframe teams should take the claim seriously and test it.

Agentic AI Could Shrink COBOL Migrations From Months to Weeks

Agentic AI is getting attention for SaaS, but the more interesting question for infrastructure teams is what it does to the COBOL estate and the mainframes that run it. My read is that the threat to these systems is real, though the timelines are likely to be longer and messier than the vendor pitches suggest.

Why SaaS isn't the only story

Earlier this year, commentators started using the word "SaaSpocalypse" to describe a set of advances: vibe coding, legacy code refactoring, autonomous agents and Model Context Protocol servers. The argument is that software vendors will shift from graphical interfaces toward headless operations run by agents. SAP's Manoj Swaminathan, president and chief product officer, has said the company is supporting that direction by exposing its APIs.

Replacing major enterprise applications with homegrown code is a different matter. Most CIOs have sunk too much money and time into making those systems work, and they rarely give the business a competitive edge. Swapping them out is a hard sell.

COBOL and the mainframe are a different case. COBOL dates to 1959, and the mainframe is from the 1950s too. Both have been on the replacement list for years, and both still run core transactions for airlines, banks, insurers, public sector bodies, healthcare providers and retailers. If you've ever seen a green screen, there's a good chance a COBOL application on a mainframe is answering it.

Why this time may differ

This isn't the first modernization push. Low-code vendors have been chipping away at COBOL for years, and the skills gap has been a known problem since the Y2K scramble, when few people could still read the code well enough to extend it past 1999. What's changed is how agents approach the code itself.

The pattern described here works in stages. One agent reads the COBOL and turns it into a detailed specification of the application and its functions. A second agent rebuilds that application on a low-code, cloud-based development stack. A third runs regression testing against the new code. Each step is a bounded task, which is what makes the approach look practical rather than speculative.

That specification step is the part I'd watch most closely. Documentation for decades-old COBOL estates is often thin or wrong, and business rules tend to live in code nobody has touched in twenty years. If agents can pull those rules out reliably, the migration gets much easier to scope. If they can't, you're just moving the guesswork into a new platform.

What OutSystems is claiming

Woodson Martin, CEO of OutSystems, talked me through the company's work in this area last month. OutSystems is a low-code platform that focuses on complex industry use cases. Martin said the company is showing enterprises that requirements can be developed with an agent, the application built with agents, and agentic capabilities added to existing COBOL programs.

The company launched Legacy Modernization Services, powered by AWS Transform, in June. The service migrates COBOL applications in a multi-step process. Agents scan the code, build dependency graphs, extract business logic and keep a human in the loop at each stage. That last point matters. Nobody serious is proposing an unattended cutover of a core insurance system.

Martin's most concrete figures involve insurance, a heavy COBOL-on-mainframe sector. He said a manual migration of a COBOL application can take six months to a year. OutSystems is using agentic AI to bring that down to eight weeks, and Martin wants to reach seven days. Those are vendor figures, drawn from the company's own projects, and I haven't seen the test conditions or the size of the applications involved. Treat them as a target to verify, not a benchmark.

Asked directly whether this means the end of COBOL, Martin said yes. "What we are doing will lead to the death of COBOL and the death or relegation of the mainframe." He also framed the approach as the path many organizations will take to agentic AI. Expect vendors to make similar claims, and expect each one to come with its own caveats.

What to check before you believe the timelines

Here's the thing: a migration measured in weeks for one application is not the same as an estate-wide program. Mainframe environments carry batch schedules, integrations, data migration, audit obligations and cutover risk that a demo rarely shows. Regression testing is only as good as the test cases behind it, and those cases often come from the same undocumented logic the agents are trying to extract.

That said, the direction is clear enough to plan around. Systems of record aren't going away, but the value of SaaS and legacy applications is shifting toward systems of action, where software does work on the business's behalf. Over time, that could also change the economics vendors have relied on with their installed base. Organizations that free up budget by retiring expensive legacy run costs will have more room to fund AI work, and that's the real prize for CIOs.

My advice for infrastructure and application teams is simple. Inventory your COBOL and mainframe workloads now, including the ones nobody wants to talk about. Find out which applications have usable documentation and test coverage. Then run a pilot on a bounded, low-risk application and measure the result against your own baseline, not the vendor's. The next wave of mainframe decisions will be made on evidence from those pilots, and it's better to have your own numbers when that conversation starts.

Outsystems' Legacy Modernization Services is built to support COBOL migration, according to the company. Photo courtesy Outsystems.

Source: Myles Suer · www.informationweek.com

Neighbouring cells