Agentic AI doesn't fix the mainframe problem, but it exposes hidden dependencies

IT modernization isn't a project, it's a chronic condition. A Kyndryl survey of 2,000 senior decision-makers shows only 10% of CTOs are using agentic AI tools for the effort, and despite decades of attempts — network services, microservices, cloud, virtualization — the finish line stays five to ten years out. Andy Thurai, former IBM chief strategist and now founder of The Field CTO, puts it simply: organizations are addicted to buying new software, but the old systems underneath never get retired.

The numbers behind the failure are unflattering. Only about half of organizations modernizing report improved IT operations, fewer than half see innovation gains. Eighteen percent cited limited or unclear value, and nearly half admitted they're behind schedule and over budget. Motivation has shifted accordingly: cost reduction and legacy escape are no longer top of the list; growing AI demand and sovereignty concerns are pushing leaders to demand the mix of old and new simply work better together.

That's where the agentic promise enters. Fewer than 10% of organizations are confident they know the interdependencies between their systems — a significant technical blind spot. Agents can map those hidden links, generate code, document processes that traditionally consume time and are error-prone. According to the report, early adopters of these tools show significantly smaller schedule delays than those that haven't.

Thurai isn't buying the hype. He calls AI just the latest "shiny distraction" from the grunt work of untangling technical debt. Four in ten organizations still run mainframes, and the old systems remain under layers of new software with no one decommissioning them. The bottom line: modernization economics don't hold water, and agents may help see the mess, but they don't solve the fundamental addiction to buying instead of cleaning.