An AWS practitioner who has worked with over four hundred enterprise mainframe customers put it plainly in 2025:
"The first half is where mainframe projects live or die. However, coding assistants are genuinely good at only the second half."
The two halves: reverse engineering, understanding what existing systems actually do, and forward engineering, building the replacements.
The current wave of AI-assisted COBOL modernization tools is aimed at closing that gap. IBM's watsonx Code Assistant for Z uses AI to analyze and document existing COBOL applications before attempting translation. One customer, Egypt's National Organization for Social Insurance, reported a 79% reduction in the time developers needed to understand complex applications. AWS launched Transform for Mainframe in May 2025, calling it the first agentic AI service for modernizing mainframe workloads at scale, promising to cut timelines from years to months. The tooling is real. The money behind it is serious.
Those comprehension gains may be genuine. Knowing what code does, line by line, has always been easier than knowing everything it touches.
The same COBOL source code behaves differently depending on the compiler and runtime: how numbers get rounded, how data sits in memory, how programs talk to middleware. In financial systems, a rounding difference becomes a material error. And then there's everything that lives outside the code entirely. A single business function might stretch across COBOL, PL/SQL, shell scripts, and Python components, stitched together by job schedulers, flat files, and human procedures that aren't visible in any codebase. Test data is another gap AWS's own practitioners flag: teams get all the way through code conversion and stall because nobody planned for data capture against real production scenarios.
TSB Bank learned this in 2018. Three years of planning. Eighty-five specialized subcontractors. Board-level reviews. When they migrated from their legacy platform, a significant portion of their 5.2 million customers lost access to their accounts. The eventual fine was $62 million. The FCA investigation found that TSB had begun migration before finishing the work of defining what the system was supposed to do.
Legacy modernization projects fail at a widely cited rate of roughly 70%, a number that has held steady across multiple technology generations.
The AI vendors know this, at least in their documentation. AWS explicitly requires "human in the loop" validation throughout. IBM's tooling focuses heavily on the comprehension phase. But there's something worth sitting with. If the tools' primary contribution is making COBOL systems more understandable, more documented, more legible to the humans who maintain them, that could accelerate replacement. Or it could make the systems easier to keep.
One practitioner framed the concern sharply:
"If you use AI to simply rewrite a messy COBOL system into a messy Python system, you have not solved the problem. You have just made it easier for people to read the mess."
Previous modernization waves followed a recognizable sequence. New tools arrive that promise to finally make migration feasible. The tools encounter complexity that lives outside the code. The effort stalls or partially succeeds. The systems survive, sometimes better documented than before, sometimes with new integration layers wrapped around them. The dependency deepens.
AI coding agents are genuinely more capable than previous generations of tooling. They may break this cycle. But the pattern has survived capable tools before, and the place where it survives is in the gap between understanding a system well enough to translate its code and understanding it well enough to know what you'd lose.