Every vendor deck now promises the same thing: GenAI will read your COBOL, rewrite it, and march you off the mainframe. Boards love it because it turns forty years of technical debt into a tooling purchase. Gartner just put a number on how that story ends: more than 70% of mainframe exit projects initiated in 2026 will fail to deliver their intended benefits.
The misconception
The myth is that AI is an exit ramp. It isn't. GenAI is genuinely good at understanding legacy systems — JPMorgan reports a 75–85% improvement in the speed of comprehending legacy code — but understanding is not migration. Gartner VP analyst Alessandro Galimberti calls it "a widening gap between the marketing promise of GenAI and its real-world ability to transform and migrate complex legacy code."
The code was never the hard part. The hard part is the data semantics, the undocumented dependencies, and the failure modes nobody has read in twenty years. Combine that with the too-big-to-fail nature of core systems and the retiring COBOL workforce, and you get what Galimberti calls "a perfect storm of risk."
The correction
Gartner's platform-smart math is blunt. For estates above 25,000 MIPS, a total exit is "nearly impossible" and rarely justifiable — the strategy is optimization in place, including running AI on the mainframe. For mid-size estates, only 15% should pursue rearchitecture; the rest should modernize tactically or adopt large-environment playbooks. And by 2030, Gartner expects 75% of mainframe-exit vendors to have pivoted or shut down as demand for one-size-fits-all migration collapses.
Treat GenAI as a modernization assistant — documentation, test generation, code comprehension — not a migration engine. Budget and govern it accordingly, and you won't join the 70%.