Gartner's April 2026 paper "Too Big to Fail" makes a bet you should take seriously: more than 70% of mainframe exit projects initiated this year will fail to deliver intended benefits — because leaders overestimated what generative AI code-transformation tools can do. By 2030, the firm says, 75% of "mainframe exit" vendors will pivot or shut down. The AI mainframe-exit market is a bubble. Here's how to keep your estate out of it.
The trap
GenAI is genuinely useful for discovering and describing technical debt in legacy code. Where it breaks is automated conversion: undocumented business logic, decades of patches, high coupling. Gartner is blunt — AI "does not account for the unique capabilities the mainframe offers," like throughput and transactional integrity. A translation that compiles is not a migration that survives year-end close or an audit.
The triage: three questions
Run every workload through these before approving an exit.
- Can you prove equivalence? No golden data, no test oracle, no audit trail to verify the migrated system matches the old one? Then you don't have a migration — you have a bet.
- How dense is the business logic? Boilerplate and I/O convert cheaply; thirty years of patched rules don't. Low coupling means AI-assisted is safe. High coupling means human-led re-engineering — or no exit.
- Does the workload actually need to leave? If mainframe throughput and integrity are the requirement, exiting is a cost-and-risk decision you can't win. Gartner calls a miscalculation here "a threat to business and operational continuity."
The call
Triage workloads, not portfolios. AI-assisted where you can verify, phased human-led re-engineering where you can't, and stay put where the platform still fits. Morgan Stanley modernized 17 million lines with its own DevGen.AI — but that's years of internal investment and verification, not a vendor demo. If the only thing that changed since last year is that a vendor showed you a COBOL converter, that isn't a strategy. It's the bubble.