On finally reading the mainframe

Define the terms first, because a budget is about to turn on them. Discovery is learning what a system does. Migration is moving the system somewhere else without changing what it does. Large language models have collapsed the cost of the first, and a research note published through the Open Mainframe Project this July draws the line cleanly: the models excel at documentation and code understanding, falter at conversion that preserves business semantics, and the useful question is where, specifically, AI creates value that survives contact with production.

The market is answering with its wreckage. Gartner's estimate, reported in late July, is that more than 70% of the mainframe migrations started this year will fail, mostly from overestimating what generative AI can do. The shape of the disappointment is familiar; the CASE tools of the 1990s also read code better than they rewrote it. What has changed is the asymmetry's scale. A model now summarizes forty years of COBOL in an afternoon, and the business rules it surfaces still have no test oracle except the production system that embodies them.

So the honest reading of this year is narrower and better than the roadmap slides: the estates became legible, which is worth a great deal, and they did not become portable, which was never the same claim. Comprehension is now cheap. Correctness is priced where it always was. Institutions that treat those as one line item are funding the difference with their integration budgets.