For decades, the mainframe has been the boring, invisible machine that quietly runs the world's banking. IBM says these systems still handle more than 40 percent of online banking, 80 percent of in-person credit card swipes, and 95 percent of ATM withdrawals. Nobody sees them, but almost everyone touches one every week.
The problem is the code. Much of it is written in COBOL, a language from the 1960s that almost nobody under 40 learns anymore. The engineers who understand these systems are retiring, and companies have spent years afraid to touch code that, if it breaks, can take down payment systems for millions of people.
AI companies smelled an opportunity. In February, Anthropic said its coding assistant, Claude Code, could handle COBOL modernization for "systems of any size." Markets reacted instantly and badly for IBM: its stock had its worst one month decline in decades, as investors bet that AI would let banks finally ditch their mainframes and IBM's most reliable revenue with them.
That panic looks overdone in hindsight. Gartner, the research firm that advises most large companies on technology spending, now predicts that more than seven in ten mainframe exit projects started this year will fail, because leaders are overestimating what today's AI tools can actually do with old, undocumented business logic. As one Gartner analyst put it bluntly, these tools are useful, but they are still just another automation tool, not a replacement for the hard business decisions around cost, risk, and retraining staff.
The more interesting shift is not replacement, it is renovation. IBM, AWS, and Microsoft are all racing to sell tools that use AI to explain what old code actually does, translate it into modern languages, and speed up testing, while leaving the mainframe running underneath. Morgan Stanley built its own version of this in house and used it to modernize more than 17 million lines of old code, saving developers over 1 million hours of manual work. Toyota used a similar approach from AWS to convert 40 million lines of COBOL to Java.
There is a catch worth flagging for any company shopping for these tools: every vendor's modernization software is built to steer customers toward that vendor's own systems. Choose IBM's tool and you likely end up deeper inside IBM's stack. Choose AWS's tool and you likely end up on AWS's cloud. None of them are neutral.
The real lesson for any business running old, critical software, whether that is a mainframe or an ancient internal ordering system, is that AI has changed the economics of modernization without removing the risk. It can cut the grunt work from months to days. It cannot answer whether the project is worth doing, whether staff can be retrained fast enough, or whether the business case actually holds up once the invoice arrives. Those are still management questions, not technology ones, and companies that treat them as solved by AI are the ones Gartner expects to end up in the failure column this year.