AI can now rewrite COBOL — and the market noticed

Two hundred twenty billion lines. That is how much COBOL code, by various estimates, still processes transactions in banks, insurance companies, government agencies, and payment systems around the world. A language created in 1959—a decade before the moon landing—remains the invisible foundation of global financial infrastructure. And this foundation is cracking: the number of specialists capable of maintaining it dwindles each year, while the cost of errors in its replacement is measured in billions. Now artificial intelligence offers the first realistic path out of this deadlock—and financial markets have already begun repricing their stakes.

The COBOL problem is not simply a technical curiosity for vintage computer enthusiasts. It is a systemic risk that people prefer not to discuss openly. According to Reuters, roughly 95 percent of ATM transactions in the US and around 80 percent of in-person financial operations pass through COBOL systems in one way or another. The average age of programmers proficient in this language has long surpassed sixty. Every year the industry loses bearers of critically important knowledge—people who understand not merely the syntax, but the business logic accumulated over decades in millions of lines of code written without modern documentation, without version control systems, often without any comments whatsoever.

Modernization attempts have been undertaken repeatedly. Major banks spent hundreds of millions of dollars on legacy system rewrite projects—and failed time and again. The famous collapse of Commonwealth Bank of Australia's initiative, which spent over a billion dollars replacing its core system, has become a textbook example. The problem is not in writing new code, but in precisely reproducing the behavior of the old code—with all its non-obvious quirks, workarounds, and decades of accumulated business rules that no one remembers.

This is precisely where large language models enter the picture. Modern LLMs have demonstrated unexpectedly strong capabilities for analyzing and translating legacy code. They can read COBOL programs, recover the business logic embedded within them, generate equivalent code in Java or Python, and—critically important—explain what each fragment does. In essence, AI acts as a translator between eras, compensating for the loss of institutional memory. Several companies have already brought specialized tools to market: IBM has integrated AI migration capabilities into its watsonx platform, and startups like Phase Change and Modern Systems are attracting significant investments precisely on the promise of automated COBOL modernization.

The market reacted with characteristic directness. Stock prices of a number of consulting companies whose business was built on manual legacy system modernization—multi-year projects with massive teams and predictably ballooning budgets—came under pressure. Investors are rightfully asking themselves: why pay thousands of consultants for a five-year project if an AI tool can accomplish much of the work in months? This does not mean that human expertise becomes unnecessary—validation, testing, and integration still require deep understanding. But the ratio of manual to automated labor in such projects is changing radically.

However, it is important not to succumb to euphoria. COBOL modernization with AI is not a magic button. Language models can hallucinate, generating code that looks correct but behaves differently in edge cases. In financial systems, where an error of one cent across billions of transactions can lead to catastrophic consequences, every line of translated code must pass through multi-level verification. Regulators—from the Federal Reserve to the European Central Bank—have not yet established clear standards for AI-assisted migration of critical infrastructure. This creates legal uncertainty that could slow implementation even with technical readiness in place.

Nevertheless, the direction of movement is clear. The COBOL problem is a problem worth trillions of dollars, and for the first time in a decade, a tool has emerged capable of making its solution economically and technically feasible. For the financial industry, this means the beginning of the largest technological migration in its history. For AI companies—the opening of a vast and exceptionally solvent market. And for all of us—hope that the system through which our salaries, payments, and savings pass will finally stop depending on code written before the internet was invented, and on a handful of specialists who remember how it works.