AI Can Now Read Your Old Software and Explain What It Does

A new research paper describes a team of seven AI 'agents' — software that can look things up, run tools, and take actions without a person clicking for it — that does something companies pay consultants fortunes to do: figure out what old software actually does. Most big companies run programs built decades ago, and the rules inside them (interest rates, refund policies, approval chains) were never written down anywhere except in the code itself. The AI reads that code, plus configuration files and tests, and writes a plain-English document listing every business rule it finds.

Why should you care? Because this kind of detective work is brutally expensive and slow. The paper says a comparable past migration at the same company took more than two years of human effort. With the AI doing it, each service took under nine minutes. The authors calculate a cost reduction above 98%. For any industry running on aging systems — banks, insurers, airlines, government, hospitals — that is the difference between a project nobody can afford and one a mid-size team can finish in a quarter.

There is a real catch. The results come from a single company's deployment, so we don't yet know how well it travels to messier code. It also needs decent code, tests, and documentation to work from. Rules that exist only in a retiring employee's head still can't be recovered by any machine, and the process keeps human reviewers at checkpoints — because a wrong rule written into the new system could quietly break billing or compliance for years.

What to watch next: whether other companies reproduce these numbers. If they do, expect this approach to spread fast, and expect IT contractors who bill by the hour for 'legacy modernization' to feel it. If you work somewhere with old software, the useful question to ask your tech team is simple: how much are we spending to understand systems we already own?