AI Helped Update Old Business Software — With Mixed Results

Imagine your bank or insurance company still runs its business on software written in the late 1990s. Updating that software is expensive, risky, and usually takes years. But what if an AI could do it automatically? A new study put that idea to the test — and the results are both promising and humbling.

Researchers asked Claude Code, an AI that writes computer code, to convert 12 features of a real corporate system from an old programming language called Visual Basic to a modern one called C#. The AI finished all 12 tasks. On average, the new features matched the old ones about 70% of the time. For simple features, the AI nailed it — 92% accuracy. But for highly complex features, accuracy dropped to 47%, meaning nearly half the work still needed human fixing.

The cost varied just as wildly. Simple features used about $1.66 in AI computing time; the hardest ones cost $10.28. That's incredibly cheap compared to paying human programmers, even if you double the numbers for oversight. But cheap doesn't mean perfect. The study shows AI is great at the boring, repetitive parts of updating old software — the parts that take up most of a human developer's time — but it struggles when the logic gets intricate and requires deep understanding of how a business works.

So what does this mean for you? Companies that modernize their aging systems with AI could save money and time — savings that might eventually mean better digital banking, faster tax filing, or fewer "system down for maintenance" messages. But it also means human programmers aren't out of a job yet. Their role is shifting from writing every line of code to supervising the AI and fixing the hard stuff. The lesson? AI is a brilliant assistant, but not yet a solo brain surgeon.