New AI Test Shows Which Coders Fix Business Software Best

Businesses run on enterprise resource planning (ERP) systems — software that handles accounting, inventory, and orders. Microsoft's Business Central is one of those systems, and it uses a special language called AL. Now, researchers have built a test called BC-Bench to see how well AI coding agents (AI that can write or fix code on its own) handle real AL tasks. They pulled 101 actual tasks from Microsoft's own production code, including bug fixes and writing tests.

The results are a reality check. When these AI agents were asked to fix bugs, the biggest differences were between the AI models themselves — not the tools that run them. More importantly, an AI that scored well on general coding tests didn't necessarily do well on AL. That means companies can't just assume the latest AI will handle their specific business software. The AI needs testing in the real environment it will be used in.

Why does this matter? If AI can fix business software reliably, it could save many hours of expensive developer time. But if it can't, businesses might end up with broken customizations or hidden errors. The researchers also included visual tasks, since developers often work with screenshots and user interfaces, making the test more realistic.

This is one of the first benchmarks focused on AI for a specific business coding language. It's a reminder that general AI praise doesn't always translate to the messy, specialized systems that keep companies running. The test could soon become a standard way for businesses to check whether AI coding tools are truly ready for their software.