AI Can Click Through Your Office Software — But It Keeps Typing Wrong Numbers

Businesses run on big behind-the-scenes systems that handle invoices, orders, payroll, and inventory. These screens are dense, the steps are linked together, and one wrong click can quietly change a company's official records. Until now, most AI testing focused on simple everyday desktop tasks, not this kind of high-stakes bookkeeping work.

A team of researchers created ERPBench, a test that runs AI agents inside a live, working version of this kind of software. The AI sees only screenshots — no special access — and does the job by clicking and typing, just like a person. Then the researchers check the database itself, not the screen, to see whether the right number actually landed in the right place.

What they found is uncomfortable. Six different AI systems, both paid and free, were tested. Strong performance on ordinary computer tasks did not carry over. The scariest detail: some agents saved a record in up to 85% of attempts, but wrote the correct figure in as few as 3%. To a human watching, the work looked finished. The screen confirmed success. The company's records were wrong.

The team also built a safety version of the tool that pauses and asks a human to approve each action before it's saved — a realistic way companies could use this today. If you work in finance, HR, or operations, this is the near-term reality: AI can absolutely help you, but someone still has to check its homework. Don't let software sign off on numbers it can't verify.