Unizel builds simulated finance back offices to train and test AI agents on real accounting work.

The agent takes over a company whose last accountant left errors in four months of closed books: a returned check never recorded, a bill paid twice, a bill posted to the wrong account. It has to find each one, fix or escalate it as the company's policy requires, then close the month and report what it changed.

We ran Claude Opus 5.5 eight times on each of six companies. A run passes only if every check passes, to the cent. Every failure is a genuine model error.