Accounting: AI Advances but Remains Under Supervision

AI models outperform accountants on structured tasks, according to Mercor, with the best achieving between 57.9% and 61.8% of the criteria. However, nearly 60% of tasks remain without a complete solution, and book closing still requires supervision. Some key dimensions of the profession have been overlooked.

No Autonomous Closing and a Limited Testing Scope

Mercor indicates that AI models are not capable of closing books without supervision. No model has fully solved nearly 60% of the benchmark tasks. The tested tasks correspond to what AI handles best, such as accurately following instructions and searching for details. Essential aspects of the profession, such as communication with clients, collaboration with colleagues, or utilizing context accumulated over the years, were not evaluated. According to Mercor, these omissions explain why accountants cannot be replaced, even though significant productivity gains are expected in the sector.

Current Scores and Leap in Eighteen Months

Claude Opus 5.5 achieves the highest score with 61.8% of the rating criteria met, followed by Fable 5.1 at 61.0% and GPT-6 Astra at 57.9%. Eighteen months ago, the top-performing AI models had results below the average for accountants, which was set at around 37%. Now, these same models are able to handle simplified tasks almost perfectly.

What APEX Measures and How the Comparison Was Made

The APEX benchmark consists of 160 exercises spread across 10 fictitious companies, designed by over 40 specialists with an average of 11 years of experience. As part of the study, twelve certified public accountants, averaging five and a half years of experience, performed simplified versions of these tasks. For these standardized exercises, Mercor indicates that AI systems outperform professionals in speed, accuracy, and cost. The accuracy of the AIs was compared to that of accountants operating without assistance, with detailed results according to the launch date of the models.