AI models outperform licensed accountants in speed and accuracy on structured bookkeeping tasks, according to a recent study conducted by Mercor. In the evaluation, twelve licensed CPAs with an average of five and a half years of experience completed simplified tasks from the APEX Accounting Benchmark. While top models previously scored well below the human average of roughly 37 percent just eighteen months ago, current models now complete these identical tasks nearly flawlessly.
The full APEX Accounting benchmark evaluates systems across 160 tasks spanning ten simulated companies. Claude Opus 5.5 currently leads the benchmark by meeting 61.8 percent of grading criteria, followed closely by Fable 5.1 at 61.0 percent and GPT-6 Astra at 57.9 percent. Despite these high scores, Mercor notes that no tested model fully solved nearly 60 percent of the tasks autonomously, indicating that artificial intelligence still cannot close the books without human supervision.
Mercor emphasizes that the study deliberately tested areas where AI excels, such as retrieving precise details and following explicit instructions. The evaluation excluded essential components of professional accounting like client communication, colleague collaboration, and years of contextual institutional knowledge. Consequently, experts explain that accountants remain irreplaceable for comprehensive roles, even as the industry anticipates major productivity gains.
This research matters because it highlights the shifting capabilities of artificial intelligence in professional services, demonstrating that while models can automate specific structured tasks, human oversight remains vital for complex workflows.