A study by Mercor shows AI models now surpass licensed CPAs in speed and accuracy on structured bookkeeping tasks, though they still require human oversight to complete full accounting processes.
Key facts
- •The study compared AI performance against 12 licensed CPAs with an average of five and a half years of experience.
- •Eighteen months ago, AI models scored below the accountants' average of approximately 37 percent.
- •The APEX Accounting benchmark was developed by more than 40 professionals with an average of 11 years of experience.
- •Mercor anticipates that AI will drive significant productivity gains across the accounting industry.
- •AI models currently struggle to complete tasks that require professional context and interpersonal communication.
A new study by Mercor indicates that AI models have surpassed licensed CPAs in speed and accuracy regarding structured bookkeeping tasks. While AI models previously lagged behind human accountants, they now perform these specific tasks with high precision.
By the numbers
Performance Benchmarks
The study utilized tasks from the APEX Accounting Benchmark, which includes 160 tasks across 10 simulated companies. In current testing, Claude Opus 5.5 achieved the highest score with 61.8 percent of grading criteria met, followed by Fable 5.1 at 61.0 percent and GPT-6 Astra at 57.9 percent.
Limitations of AI in Accounting
Despite improvements, Mercor reports that no model fully solved nearly 60 percent of the tasks, indicating that AI cannot yet close books without human supervision. The study noted that while AI excels at following instructions and finding details, it lacks the ability to handle essential professional functions such as client communication, colleague collaboration, and the application of long-term contextual knowledge.
Advertisement
This article was independently rewritten by ManyPress editorial AI from reporting originally published by The Decoder.
