AI accounting tools outperform human CPAs on routine tasks but require oversight

Mercor's study shows AI handles simplified bookkeeping faster and more accurately than licensed accountants, yet still needs human supervision for full financial closure.

Mercor's study found AI models like Claude Opus 5.5 (61.8%), Fable 5.1 (61.0%), and GPT-6 Astra (57.9%) consistently outperformed 12 CPAs with five-and-a-half years of experience on simplified accounting tasks from the APEX Benchmark. These models solved structured bookkeeping work almost flawlessly—significantly exceeding CPAs' previous average of 37 percent. Crucially, no AI achieved 60 percent on all tasks, and Mercor emphasizes the study tests only AI's strength in detail-hunting and rule-following, excluding client communication, coordination, and long-term contextual understanding. This reshapes accounting workloads without full automation: human oversight remains essential for complex financial closure, though productivity gains are expected across the industry. The study's simplified tasks and exclusion of real-world practices mean this doesn't reflect full accounting workloads. For abundance, it reduces the time accountants spend on routine tasks—freeing them to address more urgent financial needs while keeping core oversight human-led. The source explicitly states this doesn't account for real-world accounting practices like client interaction or contextual knowledge accumulation over time.

Source: The Decoder