AI beats licensed accountants on speed and accuracy, but still can't close the books without supervision
decoder ·
A recent study by Mercor reveals that AI models now outperform licensed accountants in speed and accuracy when handling structured bookkeeping tasks. The research compared twelve CPAs with an average of five and a half years of experience against leading AI systems using the APEX Accounting Benchmark. While previous models scored poorly eighteen months ago, current top performers like Claude Opus 5.5 achieve high accuracy rates on these specific tests. Despite these impressive metrics, no AI model has yet fully solved nearly 60 percent of the benchmark's complex tasks. Consequently, human supervision remains essential for closing books, as AI cannot yet operate entirely autonomously in this domain. The study acknowledges that it focuses on detail-oriented instructions where AI excels, omitting critical interpersonal aspects of accounting such as client communication. Therefore, while full replacement is not feasible, the industry is expected to see significant productivity gains through AI assistance.
decoder