A new study from Mercor pitted frontier AI models against 12 licensed CPAs, each averaging 5.5 years of experience, on real month-end close tasks. These are the detailed, repetitive accounting steps that finalize a business’s books every month. The results were not close: Claude Opus 5 scored 100% accuracy on all 20 attempts, completing each in under 10 minutes. By comparison, the human CPAs scored between 0% and 90% accuracy, taking anywhere from 30 to 180 minutes per task.
Before jumping to obvious conclusions, it is worth examining the fine print flagged by the researchers. The study focused on a specific slice of accounting work, emphasizing detail-oriented, well-defined, instruction-following tasks like file navigation and reconciliation. It did not test client communication, account for messy mid-close judgment calls, or capture the context an accountant builds over months on a specific account. The researchers noted that these findings do not make accountants replaceable, though they do signal a shift in the nature of the role.
For small business owners, whether using an in-house bookkeeper or an outsourced firm, the takeaway is clear: the routine, mechanical parts of financial work are rapidly becoming faster and cheaper. That is great news if you currently pay for hours spent on file-matching and data entry. It is also a reason to ask your finance team or accounting firm a key question: which parts of our books are candidates for AI assistance, and which parts still require a person who knows our business?
Businesses that figure out this balance early will redirect their accounting budget toward high-value judgment and advice. Those that wait will keep paying premium rates for work a model can finish in under 10 minutes.
At CMIT Solutions, we help businesses evaluate, integrate, and secure the latest AI capabilities to streamline workflows and boost productivity. Discover how we can help you navigate AI integration safely and strategically.
