Mercor study: AI models now outpace licensed CPAs on structured accounting work, but fail to close full books
A new Mercor study finds AI models beat licensed accountants on speed and accuracy for structured tasks, yet none can complete the full APEX Benchmark without human oversight.
AI systems have overtaken licensed certified public accountants on structured accounting work, according to a study from recruiting and evaluation firm Mercor. The models now finish those tasks faster and more accurately than human CPAs, a reversal from roughly 18 months ago when they trailed well behind. The finding lands as finance teams weigh how much of the close process can be automated.
The gap matters because accounting is one of the few white-collar fields where accuracy is legally binding and errors carry audit consequences. Structured work — matching ledgers, reconciling entries, classifying transactions against fixed rules — is exactly the kind of bounded task that large language models handle well. Mercor's results suggest the performance curve on those jobs has bent sharply in less than two years.
The harder test came from the APEX Benchmark, which bundles multi-step accounting problems rather than isolated entries. No model fully completed every task on it. That shortfall is the study's central caveat: models can win individual comparisons against CPAs while still failing to carry a set of books from start to finish without a person checking the work.
The distinction between task-level and process-level performance is where most enterprise AI claims break down. A model that reconciles a subledger correctly does not automatically know when a transaction looks wrong for reasons outside its training data, or when a judgment call requires an accountant's professional skepticism. Mercor's numbers put a hard boundary around what current systems can be trusted to do unsupervised.
What changes for the field is less about headcount than about workflow design. Firms that treat AI as a reviewer of first-pass structured work may see real time savings, while those expecting an autonomous close will hit the APEX wall. The open question is whether the next generation of models closes the gap on multi-step reasoning, or whether full autonomy stays out of reach until accounting-specific tooling and controls catch up.