In a Mercor study, AI outscored licensed CPAs on month-end close tasks
Mercor gave 12 junior accountants, all licensed CPAs, and Claude Opus 5 the same month-end close tasks: Claude alone scored perfectly, and accountants working with Claude were slower and slightly less accurate than Claude alone. The authors say the tasks tested what AI does best, not the whole job of an accountant.
What happened
On Oct 1, Mercor published Human Baselines for Benchmarks: AI Now Outperforms Junior Accountants, a research post by Aden Barton with a full paper behind it. Mercor sells expert human data and professional benchmarks to AI developers. It recruited 12 junior accountants, all US-based with an active US CPA license, averaging 5.4 years of accounting experience (the range was 3.8 to 8.0), six in public practice and six in industry. They worked through four month-end close scenarios adapted from Mercor's APEX-Accounting benchmark: dig through a company's working files, find the right numbers, do the math and deliver a table of results, with a three-hour limit per task, though most handed in before it. The same tasks went to Claude Opus 5 working alone and to accountants working with Claude. The tasks were built to stump models. Hard-to-spot requirements compound, so "a single overlooked number could produce a very low score", and the task authors expected about 30% from a junior accountant and 55% from a mid-level one.
The post's finding: "We find that on medium-length, well-defined accounting tasks, frontier AI models are now faster and more accurate than junior accountants, even the best one in our study." Claude alone scored 100% across its 20 attempts. The accountants without AI made 23 attempts, and Mercor writes: "Just eighteen months ago, the best AI models fell short of the average accountant's ~37% score." Working together did not help. Accountants with Claude were "slower and slightly less accurate than AI alone" and took about 15 times as long per task. In three of the four assisted sessions that missed a perfect score, Claude had the right figure at some point and it did not reach the final answer: twice Claude revised away from it and the accountant went along, and once the accountant overrode Claude's correct figure. Mercor puts the cost at $0.21 per rubric criterion met for Claude against $10.35 for the accountants, 49 times more, using the US median accountant wage rather than what participants were paid, and list token prices with no caching, which it calls an upper bound. The authors set the limits themselves: "These results do not mean accountants are replaceable, since our tasks ended up testing what AI is best at ... rather than the whole job of an accountant, which involves less structured tasks like client communication." Participants had "no coworkers to ask for help, no accumulated job context", and the authors write that they "considered not publishing" the results "for fear of misinterpretation." On the full 160-task APEX-Accounting benchmark, the best score on Mercor's leaderboard is 61.8%, and 58% of the tasks remain unsolved by any model.
Before
Eighteen months ago, by Mercor's account, the best AI models scored below the average accountant's roughly 37% on these same close tasks. Going through a company's working files, finding the right numbers, doing the math and handing over a table of results was work for a person, and the licensed CPAs in the study got up to three hours per task.
Now
On four well-defined close tasks, Claude Opus 5 working alone scored 100% in all 20 attempts, at a cost Mercor puts at $0.21 per rubric criterion met against $10.35 for an accountant at the US median wage. Working with an accountant did not raise the score: the pairs took about 15 times as long and were slightly less accurate. The rest of the job, from client communication to the context a person builds up at a company, was not tested.
Who it touches
- Small CPA and bookkeeping firms that bill for month-end close work: the study's tasks came from that work, and the human baseline was licensed CPAs averaging 5.4 years in accounting.
- Junior and staff accountants with three to eight years of experience, the range the study measured, whose well-defined close tasks are the part the model did best.
- Service-business owners who pay a bookkeeper or outside accountant every month: a fair question for the firm is which close steps it gives to AI and how a person checks the figures.
- Any office that has a person review AI output before it goes out: in this study, three of the four imperfect assisted sessions lost a correct figure between the first answer and the final one.
What it means for the business
The finding to carry into Redwater's builds is the part where people and AI worked together. In three of the four assisted sessions that missed a perfect score, the right figure appeared and was lost: twice the model revised away from it and the accountant went along, once the accountant overrode it. Redwater has built an AI bookkeeping step that sorts bank transactions, applies the ones the model is confident about and puts the rest in an Unsure list for a person to accept or change. This study is a reason to make that review screen show what a figure was, what it changed to and the document it came from, so a person checks a change against the source instead of trusting whichever answer came last.
For talks and content it is a sourced proof point that carries its own limits, and the honest version is the one to tell: the authors themselves say the results "do not mean accountants are replaceable." Bookkeepers and small CPA firms are both an owner type Redwater speaks to and a referral partner it looks for, and the question this study hands them is concrete: which steps of the close are well-defined enough to give a model, and who checks a changed figure against the file. Mercor's cost figures are per rubric criterion at a median wage, not a bookkeeping bill, so no pricing claim should be built on them.
