AI Beat CPAs on Month-End Close Tasks, but Dropped Figures Anyway
A new study found an AI model met every criterion on four month-end tasks, while CPAs without it averaged about 37%. Twice it dropped a correct figure.
A study circulating among accountants in early October 2026 puts a sharp point on the AI question. The model in question, Claude, met every scoring criterion across four month-end close tasks. Twelve CPAs working without AI on the same tasks averaged about 37%. The detail that matters most is quieter: twice, the model dropped a figure that was right, and the CPA reviewing the work went along with it.
The results that got attention
On the headline numbers, the AI won comfortably. Claude met every criterion on the four tasks. The twelve CPAs who worked without AI averaged about 37% against the same scoring. That gap will fuel arguments on both sides of the automation debate.
Treat the numbers as directional. The public summary does not say how the criteria were weighted or how hard the tasks were. What it does establish is simpler: on this test, the machine produced more complete work than trained humans working alone.
The failure nobody caught
The score is not the part that should worry you. Twice, Claude dropped a correct figure, and the CPA reviewing the output did not catch the omission. That is the classic shape of automation risk. An omission produces no error message. Nothing turns red. A report can be internally consistent and still be missing a loan, an invoice, or an entire account.
Humans check arithmetic far better than they check completeness, especially when a draft looks polished. A wrong number invites scrutiny. A missing one just looks tidy.
So can AI run your month-end close?
Not by itself. The study covered four specific month-end tasks, not a full close. A real close involves source documents, judgment calls, and follow-up questions nobody scripted. And the model failed in the most dangerous way possible: silently.
The more useful takeaway concerns you, not the software. If a CPA missed a dropped figure, a tired owner reading a clean-looking report at 9 p.m. is less likely to catch one. Whatever prepares the numbers, your review has to assume something was omitted.
Five checks before you sign off
The commentary around the study pointed QuickBooks and Xero users toward review checks; ours below are written for QuickBooks Online. They take minutes, and they target the exact failure the study exposed: a correct figure that quietly vanished.
- Agree every balance sheet account to a statement. Cash, loan, and credit card balances should match the month’s statements. A dropped figure typically surfaces as a reconciliation difference.
- Clear Undeposited Funds before you close. Customer payments sit in that account until they are matched to a bank deposit. A leftover balance usually means a payment was never banked, or was posted to the wrong place.
- Compare this month’s trial balance to last month’s. Look for accounts that should carry a balance and show zero. A dropped line leaves a hole that this comparison catches quickly.
- Review anything in uncategorized income or uncategorized expense. Items parked without a proper account hide both errors and omissions.
- Read the audit log. QuickBooks Online records added, changed, and deleted transactions with dates and users. Look for deletions near the close date and ask about them.
The practical next step
Run the five checks on your next close before you sign off, no matter how the numbers were prepared. Note what each one catches. A single cycle will show you where an hour of review protects you most, and whether any AI assistance you use has earned a place in your workflow.