Key takeaways
- AI makes the close more accurate on high-volume, rule-shaped work like matching and coding, and a 2026 Journal of Accounting Research study found AI assistance “improves classification accuracy on average.”
- AI isn't accurate enough to run the close alone: in the APEX-Accounting test of month-end close tasks (July 2026), the top model met 56.4% of an expert's checklist on average, and the most consistent model solved just 2.6% of tasks correctly in all eight runs.
- You can prove accuracy in your own close by planting known errors in a month you've already closed and tracking how many the AI catches, how many of its flags are real, and how many adjustments land after sign-off.
AI makes the month-end close more accurate when it prepares high-volume, rule-shaped work like matching and coding, and a person reviews it, but in the APEX-Accounting test of close tasks, no model got even 3% of tasks right in all eight runs. Using AI this way gives you fewer manual errors, earlier warnings, and a named reviewer behind every posting. This guide covers the evidence, the close tasks AI handles well, the errors it makes, and how to review and measure its work.
We make AI agents for finance teams, including a Close Agent, so we've kept this guide to public evidence. Every source was checked on its original page on October 13, 2026.
What does the evidence say about AI and month-end close accuracy?
The evidence says AI improves accuracy on average for structured tasks, but it still makes errors that need a reviewer. Knowing what each study measured helps you separate tested results from vendor claims.
- AccuracyWhether the numbers in the close are right: correct account, period, amount, and support
- ConsistencyWhether the AI gets the same task right every time it runs
- RecallThe share of real errors the AI catches
| Source | What it tested | What it found |
|---|---|---|
| Choi and Xie, Journal of Accounting Research, 2026 | Data from an AI accounting platform serving 79 small and mid-sized firms, plus a survey of 277 accountants | AI “improves classification accuracy on average,” but “reliance on non-consensus AI recommendations can increase the risk of error” |
| APEX-Accounting, Mercor with Ramp, July 2026 | 160 month-end close and bookkeeping tasks across 10 simulated companies, nine AI models | Best model met 56.4% of the expert checklist on average; the most consistent model solved 2.6% of tasks correctly in all eight runs |
| FinBalance, academic paper, June 2026 | Six AI models turning bundles of source documents into journal entries and a balance sheet | The best model built an exactly correct final balance sheet in at most 46% of cases |
| Gartner survey of 497 controllership staff, published February 2024 | How often accountants make errors | 18% of accountants make financial errors at least daily, and 59% make several a month |
The Choi and Xie study is peer-reviewed and draws on real accounting records, not only a lab test. It also found that accountants “selectively intervene when AI confidence scores are low.” That's the pattern that makes AI accurate in practice: the AI does the volume, and people step in where it's unsure.
The APEX-Accounting results show why that review step matters. Its 56.4% is the share of an expert's checklist the best model met on average, not the share of tasks it finished correctly. The paper also found that “58% of tasks were never fully solved by any model on any run.” It added that among the top three models, “roughly seven in ten failures came from flawed reasoning, not an inability to find the right information.” APEX-Accounting was built by Mercor and Ramp, which both sell AI products, and it doesn't test consolidation, multi-entity, or multi-currency work.
Will AI ever be 100% accurate? No test so far comes close, and people aren't either: the Gartner figures show errors are already part of today's close. The useful question is whether AI plus review produces fewer errors than your current process, which you can measure.
Is AI replacing accountants? None of these results support a close that runs without them. For how AI agents split work with your team, see best AI agents for finance teams.
What parts of the close can AI automate accurately?
AI is accurate on tasks with clear inputs, a right answer you can check, and high volume. It's weakest where the answer depends on judgment. Sorting your close this way shows you where to start.
| Close task | How well AI fits today | What a person still checks |
|---|---|---|
| Bank and card matching | Strong: exact amounts, dates, and references | Partial payments, transfers, and anything left unmatched |
| Invoice checks against PO and receipt | Strong: quantity, price, and duplicate checks are rule-based | Price and quantity variances above tolerance |
| Transaction coding | Strong on average, per the Choi and Xie study | Unusual items and low-confidence suggestions |
| Recurring schedules (prepaids, depreciation) | Strong on the math | Contract terms, start and end dates, and setup |
| Variance explanations | Good first drafts | That every explanation traces to actual transactions |
| Accrual estimates | Weaker: estimates need judgment and current facts | The basis for each estimate and its amount |
| Revenue recognition, impairment, reserves | Not reliable enough to prepare alone | The judgment itself, with AI gathering support |
| Consolidation, intercompany, multi-currency | Reliable automation isn't established by these results | All of it, until you've tested it on your own data |
When should AI be avoided? Avoid it for any judgment you can't check against a source document, such as a reserve estimate or a revenue recognition call, unless a person makes the final decision. For the full task list, see the month-end close checklist.
What errors does AI make in the close?
AI makes familiar errors, such as a wrong account, period, or amount, and it can repeat one error across hundreds of lines. Knowing the error types tells your reviewers exactly what to look for.
| Error type | What it looks like | The check that catches it |
|---|---|---|
| Wrong account | Software subscription coded to office supplies | Review entries by account against the chart of accounts and prior months |
| Wrong period | June invoice accrued in July | Compare service dates on the invoice to the posting period |
| Wrong amount | Digits transposed or a partial payment treated as full | Tie each amount to the source document |
| False match | Payment matched to the wrong invoice with the same amount | Check references and counterparties, not just amounts |
| Plausible but wrong explanation | A variance explained as “seasonal” when it's a miscoding | Trace each explanation to the underlying transactions |
| Repeated error | One wrong mapping rule applied to every similar transaction | Test a sample across vendors and months, not just one |
| Missing data | Transactions dropped by a filter or a missing statement | Agree counts and totals to the source system |
Example
An AI prepares a June accrual for a $4,800 utilities invoice dated June 28. It makes three errors: it codes the entry to rent instead of utilities, dates it July 1, and records $8,400 by transposing digits. A reviewer who checks account, period, and amount against the invoice catches all three. Left unreviewed, June expense is understated by $4,800 and July is overstated by $8,400, a $13,200 swing across two months.
| Field | As the AI prepared it | As the invoice supports it |
|---|---|---|
| Account | Rent | Utilities |
| Date | July 1 | June 30 |
| Amount | $8,400.00 | $4,800.00 |
How can finance teams keep control over AI-generated entries?
Keep control by treating AI as a preparer that never approves its own work and never posts without a named person's sign-off. This keeps your existing controls intact while the AI does the preparation.
In the PCAOB's July 2024 staff spotlight on generative AI, audit firms told PCAOB staff that “human involvement remains essential for auditors and preparers and is needed to review the output from GenAI.” Some warned that GenAI “may sometimes generate new content that is false or misleading.” And an engagement team member who uses a GenAI tool, they said, “is still responsible for the results and documentation of the work.”
In practice, that means:
- Approval before postingNothing reaches the ERP until a named approver approves the exact lines.
- Separate preparer and reviewerThe AI prepares, and a different person reviews. The AI can't approve.
- Evidence on every entryEach entry links to its source documents and the reconciliation it came from.
- A record of every stepEvery query and calculation the AI ran is logged, so you and your auditors can retrace it.
- Stop when data is missingThe AI asks for a missing statement instead of estimating a balance.
- Tiered reviewLarger or unusual entries get a second reviewer, as they would today.
Reclass INV-88213 to 2000 Accounts payable
- Dr 2100 · Accrued expenses7,200.00
- Cr 2000 · Accounts payable7,200.00
Can AI help with compliance during the close? Yes, when it checks every line instead of a sample and flags exceptions with their source. It doesn't replace your control owners. For what to review on each entry, see month-end journal entries.
How do you check AI-prepared work without redoing it?
Check the edges and a sample, not every line: tie the totals, review every exception, and trace a spread of matched items back to source. This gives you confidence in the whole reconciliation for a fraction of the effort.
- Tie the totalsAgree the ending balances to the bank statement and the general ledger.
- Review every exceptionRead each item the AI flagged or couldn't match.
- Trace a samplePick matched items across different amounts, dates, and counterparties, and trace each to both sides.
- Check the explanationsConfirm each reconciling item names its amount, date, and source.
- Look for repeatsIf one item is wrong, check whether the same rule touched other items.
Example
Your operating account has 1,200 June transactions. The AI matches 1,180 and flags 20 exceptions. You tie the two ending balances, review all 20 exceptions, and trace 25 matched items chosen across amounts and dates. That's 47 items reviewed instead of 1,200. If any traced item is wrong, you widen the sample before you sign off.
| What you review | Items |
|---|---|
| Ending balances tied | 2 |
| Exceptions reviewed | 20 |
| Matched items traced | 25 |
| Reviewed, of 1,200 transactions | 47 |
- 1010 · Operating accountBank statement2,184,920.000.00Tied
- 1020 · Payroll accountBank statement312,480.000.00Tied
- 1200 · Accounts receivableCustomer subledger1,642,300.000.00Tied
- 2000 · Accounts payableVendor subledger904,160.00−7,200.001 exception
Reconciling item · 1 of 1
INV-88213 · Brightwater Freight
7,200.00In the vendor subledger, but posted to 2100 Accrued expenses in the ledger. Reclass to 2000 proposed.
What metrics show AI is working?
The metrics that show AI is working measure errors caught and errors that get through, not hours saved. Tracking them before and after gives you evidence instead of a vendor's percentage.
| Metric | How to calculate it | What it tells you |
|---|---|---|
| Recall on planted errors | Planted errors caught ÷ planted errors | Whether the AI finds the errors you know about |
| Precision of flags | Genuine issues ÷ items flagged | Whether its flags are worth your reviewers' time |
| Reviewer change rate | AI entries changed or rejected ÷ AI entries prepared | How often the AI's work needs fixing |
| Post-close adjustments | Entries booked for the period after sign-off | Whether errors are getting past the close |
| Unsupported reconciling items | Items without a source document ÷ all reconciling items | Whether the work is audit-ready |
| Escaped errors | Errors found later by audit or next month's close | Whether review is catching what matters |
Example
You plant 12 known errors in a month you've already closed. The AI flags 30 items, including 10 of your 12 errors and 2 real errors your team missed at the time. Recall is 10 ÷ 12, or about 83%. Precision is 12 genuine issues ÷ 30 flags, or 40%. That tells you to check how it missed the other 2, and whether 18 false flags a month is a review load your team can carry.
Review hours per account still matter, but they measure speed, not accuracy. For speed benchmarks, see how long should month-end close take.
Is it safe to use AI for accounting?
It's safe to use AI for accounting when your data stays in an environment you approve, the model doesn't train on it, and nothing posts without a person's approval. Checking these points first protects your data and your books.
Ask any vendor where your data runs and whose AI model and keys it uses. Then ask whether the model can see your credentials and what it can write to your ERP. Pasting ledger exports into a general chatbot can fail several of these checks. Can you use ChatGPT for accounting? It's useful for analysis and drafts. A standalone chatbot does not provide an approved accounting posting workflow by default. For how general assistants compare with finance agents, see best AI agents for finance teams.
Where should you start with AI in the close?
Start with one high-volume, checkable task, run it in parallel with your current process, and measure it before you rely on it. This proves accuracy on your own data before it touches your books.
- Pick the taskBank reconciliations or invoice checks are good first choices because the right answer is easy to verify.
- Set a baselineCount post-close adjustments and reviewer corrections over your last three closes.
- Test on a closed monthPlant known errors and measure recall and precision.
- Run in parallelConnect the AI to your ERP and bank feeds in read-only mode, and let it prepare two or three closes while your team works as usual.
- Turn on posting with approvalOnce the numbers hold up, let it post approved entries, one task type at a time.
Expect results to build over time. Gartner's November 2025 survey of 183 CFOs and senior finance leaders found 59% use AI in their finance function. Among those using it, 34% apply it to error and anomaly detection. Even so, 91% of respondents reported low or moderate impact initially.
How we build for accuracy
Our Close Agent is built for this kind of review. Every difference it finds shows its amount, date, source, and reference, and a named approver approves the exact lines before anything posts. When data is missing, it stops and asks instead of guessing. If you want to see AI prepare your close for review on the ERP you already run, explore the Close Agent.
Share this article