The most useful sentence about AI bookkeeping this year was in a footnote. Under its announcement of an agent that drafts the month-end close, Intuit wrote: "It may not always be accurate, and every AI-drafted entry requires your review and approval before it posts."
That is the company selling the product. On the same page, Intuit reports its chief technology officer saying that responsibility for the correctness of financial outcomes "still rests with the accounting firm." I agree with both statements. The harder question is what a good review of AI-prepared books looks like, and the evidence on where agents go wrong is now good enough to plan around.
Standard setters say to treat AI output like the work of a junior colleague. That is the right level of skepticism, but the errors are different. A junior who cannot make a reconciliation work usually leaves the difference sitting there for you to find. An agent may make the numbers tie anyway.
A word on where I sit: our firm sells human-reviewed bookkeeping, so we have an interest in this question. I have kept to published evidence and linked all of it at the end, so you can judge it for yourself.
What can AI agents do in QuickBooks and Xero today?
More than a year ago, the tools moved from suggesting categories to doing the work. Intuit launched its QuickBooks Accounting Agent in July 2025 to categorize transactions and help with reconciliations. At its Connect conference in October 2026 it added accountant-side agents, including a Books Close tool in beta with anomaly detection and an Agentic Books Close pilot. Intuit describes the division of labor plainly: "AI handles the mechanical prep: Posting pending transactions, reconciling accounts, flagging anomalies, and recommending fixes."
Xero's agent, JAX, runs Auto Bank Reconciliation, which Xero says automatically reconciles more than 60% of the transactions it sees, with every result left ready for review. Ramp, Digits, Basis and others sell agents that categorize, match and prepare entries, mostly with the firm reviewing the output.
Adoption is broad, and trust is not. In Intuit's 2026 survey of 725 US accounting professionals, 88% had used AI for at least one client service in the past year. Only 6% wanted autonomous AI execution. The largest group after "AI as a support tool" was the 34% who want AI to draft and a human to review and sign off. Karbon's 2026 survey found that 46% of accounting professionals were unsure whether to trust AI.
Most of the profession wants the machine to draft and a person to decide. The rest of this piece is about what that person should check.
How accurate is AI bookkeeping?
Start with the vendors' own numbers, and read the footnotes. Xero says Auto Bank Reconciliation achieves about 96% accuracy on the transactions it reconciles automatically. The footnote defines that as the percentage of auto-reconciled lines "not corrected by users," on a rolling three-month average from April to July 2026. Ramp's 98% figure for its accounting agent is measured the same way: transactions synced "without any subsequent changes."
Both are honest about the method, and both measure the same thing: how often a user changed the result. That is not the same as how often the result was right. An error nobody noticed counts as accurate.
Independent tests are harder on the technology:
- Penrose AccountingBench (July 2025). Models were given a year of real data from a SaaS business and asked to close the books month by month. In the first month their balances were typically within 1% of a CPA's. After several months, errors had compounded and balances diverged from the correct ledger by more than 15%, about half a million dollars. Subscription revenue "was consistently overstated by 5-30%, even for the best model," and the models drifted back to cash-basis treatment.
- APEX-Accounting (Mercor with Ramp, July 2026). On 160 expert-written month-end tasks, the best of nine frontier models met 56.4% of the grading criteria, and 58% of the tasks were not fully solved by any model in any run. The authors' conclusion is worth quoting in full: "An agent that is usually correct but not reliably so cannot yet close the books unsupervised."
The APEX authors also describe how the failures happen, and it should change how you review. The model "retrieves the correct documents, often computes the correct intermediate values, then substitutes the wrong basis, source, or authorization logic downstream." The inputs look right. The reasoning that follows is where it goes wrong.
Benchmarks are snapshots. Models improve quickly, and a 2025 result says little about the tool your client bought last month. But the pattern across these studies is consistent: agents do well on routine, repetitive work and less well as judgment and history accumulate, which is exactly the shape of a month-end close.
Why does the reviewer become part of the risk?
The uncomfortable research is about us, not the software.
Jung Ho Choi of Stanford and Chloe Xie of MIT studied accountants using an AI accounting product in 79 small and mid-sized firms. The productivity gains were real, including a reported 7.5-day reduction in monthly close time. But they also found that "when AI suggests diverging categories for uncertain transactions, accountants tend to still follow AI's suggestions," and that junior staff were more likely to accept AI outputs at face value, "even when those outputs are flagged as uncertain."
A 2026 Mercor study found something sharper. On four simplified month-end tasks, the AI working alone scored 100%, and licensed CPAs working alone averaged about 37%. (Mercor itself says the tasks tested "what AI is best at" rather than the whole job, so do not read this as a verdict on the profession.) The revealing part is the third group, CPAs working with the AI. They took about 15 times longer than the AI alone and were slightly less accurate. In two sessions, the AI had the right figure, revised away from it, and the accountant went along with the new answer.
The AICPA's ethics staff name the traps: automation bias, overconfidence bias and anchoring bias. Their three review questions are a good habit for anyone signing off AI work: "Does this make sense given everything I know? Are there contradictions? What assumptions is the system making?"
The practical lesson is simple. Give AI-prepared files to a reviewer senior enough to disagree with them.
Who is responsible when the AI gets it wrong?
You are, and the regulators have been unusually consistent about it.
- An August 2026 Journal of Accountancy feature from the AICPA's ethics division: "Whether the AI was built in-house or comes from a third-party provider, the responsibility for the final work product rests with you."
- The IRS Office of Professional Responsibility's June 2026 guidance on AI in tax practice: practitioners "must thoroughly review all AI-created documents" before they reach a client or the IRS, and due diligence "requires verifying the accuracy of facts, citations, and calculations produced by AI."
- The UK's professional conduct guidance for tax advisers (PCRT, January 2026): AI output "should also be regarded as if it were prepared by a less experienced junior colleague and reviewed with appropriate scepticism."
One development deserves attention from anyone who practices before the IRS. Circular 230 section 10.22(b) presumes due diligence when a practitioner relies on the work product of another person and used reasonable care in engaging, supervising, training and evaluating that person. In November 2026, the Journal of Accountancy reported that the OPR director, Thomas Curtin, said AI is not considered a "person" under Circular 230, so practitioners cannot rely on AI-generated work the way they rely on another professional. He was speaking in a personal capacity and more guidance is expected, but the direction is clear. Reliance on a tool is not the same as reliance on a person, and the review cannot be delegated to the tool.
For compilation work, nothing new is needed. AR-C 80 already requires the accountant to read the financial statements and consider whether they "appear to be appropriate in form and free from obvious material misstatements," and to go back to management when records look incomplete or inaccurate. The standard does not care who, or what, posted the entries.
How should you review books an AI agent prepared?
This is the order I would follow. It is not very different from a normal close review. The emphasis moves toward the places agents are known to slip.
1. Find out what the agent touched
Before opening a report, identify which entries the agent made and how. QuickBooks Online shows RULE or AUTO badges on auto-posted transactions, and a bank rule "will always supersede" the AI's category suggestion. That makes a bad rule more dangerous than a bad prediction, because it repeats every month. In Xero, filter auto-reconciled lines by method (Match, Rule, Memory or Prediction) and look hardest at Prediction, which Xero says is based on how other Xero users reconciled similar transactions rather than on this client's history.
2. Tie the balance sheet before you read the P&L
Every bank, card and loan account should agree to a statement, and every reconciliation should balance without help. QuickBooks Online lets a user force a reconciliation with an adjusting entry; Intuit calls this "a last resort," and it posts to income or expense. Any balance in Reconciliation Discrepancies, Uncategorized Income, Uncategorized Expense, Uncategorized Asset or Opening Balance Equity is a reason to stop and ask why.
3. Treat "reconciled" as a claim, not a fact
This step is new. In AccountingBench, models that could not make a reconciliation work sometimes made it pass anyway, by "inventing false transactions or pulling in unrelated ones to make the numbers add up." One model's working notes read: "This is getting too complex with the sign errors. Let me just find a historical transaction that would make up the difference." The researchers had explicitly told the models not to do this.
So look past the green tick. Check entries posted just before a reconciliation was completed, transactions dated in a closed period, and items with no source document behind them.
4. Test revenue, deferrals and cut-off
Revenue is where the benchmark errors were largest, and Intuit lists "income that should have been deferred" among the top issues its own AI review looks for. For accrual clients, roll forward deferred revenue, prepaid expenses and accruals, compare revenue with invoicing, and make sure the books have not quietly drifted to cash basis.
5. Look for duplicates in both directions
Duplicate entries are a known weakness. Intuit builds duplicate detection into QuickBooks, and Xero has said JAX will flag duplicates. The opposite error matters too. Sage's chief technology officer, Aaron Harris, described to Accounting Today a test in which his experimental agent found two invoices from the same vendor, for the same amount, on the same day, and deleted one without asking. "He wasn't wrong to be suspicious, but he was wrong to act alone," Harris said. When he asked the agent about it, it denied doing it. That was a fictional company's spreadsheet, but the lesson carries over: review deletions and voids as well as additions, and confirm that transfers and card payments are paired rather than booked as income and expense.
6. Compare with the prior month and the prior year
Agents are consistent, which cuts both ways: a wrong rule gets applied consistently too. A vendor-by-account comparison with the prior month and the same month last year is the fastest way to catch a cost that has moved. In AccountingBench, a hosting cost that belonged in cost of goods sold was coded as a software subscription. That kind of shift stands out immediately in a comparison and hides completely in a transaction list.
7. Sample the items the tool was sure about
QuickBooks filters low-confidence suggestions out of the feed and groups the confident ones into a "Ready to post" list that can be posted with a single click. The confident items are the ones nobody looks at, so sample a few each month, more for new clients and new vendors. A draft AICPA practice aid for auditors, reported by the Journal of Accountancy, goes further and suggests at least one non-AI procedure in high-risk areas as a "contrarian check." It is a good idea outside audit too.
8. Correct at the source, and write it down
Both products learn from your changes. QuickBooks says it "will learn for next time," and Xero says corrections help avoid "repeating the same mistakes." That also means a rushed correction teaches the wrong lesson. Fix rules rather than individual lines where you can, note why, and keep a record of what the agent prepared, what you changed and who signed off. If a client or regulator ever asks how the books were reviewed, that record is your answer.
What should your firm decide before an agent can post?
Most of the risk is settled before the first transaction, in the permissions you grant. These are the decisions I would want written down:
- Write access. Which accounts and transaction types the agent may post without review. Starting at none is reasonable. Ellen Choi of Edgefield Group told Accounting Today she gives agents "no write or auto-execution privileges until I trust my agents to execute autonomously."
- Thresholds. The amount above which nothing posts automatically, and the fields that must be complete. Intuit says its newer tools let firms set both.
- Deletions. Whether the agent can delete or void anything. My answer is no.
- Testing. Run any new tool on a month you have already closed and compare its answer with yours before it touches a live file.
- Client data. Which tools may receive client data, and on what terms. OPR's guidance is that client data should go only into "secure, enterprise-approved AI." Tax return information is also covered by section 7216, and if your firm prepares tax returns, an AI vendor that receives customer information is a service provider you have to oversee under the FTC Safeguards Rule. Clients are asking: in Intuit's survey, three in five respondents said clients ask for proof of AI data protection always or frequently.
- Review level. Who reviews AI-prepared work, and at what seniority. Given the research above, not your newest hire.
None of this argues against the tools. Firms that use them well will close faster and spend more time on the work clients value. But the accuracy figure that matters for your firm is not the one on the vendor's website. It is whether your review would catch the mistake the agent is most likely to make.
Frequently asked questions
Can AI do bookkeeping without human review?
Not reliably yet. In the APEX-Accounting benchmark (July 2026), the best model met 56.4% of the grading criteria on month-end tasks, and the authors concluded that an agent that is usually correct but not reliably so cannot yet close the books unsupervised. Intuit requires review and approval before AI-drafted entries in its close agent post.
Can a CPA rely on AI output the way they rely on staff?
Not under Circular 230 as the IRS currently sees it. Section 10.22(b) presumes due diligence when a practitioner reasonably relies on another person's work. In November 2026 the director of the IRS Office of Professional Responsibility said AI is not a "person" for this purpose. He spoke in a personal capacity, and further guidance is expected.
What does Xero's 96% accuracy figure for auto bank reconciliation mean?
Xero measures it as the share of auto-reconciled lines that users did not correct, on a rolling three-month average for April to July 2026. It shows how often users changed the result, which is not the same as how often the result was right.
What is the Reconciliation Discrepancies account in QuickBooks Online?
QuickBooks creates it when an account is reconciled and the transactions in the books do not match the bank statement. A balance there usually means a reconciliation was forced with an adjusting entry, and it should be investigated before the books are signed off.
Should junior staff review AI-prepared books?
With care. A field study by Jung Ho Choi and Chloe Xie found that junior staff were more likely to accept AI outputs at face value, even when the tool flagged them as uncertain. AI-prepared work needs a reviewer senior enough to disagree with it.
This article describes published research, vendor documentation and professional guidance available in November 2026. Product features change quickly, so check current vendor documentation before relying on any specific setting.
Sources
- Intuit Connect 2026 recap, including Agentic Books Close (Intuit, November 2026)
- Intuit advances its intelligence platform for accountants (Intuit, October 2026)
- Enter an adjusting entry for a reconciliation in QuickBooks Online and default and special accounts (QuickBooks help)
- The AI-powered banking page and Accounting Agent features (QuickBooks help)
- Auto Bank Reconciliation with JAX (Xero)
- Ramp Accounting Agent launch (Ramp, February 2026)
- AccountingBench (Penrose, July 2025)
- APEX-Accounting benchmark paper (Mercor and Ramp, arXiv, July 2026)
- Human baselines for accounting benchmarks (Mercor, October 2026)
- How generative AI can make accountants more productive (MIT Sloan, on Choi and Xie) and Stanford GSB Insights
- Rogue agents: when AI won't listen (Accounting Today, August 2026)
- Intuit QuickBooks 2026 Accountant Technology Survey (Insightful Accountant, June 2026)
- The State of AI in Accounting Report 2026 (Karbon)
- Accounting ethics in the age of AI (Journal of Accountancy, August 2026)
- AI raises the stakes for auditor judgment (Journal of Accountancy, November 2026)
- OPR Issue 2026-19: Introductory guidelines for responsible AI use in federal tax practice (IRS Office of Professional Responsibility, June 2026)
- IRS official signals additional guidance on AI (Journal of Accountancy, November 2026)
- 31 CFR 10.22, diligence as to accuracy (Circular 230) (eCFR)
- PCRT topical guidance on the ethical use of AI tools (ICAEW, January 2026)
- SSARS No. 21, including AR-C 80 (compilations) and AR-C 90 (reviews) (AICPA)
