AI Scored 97.8% at Bookkeeping vs 79.1% for Human Accountants — What Owners Should Check

A new round of coverage is putting a striking number in front of business owners: in one benchmark, an AI bookkeeping system correctly categorized 97.8% of transactions, while outsourced human accountants got 79.1%. The test comes from Digits, a company that sells AI accounting software, so the headline deserves a close read. Here is what was measured, what was left out, and what it means if you hand your books to an outside bookkeeper or an AI tool.

What the Benchmark Tested

Digits ran its "Beyond the AI Hype" benchmark in June 2026. It compared three groups on the same job: Digits' own Agentic General Ledger, 13 general-purpose frontier AI models, and 12 outsourced human accountants. Every participant worked through the same 2,000 real transactions from four small businesses. Answers were checked against a U.S. GAAP ground-truth set that professional accountants had reviewed.

The task was narrow: one-shot transaction categorization, meaning deciding which account each purchase or deposit belongs in, with no back-and-forth questions.

The Numbers

• Human accountants: 79.1% accuracy.
• Digits' AI system: 97.8% accuracy.
• General-purpose models: Per press coverage, five publicly available frontier models also beat the 79.1% human score. The coverage calls that a first.

Digits also claims its system made roughly ten times fewer mistakes than the human accountants, ran about 8,500 times faster, and cost about 24 times less per transaction. Those are the vendor's figures, as reported by trade press.

Why You Should Read It With Caution

The 97.8% score belongs to the system built by the company that ran the test. That does not make it wrong, but it does mean the headline comparison is vendor-reported rather than independently verified. Coverage of the benchmark also notes that it measured only categorization. It did not test tax strategy, audit judgment, client advisory work, or fraud detection. Digits itself says judgment calls and controls still rest with people.

A separate benchmark by DualEntry Labs, covering 101 real accounting tasks, reportedly found that top models still miss enough on harder, multi-step work to need human review. In other words, AI looks strong at sorting transactions into the right buckets and weaker at the messy, multi-step judgment that makes up the rest of accounting.

What 79.1% Means for Your Current Bookkeeper

The less-discussed number is the human one. If outsourced accountants in this test averaged 79.1%, roughly one transaction in five landed in the wrong category. That does not mean your bookkeeper is that inaccurate, since the test used one-shot categorization without the ability to ask you questions. But it is a useful reminder that errors in books are normal, and that owners rarely look for them.

Miscategorized transactions are not harmless. They can distort your profit, hide a cost that should be tracked, or put a personal expense in a business account. Cleaner categories make a profit and loss statement reliable enough to make decisions from, whether that is a loan application, a hiring choice, or a tax filing.

How to Check Your Own Books in 30 Minutes

Whether a person, software, or an AI tool does your categorization, a short monthly review catches most problems. This is general practical advice, not a guarantee.

• Scan the "uncategorized" and "miscellaneous" accounts. A large balance in either one means the sorting was not finished.
• Sort by largest transactions. Check the ten biggest expenses and deposits each month. Big errors do the most damage.
• Look for personal spending. Mixed accounts are a common source of mistakes. A dedicated business bank account makes this review much easier.
• Check recurring charges. Software subscriptions, insurance premiums, and loan payments should land in the same category every month. Insurance, for example, belongs with your business insurance costs.
• Compare your bank balance to the books. If they do not match, something is missing or duplicated.
• Review contractor payments. Payments to workers can end up in the wrong place, which matters for 1099 reporting. Our guide on independent contractors vs employees explains why classification matters.

If You Are Thinking of Using AI for Your Books

The benchmark suggests the technology has reached a point where it is worth testing, but test it carefully. A sensible approach for a small business looks like this:

• Start with a pilot. Run the tool on one month of old transactions and compare its categories with your accountant's final version.
• Keep a human reviewer. Let AI do the first pass and have a bookkeeper or accountant approve it, especially before tax time.
• Ask about data handling. Find out where your financial data is stored and whether it is used to train models.
• Do not skip the advisory part. Categorization is only one piece. Entity choice, deductions, and cash-flow planning still benefit from professional advice. If you are weighing how to structure the business, start with LLC formation.

The Jobs Question

Trade coverage of the results suggests junior and offshore bookkeeping roles are the most exposed, while also noting that adoption in accounting has historically lagged the technology because of client trust and liability concerns. For owners, the practical effect is likely to be lower prices and faster turnaround for routine bookkeeping, with human accountants shifting toward review and advice. Whether that happens in your market, and how quickly, is not something this benchmark can answer.

The Bottom Line

The 97.8% versus 79.1% result is a vendor-reported finding on one narrow task, but it points in a clear direction: AI is getting good at sorting transactions. Use that as a reason to look at your own books this week, ask your bookkeeper how errors are caught, and trial any AI tool against a month you already trust. Whoever does the work, you are the one who signs the tax return, so a quick monthly review remains the cheapest insurance you can buy.

Photo by Carlos Muza on Unsplash