AI for accountants: what it can do on the close, and where it fails

Most writing about AI for accountants either promises a faster close or warns about hallucinations. Neither helps a controller decide what to try on Monday. This article uses recorded runs on the books of Northwind Supply Co., a fictional industrial-supplies distributor in Cleveland, Ohio, with about 95 staff and roughly $45 million a year in sales. The data is invented, so every figure here can be checked against an answer key, and the failures are as useful as the successes.

What worked on the October close

In the Week 3 recording, an AI agent prepared the October accrual entries. It had to read the ledger, apply the rules, calculate five accruals and draft the journal entries for a controller to approve. All five accrual amounts matched the answer key. The draft entries balanced, with debits and credits both at 169,373.31, and every account it used existed in the chart of accounts.

The agent also caught its own small error. Its first script produced a depreciation figure of 40,410.00 and a total of 169,373.47. It then recommended 40,409.84 itself and flagged a 0.16 residual on one fixed asset. That is the kind of work AI does reliably well: repetitive arithmetic across many lines, with a consistent method, shown step by step so a person can inspect it.

In Week 4, an agent-assisted bank reconciliation tied to 1,779,474.31 on both the bank side and the book side. A screen over the 78 journal entries for October ranked the planted red-flag entries first, in the same order as the answer key.

What a controller should take from that

Three habits separate a useful result from a risky one.

1. Ask for the working as well as the answer. The accrual run was trustworthy because the arithmetic could be re-performed line by line. A bare total from a chat window gives you nothing to check.
2. Check against the source. The duplicate-payment finder built in Week 2 flagged three pairs of possible duplicate payments totaling 23,610.56. The tool's own output called them possible duplicates, not a final answer. One pair differed in the invoice number (88120 against 81120). A person still has to look at each pair.
3. Keep the approval with a person. In the same Week 3 run, a second agent passed on an approval. The first agent declined it and did not write the final journal-entry file, because it could not stand in for the controller's own approval. Build that checkpoint into any workflow you design.

Where it failed

The failures matter more than the successes, because they tell you where review time belongs.

• In Week 8, an AI reviewer checked a close memo that contained 14 planted errors. It flagged 13. The one it did not flag was an estimate stated as a fact. It did quietly correct the wording in its corrected memo, but it never told the reader that the original was wrong. A reviewer that finds 13 of 14 is useful. A team that believes it finds 14 of 14 is exposed.
• In Week 5, an AI tested all 612 invoice records against five attribute checks. It could not test whether an approval really exists behind a "Yes" flag on 245 invoices, whether a shipping document is genuine, or whether the register was complete. It said so, which is the right behavior, but a reader skimming the exception count would miss it.
• The Week 4 and Week 5 recordings also contain small mistakes by the AI, such as a misstated discount dollar amount and a gross-margin definition that differed from the books. "Checked against a key" is true of every recording. "Everything matched" is not.

A fair summary

AI is good at the mechanical parts of the close: drafting entries from rules, tying out reconciliations, screening whole populations of records and finding patterns a person would not scan for. It is weaker where the answer depends on judgment, on facts that are not in the data, or on a definition that differs between your books and the model's assumption. It also fails in ways that look confident.

Each of these results is one run on one model. It shows that something is possible, not that it will happen every time. A model update, a changed prompt or a different month can change the outcome, which is why the course teaches re-testing and writing down a pass mark before you rely on a tool.

How to start without risking client or company data

Use invented data first. The Northwind books are invented, so they are safe to paste into any AI tool, and you can try a prompt, an agent or a screen on a full ledger without pasting a real record into any tool. Once a workflow works there, classify your real data as green, amber or red before anyone uses a real record. Tax return information, in particular, has its own disclosure rules under IRC §7216.

The course behind these recordings

Ledger & Agent's first course is AI You Can Sign Off On, a four-week course. The next cohort opens soon. Each week brings one 60-minute live session, recorded and posted within 24 hours, about one to two hours of coursework, and written feedback that an AI drafts, with a person approving every return for revision, a sample of completes and every capstone score. At the end of week four you hand in one workflow redesigned with AI, with its checks and controls, run on the Northwind books. The eight-week AI for Accounting Professionals program follows it, at $2,950 per seat; for team training, contact us. It adds eight weeks of 60-minute live sessions, a discussion forum shared with the whole cohort and a capstone adoption plan for your own function, with a recorded AI run every week, including the ones described above.

Enroll now

---

General education, not legal, tax, audit or security advice. Confirm the rules (AICPA Code, IRC §7216, FTC Safeguards Rule, NIST frameworks, auditing standards) against current primary sources before you rely on them. Northwind Supply Co. is a fictional company and all of its data is invented for teaching.