AI in accounting: what a 2026 study shows about human review
AI coding suggestions can help a reviewer, but they can also steer that reviewer toward a weak answer. A 2026 accounting study illustrates both effects. Before adopting a tool, test which mistakes your team catches, which it accepts and how much checking the work takes.
.jpg){kind=link}
What the study found
Jung Ho Choi and Chloe L. Xie's study, first published in the Journal of Accounting Research on 16 April 2026, combines a survey of 277 accountants with one platform's records from 79 small and mid-sized businesses. Most adoption findings describe associations rather than prove cause and effect.
A separate experiment tested a narrower question. Across 99 participants and 43 transaction-categorisation tasks, AI assistance raised accuracy by about 17.5 percentage points. It had no statistically significant effect on average total completion time.
For the same transaction, categories suggested less often across participants were less accurate on average. Participants were more likely to choose one of those categories when the AI proposed it. Such a suggestion was not always wrong, but it could influence the reviewer's choice.
The experiment concerns transaction categorisation, not tax advice, audit work or a complete month-end close. One platform's field results and a narrow experiment cannot establish every tool's performance or its long-term effects.
Sources and context, checked :
What audit regulators say about review
Guidance written for auditors does not set rules for bookkeeping, but its control thinking is useful. The FRC's March 2026 guidance on generative and agentic AI in audit describes outputs that leave out part of a task or reason incorrectly. Its mitigations include design, testing, training and review, and it accepts that review can be less extensive where testing supports the tool's quality and staff are trained.
ICAEW's May 2026 guidance for audit firms asks reviewers to stay alert to AI use the firm has not authorised, and raises client transparency and data protection safeguards. For an owner, the practical upshot is that the right amount of checking is something to measure on your own transactions, and which tools may see your records is a decision to make explicitly.
Sources and context, checked :
Run a pilot that counts the checking
Use a closed month or approved test data with a mix of routine and awkward transactions, and have someone code the same batch independently so you have a reviewed answer to compare against. Check the source document whenever the tool proposes something out of the ordinary, such as coding a regular stock supplier's bill as equipment. Do not load confidential client or employee data into a tool until its data handling has been approved.
Decide in advance what the tool may not do, such as posting journals or releasing payments, and who handles items it cannot classify.
| Measure | How to count it |
|---|---|
| Accuracy | Share of AI suggestions that match the independently reviewed answer |
| Errors caught | Wrong suggestions the reviewer corrected |
| Errors missed | Wrong suggestions the reviewer accepted |
| Unusual suggestions | Accuracy where the tool proposed a category the business rarely uses for that supplier or customer |
| Checking time | Minutes per batch, including opening source documents, compared with your current process |
Deciding after the pilot
If missed errors cluster in particular transaction types, take those items out of the tool's scope and send them to a person. Retest the narrower scope against the accuracy and review-effort limits agreed before the pilot. Keep problem cases restricted until the results support a change.