AIToday
Large Language ModelsAI Coding AssistantsZenn AI/MLPublished: Oct 23, 2026, 13:00 JST

Audit tool cuts 383 entries to 25 with 11 rules

Send on LINE
Audit tool cuts 383 entries to 25 with 11 rules

3 Key Points

  1. What happened

    Running the v0.1.1 code with 11 rules on bundled fictional data on October 15, 2026, 383 journal entries produced 25 flagged entries and 53 total rule hits, since one entry can match several rules.

  2. Why it matters

    The tool is built to prioritize which entries an auditor reviews first, not to conclude fraud; a flagged entry still needs human inspection of supporting documents.

Not sure about something? Ask the AI

Questions and answers are published on this page.

Summaries like this, every morning.

Context & Analysis

Accounting knowledge shaped three of the 11 rules, starting with threshold avoidance: when a company requires senior approval above a set amount, the tool looks for entries just below that line — within 5% of it — because people may try to avoid approval. The rule does nothing unless approval thresholds are supplied, since thresholds vary by company. A second rule pairs the days before fiscal year-end with large amounts, and it only runs when a fiscal year-end date is given.

The rare-account-pair rule relies on the observation that normal entries repeat a small number of account combinations; combinations appearing only a few times are worth checking. That rule only looks at how often a combination appears, not at specific account names, and it skips processing when the population is under 50 entries. The tool then adds up severity points — high 10, medium 5, low 2 — to order which entries to review first; the developer stresses these points are not a probability of fraud. The design deliberately leaves false positives in place, since legitimate transactions can also sit just under an approval threshold, and the developer explicitly declines to claim that AI finds fraud.

FAQ
How many rules does the tool use?
The v0.1.1 version implements 11 rules. The publicly released version is v0.3.1 with 15 rules, confirmed on October 15, 2026.
Does a flagged entry mean fraud?
No. The tool only extracts entries for review; a rule hit alone does not establish fraud, and the developer states it cannot replace audit procedures or support a fraud conclusion.
Why is Benford's law not one of the 11 rules?
Benford's law looks at the population's distribution and cannot judge individual entries, so it is not included in the 11 rules or added to the score.
Send on LINE

AI news that matters for your work, delivered every morning.

Pick your industry and the AI tools you use, and get news related to your work every day.

Free · 30 seconds with Google · unsubscribe anytimeWhat is AIToday? →

Ask AI

Ask AI anything about this article. The AI reads this article, earlier AIToday articles, and Wikipedia, and cites its sources. Q&As are published on this page for other readers too.

Questions and answers are published on this page.

Related Articles

Next articleAnthropic pursues bug-free software via formal verification