AI fraud detection in accounting: the checks that actually find things

Start with the tests that pay for themselves

• Duplicate payments and duplicate documents.
• Entries that do not balance, or totals that do not agree to their lines.
• Postings into a closed period.
• Manual journals clustered in the last days of the financial year.
• Round-number payments just below an approval threshold.
• Payments at weekends or outside business hours.
• Write-offs and credit notes above a materiality level.
• Digit distribution testing across large populations.

Tuning beats adding

An audit that produces four hundred findings is ignored. Every check needs a materiality floor, a confidence score and aggregation, so that a bulk historical import produces one summary finding rather than hundreds of identical ones. The measure of a good engine is the ratio of investigated findings to raised findings.

What AI adds over fixed rules

Rules catch what you predicted. Models catch the shape of the exception: a supplier whose payment pattern changed, a cost centre whose spend broke its own seasonality, a description that no longer matches the account it posts to. The right combination is rules for the known frauds and statistical detection for the unknown ones.

Investigation needs evidence

A finding is only useful if it carries the entries, the amounts, the dates and a link to the source document. Anything less turns into a week of asking the ledger team to pull screenshots.

Controls still matter

Detection does not replace segregation of duties, approval limits, supplier bank-detail verification with callback, and locked periods. Detection tells you the control failed.

NanoFora runs these checks on your posted ledger with materiality and confidence on every finding, plus a review workflow. Read duplicate payment detection and the financial audit guide.