Audit runs on sampling. Faced with a company posting hundreds of thousands of entries a year, the firm could only pull a small random slice and ask for backup. The clumsy errors showed up; the coordinated fraud, the quiet theft, the manipulation, sat in the 99% no one looked at, and the firm carried the exposure.
A random sample of a few dozen transactions out of thousands gave nowhere near enough coverage, and it left the firm exposed to professional-negligence claims if an audited company later collapsed on something the sample happened to miss.
Finding the outliers by hand was its own problem. Interns spent weeks combing enormous data extracts in spreadsheets, and the process was slow, tiring and easy to lose focus in, exactly the conditions under which a subtle anomaly slips past.
And the real fraud was designed to hide. Payments split to stay under an approval threshold, entries keyed in the small hours, fictitious suppliers, these are patterns a random sample is almost guaranteed to walk straight past, because they do not look wrong one line at a time.
Replace the sample with the full population, and let the algorithm decide where a human should look.
We put in an engine that ingests the client's entire year of journal entries and examines the complete population in a fraction of a second, so the audit no longer rests on a random slice of a few dozen transactions.
Anomaly-detection models flag the tell-tale patterns instantly, a fictitious supplier sharing an employee's email address, transactions posted at weekends, repeated reversing entries, and suspiciously round numbers, the coordinated abuse that never looks wrong one line at a time.
The responsible partner gets a dashboard that ranks every area of the financial statements by risk and directs the team's effort exactly to where the algorithm found a mismatch or a high-risk pattern, instead of spreading it thin across everything.
Because the whole population is analysed, the firm can show it looked at everything and document the coverage, which is exactly the standard a strict audit regime expects and a sample can never demonstrate.
The weeks that interns spent scrolling through extracts disappeared, and their time moved to the audit judgement that actually needs a person once the engine has surfaced the candidates.
The same scan that protects the firm surfaces real issues for the audited company, cash leakage and weak separation of duties, turning the audit from a sign-off into advice the client values.
Full coverage, real findings, and less time on site.
The audit moved from a random sample to analysis of 100% of the data, which lifted audit quality sharply and drove the firm's professional-negligence risk down towards zero. It also caught several real cases of cash leakage and weak separation of duties at clients, work the client valued well beyond the sign-off.
The on-site fieldwork the interns used to carry fell sharply, which lowered the cost of each engagement and let the firm bid more competitively for new audit work without cutting corners on quality.
A sample is a bet that the fraud is in the part you happened to pick. The value here is analysing the whole ledger, letting the model surface the patterns a person cannot see across hundreds of thousands of rows, and pointing the audit team's judgement precisely where the risk sits. It is the same anomaly-detection discipline behind our fraud-detection work, aimed at the audit rather than the claim.
Blunt fraud rules were punishing honest customers and still missing the real thing. A real-time model scored risk in the moment, catching more fraud while letting genuine claims through without the friction.
Read the case → Fraud detection · Public sectorA public body's manual checks let fraud through and made honest applicants wait weeks. Risk scoring cut fraud losses around 50% and false alarms 80%, moving legitimate approvals from weeks to minutes.
Read the case →
Reef TRH