How an AI application can help auditors detect fraud

Anomaly detection refers to a practice in which auditors detect accounting fraud by selecting samples among journal entries, also known as a general ledger (GL), and testing them to ensure accuracy. In a database of 100 million entries, maybe 10 will be cause for concern. This means that highly experienced auditors try and identify the needles in the haystack by sensing where audit risk might come from. They do this based on their knowledge about the clients, including their businesses, accounting policies and governance.

Naoto Ichihara, an Assurance partner for Ernst & Young ShinNihon LLC in the Tokyo office, always had a passion for programming. He develops models and systems for audit and was interested in how machine learning could be applied to accounting data. After surveying existing academic papers and algorithms, Naoto realized there was a better way to detect anomalies through machine learning, and he coded an AI solution that could sense anomalous entries in large databases — the first-of-its-kind in the auditing field. Never imagining himself an inventor, the technology was patented and Naoto built a team of auditors and developers to test and improve the solution’s detection method. This innovative tool was named EY Helix GL Anomaly Detector, or Helix GLAD.

Without the support of auditors, the tool has limited value, since it relied on them to evaluate the flagged entries and recommend action. The EY Helix GL Anomaly Detector team knew it would be difficult to convince auditors that an unproven tool could help them do their job better. So the team tested the solution against a data set where they had predetermined which journal entries were fraudulent. As the Assurance team watched the algorithm correctly uncover the fraudulent journal entries, they believed in its potential to help audit more accurately.

However, auditors had no way of knowing why the algorithm sensed any particular anomaly, which would be vital in evaluating its impact. The team devised a solution that leveraged data analytics to create visual maps of the flagged entries and the reason for their detection — the equivalent of an X-ray into its detection method.