Improve Fraud Detection in Accounting Audits with AI

Small and mid-sized accounting firms typically rely on audit software like Caseware or AdvanceFlow, supplemented by extensive Excel work. These tools are built around Generally Accepted Auditing Standards (GAAS), which permit statistical sampling. An auditor might test 200 transactions out of a 50,000-transaction ledger, leaving 99.6% of the data completely unexamined. This creates a massive blind spot where fraud can hide.

In practice, this means an auditor following a standard checklist might miss a sophisticated scheme. For example, a fraudster might create a shell company with a name very similar to a real vendor and submit multiple invoices for $4,500, just under the typical $5,000 threshold for mandatory review. A random sample is statistically unlikely to pick up these specific transactions, and a rules-based system won't flag them because they fall below the review limit. The auditor signs off on the financials, unaware of the systematic theft.

The structural problem is that traditional audit tools are designed for compliance, not for exhaustive data analysis. Their data models are rigid, making it difficult to join ledger data with external sources, like checking vendor addresses against employee home addresses. They are fundamentally reactive, applying fixed rules to small samples. They cannot learn the unique transaction patterns of a specific business and flag deviations from that norm.

Syntora's approach extends from our direct experience building an internal accounting system with an Express.js backend and a PostgreSQL ledger. We would start with a discovery phase to map your clients' data sources. We connect directly to the APIs of systems like QuickBooks Online, Xero, or NetSuite to pull the full general ledger, chart of accounts, and vendor lists. This initial audit creates a clean, standardized dataset for analysis.

The core of the system is a set of Python scripts using libraries like pandas for data manipulation and scikit-learn for modeling. Instead of simple rules, we would train an anomaly detection model, such as an Isolation Forest, on 12-24 months of transaction history. This model learns what a 'normal' transaction looks like for that specific business and assigns a risk score to every new entry. The model is wrapped in a FastAPI service and can be deployed on AWS Lambda for efficient, on-demand processing.

The delivered system is not a black box. It's a dashboard that provides your audit team with a prioritized list of the top 50 riskiest transactions for each client. Each flagged item includes a human-readable explanation, like 'Journal entry posted at 2:00 AM on a Sunday' or 'Payment to a new vendor approved by a non-finance manager.' Your team uses this intelligence to guide their substantive testing, focusing their expertise where it matters most.