Build an AI System to Detect Expense Fraud Before Approval

Small finance teams often rely on the built-in checks from tools like Expensify or Ramp. Expensify's SmartScan can digitize a receipt, but it struggles to identify a slightly altered photo of the same receipt submitted two weeks later. The system lacks the image similarity logic to cross-reference a new submission against an employee’s entire history, letting near-duplicates pass through unchecked.

Corporate card platforms like Ramp or Brex use rigid, rule-based policies. A rule might block a transaction over $500, but it cannot detect an employee who consistently submits meal expenses for $49.95 to fly just under a $50 manual approval threshold. For a small team, these platforms create a false sense of security while missing the most common forms of low-level, high-frequency fraud that add up over time.

Consider a 40-person company where a finance manager reviews reimbursements. An employee submits two Uber receipts on the same day for a similar route. The system approves both because they are unique transactions. The manager must manually open both receipts, compare the destinations, and spend 15 minutes investigating a $25 expense. This manual spot-checking is inefficient and catches only the most obvious issues.

The structural problem is that these tools operate on a per-transaction basis. They are not designed to analyze an employee’s spending behavior over time, compare it to their team's average, or identify statistical outliers. Solving this requires a system that can build a historical baseline of what is normal for your specific company and flag deviations from that pattern.

The first step is a data audit. Syntora would connect to your expense management platform's API and pull the last 12 months of approved expenses. This historical data is used to build a profile of normal spending behavior for every employee and department. The audit identifies the most common types of policy exceptions and potential fraud, which directly informs the features the detection model will use.

We built financial systems that automatically categorize bank transactions from Plaid into a PostgreSQL ledger. A similar approach applies here. A custom system would use a Python service to extract structured data from receipts and an image embedding model to create a unique fingerprint for each receipt photo. This fingerprinting allows the system to catch visually similar duplicates that OCR-based tools miss. The core logic uses an isolation forest model, deployed via FastAPI on AWS Lambda, to flag anomalies against the historical baseline.

The delivered system integrates directly into your current process. When an employee submits an expense, a webhook triggers the detection model. Within 5 seconds, if the expense is flagged, a comment is automatically added in your expense software explaining why (e.g., 'Possible duplicate of expense #4312 from 3 weeks ago'). Your finance team sees the alert in context, investigates, and denies the expense before any money moves.