Build an AI System to Stop Expense Fraud

Most SMBs start with tools like Expensify. Its 'SmartScan' OCR captures receipt totals but fails on context. The system can read a $200 dinner receipt but cannot determine if the dinner was with a client or a spouse. A finance manager must still manually check the memo, cross-reference calendar invites, and guess if the expense is compliant. The automation just moves the manual work from data entry to data verification.

Card-based platforms like Ramp or Brex enforce hard spending limits, which works for predictable costs like software subscriptions. These platforms fail with dynamic, project-based expenses. If a project manager needs to buy $2,500 in materials from a new vendor to meet a deadline, the card gets declined. The approval workflow is rigid and requires manual overrides that defeat the purpose of the platform.

Consider a 30-person consulting firm. An account director takes a client to a sporting event, a valid expense. The ticket costs $400, but the company policy limit is $250 for client entertainment. Expensify flags the amount but cannot see the attached PDF contract that pre-approved the specific expense. The finance manager wastes 15 minutes emailing the account director to confirm the exception, delaying reimbursement and wasting time on a legitimate expense.

The structural problem is that off-the-shelf tools rely on a fixed set of universal rules. They check if an expense is over a dollar limit or submitted after 30 days. They cannot learn a company’s unique spending DNA. These tools don't know that travel to a specific client’s city always involves higher-than-average hotel costs, or that the engineering team’s software purchases spike in Q4. You are forced to fit your business logic into their software’s constraints.

Syntora's approach begins with auditing your last 12 months of expense reports and your written expense policy. We map every category, approval flow, and known exception. This audit identifies the top 3-5 categories where out-of-policy spending is most likely to occur, which becomes the initial focus for the AI model.

We would build a FastAPI service hosted on AWS Lambda that connects to your bank data via Plaid, just as we did for our own financial ledger that syncs in under 3 seconds. For receipt submissions, we would add an endpoint that uses an OCR model to extract text from images. A small AI model, fine-tuned using the Claude API on your specific policy rules, then classifies each expense against your ruleset, flagging anomalies with a 95% or higher confidence score for manual review.

The delivered system provides a simple dashboard showing all submitted expenses with their risk score from 0-100. Approved expenses are batched for payment, and journal entries are automatically created in your PostgreSQL ledger. The entire process for a batch of 50 expenses would take under 60 seconds. The hosting cost on AWS Lambda for a team of 30 is typically under $50/month.