Automate Accounting Workflows with a Custom AI System

Most teams start with an accounts payable tool like Bill.com. It's effective for standard invoices but fails when approval logic gets complex. For example, an invoice over $5,000 for a specific project requires the project manager's approval, but only if the vendor is a subcontractor. Off-the-shelf tools cannot support this multi-conditional logic, forcing you back to manual email forwarding. Their per-user pricing also penalizes growth, with costs rising to over $400 a month for a small team.

Some businesses try a DIY approach with generic OCR software to read scanned invoices. These tools extract text but lack the intelligence to understand context. They confuse PO numbers with invoice numbers and struggle to parse complex line items. This results in messy, unreliable data that an AP clerk still has to manually correct, which defeats the purpose of automation.

A 30-person agency we worked with had this exact problem. Their AP clerk spent every Monday forwarding invoices to project managers on Slack and chasing approvals. A $10,000 invoice for a critical software subscription was paid two weeks late because the approval request was buried in a Slack channel, resulting in a service interruption and a 5% late fee.

Syntora would begin an engagement by designing an ingestion pipeline. This would involve setting up a dedicated email address, for instance, ap@yourcompany.com, to automatically forward incoming invoices to an AWS Lambda function. This function would be designed to identify and extract PDF or image attachments. For initial processing, the architecture would involve a call to the Claude 3 Sonnet API, which can classify documents as invoices, receipts, or other types.

For documents identified as valid invoices, the system would be configured to use a more advanced Claude 3 Opus API call. This extracts key-value pairs such as vendor name, invoice number, due date, individual line items, and the total amount into a structured JSON object. Syntora would tune this extraction model using a sample of your historical invoices to achieve high accuracy for your specific document types. The extracted data would then be stored in a dedicated Supabase database table.

The core business logic would be implemented within a FastAPI service. Syntora would codify your specific approval matrix into Python functions. For example, rules like 'if amount exceeds $5000 and the department is Marketing, then route for CMO approval' would be programmed to ensure accurate routing. This programmatic approach offers greater flexibility compared to visual workflow builders. Approval requests could be sent as interactive messages in platforms like Slack or Microsoft Teams, allowing for rapid, one-click approvals.

Upon final approval, a webhook would trigger a function designed to connect directly to your chosen accounting software's API, such as QuickBooks Online or Xero. This function would create a new bill, pre-filling vendor details, amount, and line-item information from the extracted data, and attach the original PDF. Your accounting team would then review and schedule the payment.

An engagement typically spans several weeks, factoring in discovery, development, testing, and deployment. To facilitate development, the client would need to provide examples of unique invoice formats and a detailed definition of their approval matrix. The deliverables would include a deployed, custom-built invoice processing and approval system, along with documentation and knowledge transfer for your team.