Build a Custom AI System for Accounts Payable

Most small finance teams rely on the basic features within their accounting software, like QuickBooks Online's bank rules. These rules are brittle, using simple text matching that fails if a vendor name has a slight variation. They cannot parse line items from a PDF invoice, forcing manual entry for every bill.

Tools like Bill.com or Melio offer a step up, but they lock you into their platform and their rigid workflows. If your business requires a multi-step approval process that depends on the project code or invoice amount, you cannot build that custom logic. You are forced to adapt your business process to the software's limitations, and their per-transaction or per-user fees add up quickly.

Consider a 20-person consulting firm where invoices arrive as PDFs via email. The office manager spends hours manually keying invoice numbers, amounts, and line items into QuickBooks, then chasing approvals in Slack. An invoice from a new software vendor needs to be coded to a G&A expense account, while an invoice from a subcontractor must be coded to a specific project's cost of goods sold. No off-the-shelf tool can reliably make this distinction because it lacks the context of your business operations.

The structural problem is that these tools are built for mass-market simplicity, not business-specific context. They cannot connect to your other systems to validate information, like checking a project budget in a spreadsheet before approving an expense. They are closed systems that treat AP as a generic, isolated task, not as an integrated part of your financial operations.

The engagement starts with a full audit of your current accounts payable process. We map every vendor, how their invoices arrive, and the specific business logic behind every approval. Syntora analyzes your last 12 months of invoices to understand categorization patterns and chart of accounts structure. This audit produces a data flow diagram and a technical specification that you approve before any build work begins.

The technical approach uses a dedicated intake system, often an AWS Lambda function triggered by an email. For PDF or image-based invoices, the system uses the Claude API to perform optical character recognition (OCR) and extract key entities like vendor, invoice number, line items, and amount. This extracted data is validated with Pydantic schemas and stored in a Supabase PostgreSQL database, creating a structured record of every bill.

The delivered system is a simple web application, deployed on Vercel, where you can review pending invoices. The interface shows the extracted data alongside a suggested journal entry. Approvers receive a Slack notification with 'Approve' or 'Reject' buttons. Once approved, the system uses the QuickBooks Online API to create the bill and schedule the payment, turning a 15-minute manual task into a 10-second review. The entire system costs under $50 per month to run.