An accounts-payable outsourcing firm extracting invoices, receipts and contracts into validated JSON for its clients' finance systems.
The business
Bayview Payables, based in Atlanta, Georgia, runs accounts payable for about 200 mid-sized companies. It receives roughly 140,000 documents a month and must turn each into structured data for the client's finance system.
The documents
| Type | Share | Notes |
|---|
| Invoices | 70% | Hundreds of layouts; some handwritten; some from international suppliers |
| Receipts | 15% | Photos and scans, often low quality |
| Contracts | 10% | 5 to 80 pages; payment terms buried in clauses |
| Other | 5% | Credit notes, statements, cover letters, things nobody expected |
The pipeline today
- A prompt says "respond with JSON". About 4% of responses fail to parse (markdown fences, trailing commas, extra prose).
- The schema marks
vendor_tax_id as required. About 15% of invoices have no tax ID, and the model sometimes invents one. - Dates arrive as "03/04/2026", "4 Mar 26", "March 4th" and "Q1 2026" and come out in mixed formats.
- On about 8% of invoices the line items do not add up to the stated total.
- A monthly backlog of about 30,000 historical documents is needed by the next business day, at the lowest cost.
Quality and review
- Reported accuracy is 97% overall. Nobody has broken it down by document type or field.
- Human reviewers can check about 10% of documents. Today they review every tenth document in arrival order.
- Clients want reviewers' time spent where errors are most likely.
People
| Stakeholder | Cares most about |
|---|
| COO (sponsor) | Lower cost per document without more client complaints |
| Finance systems integrator | JSON that always parses and always matches the schema |
| Review team lead | Reviewers sent to the right documents |