The problem
Someone on the team retypes vendor names, amounts, dates and terms from PDFs into a spreadsheet or accounting system. It's slow, boring and error-prone.
How it works
- Documents arrive by email, upload form or a watched folder.
- Text is extracted with OCR where needed, and the model maps it to your field list (vendor, amount, due date, terms, etc.).
- Each field gets a confidence score and is checked against simple rules like totals adding up and dates being valid.
- Clean records are written to your tracker while low-confidence fields are highlighted for a quick human check.
What it needs from you
- The list of fields you need from each document type
- 20–50 sample documents per type
- The destination (sheet, Airtable, accounting system)
- A reviewer for flagged fields
Why it's built this way
Field-level confidence means people review only the uncertain parts, not the whole document.
Rule checks catch errors the model can't see on its own, like totals that don't sum.
How it stays under control
- Per-field confidence with review threshold
- Validation rules (totals, dates, required fields)
- Original document linked to every record
- No payments or approvals triggered automatically
The path to autonomy
Document types that consistently pass validation can skip review. New layouts or vendors start supervised.