The paperwork nobody can use
Supplier documents arrive constantly - delivery notes, certificates of conformity, test reports, specification sheets, safety data. Each contains information the business needs, in a layout chosen by the supplier.
Most of it is filed and never used analytically. When a question arises - which batches came from which source, which certificates expire this quarter - it means opening files.
Concentrate on the top suppliers
Document formats follow a steep distribution. A handful of suppliers usually account for most of the volume, and each sends a consistent layout.
- Count documents by supplier over a representative period.
- Take the suppliers covering most of the volume and handle their formats specifically.
- Use a general approach for the remainder, with lower expectations.
- Keep a straightforward manual path for genuinely one-off documents.
- Revisit the distribution periodically as the supply base changes.
Trying to handle every format equally well from the start is how these projects overrun. The long tail is genuinely hard and rarely worth the same effort per document.
What makes documents hard
- Scans and photographs rather than digital PDFs
- Tables spanning pages, with headers repeated or not
- Multiple deliveries on one note, or one delivery across several
- Units and formats differing by supplier - kg against lb, dates in either order
- Handwritten annotations that change the meaning
- The same field named differently on each supplier's template
The date format issue causes quiet, serious errors. A date read day-first when it was month-first produces a plausible wrong answer rather than a failure, which is the worst kind of bug. Pin the expected format per supplier rather than guessing per document.
Validate against what you already know
Extraction accuracy improves substantially when checked against your own records. You usually know what you ordered, so an extracted quantity or part number can be validated rather than trusted.
| Extracted | Check against |
|---|---|
| Part number | Purchase order lines |
| Quantity | Expected delivery, with tolerance |
| Supplier | Supplier master |
| Date | Plausible window around the order |
| Batch or serial | Format rules for that supplier |
Mismatches become exceptions rather than errors, which is the right outcome. An extraction disagreeing with the order is exactly what someone should look at.
Fail loudly
The dangerous failure is a confident wrong value posting into a system. A low-confidence extraction must go to a person rather than defaulting, and the original document should always be one click away.
Store the source document reference with every extracted value. When something looks wrong six months later, being able to see the page it came from turns an investigation into a glance.
A misread date does not throw an error. It just quietly puts the wrong month in your system.