The people in this story are made up. The problems are the ones we hear most.
Flowchart
The workflow in Document Processing Factory
- StartDriver photo or scan, sent by email
- StepIngestion
- StepClassification (bill of lading / delivery receipt / other)
- StepExtraction (shipment number, date, consignee, pieces, weight, signature present)
- StepSave to database
- DoneBilling queries
The billing clerk and the box of receipts
Before
- Driver hands in paper
- Clerk retypes
- Typos and missing receipts
- Invoice delayed
After
- Photo or scan emailed
- Fields extracted
- Row in database
- Invoice ready
Marcus does billing for a regional trucking company with a couple dozen trucks. Customers are invoiced only after delivery is proven, and proof is paper: a signed delivery receipt, sometimes a bill of lading with handwritten notes about shortages or damage.
Drivers hand in paper at the end of the week or snap photos on their phones. Marcus types the shipment number, delivery date, consignee, piece count and any exception notes into the billing system. A missing receipt means an invoice can't go out. A typo in a shipment number means it goes to the wrong customer. When a customer disputes a charge months later, someone digs through boxes for the original.
The information exists. It's just locked in images.
OCR alone isn't the whole answer
Plain OCR
- Image
- Wall of text
Document processing
- Image
- Document type
- Named fields
- Database columns
Plain OCR turns an image into text. That helps with searching, but Marcus doesn't need a wall of text; he needs the shipment number in the shipment number column. Getting from page to field means knowing what kind of document it is and where each value lives on it, even when two customers' forms look different. (We cover the difference in OCR vs AI document processing.)
The workflow in Document Processing Factory
- Scan or photo by email
- Ingestion
- Classification: bill of lading / delivery receipt
- Extraction
- Database row by shipment number
- 01Ingestion. Scans and driver photos are sent to a billing email address, and an ingestion widget brings them in from there.
- 02Classification. A classification widget, trained on Marcus's own documents, separates bills of lading, delivery receipts and everything else.
- 03Extraction. For each type he picks the fields:
| Document | Fields |
|---|---|
| Delivery receipt | Shipment or PRO number, delivery date, consignee, pieces delivered, exception notes |
| Bill of lading | Shipment number, shipper, consignee, pieces, weight, ship date |
- 04Save to a database. Every document becomes a row keyed by its shipment number, so billing can query it.
What changes for billing
- Invoices wait for the pileBillable when processed
- Missing proof unnoticedQuery shows gaps
- Disputes mean digging boxesSearch by shipment number
- Damage notes buriedFilterable field
- Invoices go out sooner. A delivery receipt becomes billable data when it's processed, not when the clerk reaches it in the pile.
- Missing proof is visible. A simple query shows shipments delivered with no receipt on file.
- Disputes get easier. Search by shipment number instead of digging through boxes.
- Exceptions stand out. Damage and shortage notes become a field you can filter on, so a person reviews those before the invoice goes out.
The same pattern elsewhere
Bring in → Sort → Extract → Store
- Packing slips
- Field service tickets
- Inspection checklists
- Paper archives
Any business with paper that must become rows has this problem: warehouse packing slips, field service tickets signed by customers, inspection checklists, old paper files a company wants searchable. The steps are the same: bring in, sort, extract, store.
Before you trust the numbers
- Run side by side
- Test real handwriting
- Add validation checks
- Keep a pointer to the original
- Run it side by side. For the first weeks, compare extracted values against what a person would have typed.
- Watch handwriting. Handwritten notes are the hardest part of any scan. Test with your real ones.
- Validate what you can. A shipment number should match a shipment that exists; a delivery date shouldn't be before the ship date. Checks like these catch errors before they reach an invoice.
- Keep the original. Store where each original file lives alongside its row, so anyone can check the source.