The challenge
A 16-person back-office team keyed carrier invoices, delivery slips and customs declarations into the company’s logistics system and SAP by hand. Volume had grown 35% in two years, and month-end closing regularly ran late.
Documents arrived by email, fax and paper scan, in Japanese and English, from more than 900 carriers and suppliers. Each used its own format, many carried hanko stamps and handwritten corrections, and an earlier OCR-template tool could not keep up.
Our approach
- 01
Measure before building
We sampled 5,000 real documents and labelled 2,000 of them. That gave us a test set before we wrote any product code, so every design decision could be measured.
- 02
Confidence scoring on every field
The system scores every field. High-confidence documents go straight to posting. Anything uncertain goes to a review queue with the doubtful fields highlighted.
- 03
Validate against the business
Extracted values are checked against orders, shipment records and carrier rate contracts. A plausible value that does not match the shipment record is still flagged.
What we built
Ingestion pipeline
Monitors shared mailboxes, the fax server and scanner folders, de-duplicates, splits multi-document PDFs and straightens scans.
Classification & extraction
Layout-aware parsing for Japanese tables, stamps and vertical text, plus LLM extraction into strict per-document-type schemas.
Validation engine
Business rules, three-way matching against orders and shipment records, and qualified-invoice registration number checks.
Review workspace
A web app where operators approve, correct or reject flagged fields. Every correction is fed back into the test set.
How it flows
- Email, fax & scans
- Classify
- Extract
- Validate vs core system
- Review queue
- SAP posting
How we measured it
Before release, every change runs against the 2,000-document test set. Accuracy is tracked per field, per document type and per source channel, with faxes tracked separately. Production corrections are sampled weekly and added back to the set.
Results
- 62% of documents flow from inbox to SAP without anyone touching them.
- Median processing time dropped from 2 days to 35 minutes.
- The team absorbed a 20% volume increase without new hires; four people moved to carrier and billing inquiries.
- Month-end closing now finishes on schedule.
Timeline
Weeks 1–2
Discovery
Document sampling, process mapping, test-set labelling
Weeks 3–5
Prototype
Extraction on the top 3 document types, accuracy baseline
Weeks 6–13
Build
Validation engine, review app, core system and SAP integration
Weeks 14–16
Rollout
Shadow mode, then staged go-live by document type