Client receives 400 mixed documents daily via email. Invoices, contracts, forms, receipts, purchase orders all hitting same inbox.
Manual sorting: 45 minutes daily. "Is this an invoice or PO?"
Built 3-node classification system. Automatically routes everything.
THE SYSTEM:
Document Analysis → examines structure, phrases, layout
Classification Switch → routes based on type confidence
Error Fallback → uncertain items to review queue
2 seconds per document. 400 documents in 15 minutes. Zero human sorting.
HOW IT WORKS:
Document nodes identify types before extraction. Structure, phrases, layout patterns.
Invoices: vendor info top, line items middle, total bottom
Contracts: parties identified, terms sections, signatures
Forms: field labels, checkboxes, submission info
Purchase Orders: PO number prominent, quantities, shipping
THE ROUTING:
Invoice (>90% confidence) → Invoice workflow
Contract (>90% confidence) → Contract workflow
Form (>90% confidence) → Form extraction
Receipt (>90% confidence) → Expense tracking
PO (>90% confidence) → PO verification
Uncertain (<90%) → Review queue with preview
REAL NUMBERS:
Daily: 400 mixed documents
Invoices: ~180 (45%)
Contracts: ~80 (20%)
Forms: ~60 (15%)
Receipts: ~50 (12%)
POs: ~20 (5%)
Review: ~10 (2.5%)
97.5% classified automatically. Only 10 daily need human classification.
THE IMPLEMENTATION:
45 minutes setup classification. 2 hours connecting downstream workflows. Total 3 hours for system saving 45 minutes daily forever.
ROI: Positive after 4 days.
ACCURACY IMPROVEMENT:
Week 1: 91% accuracy
Week 4: 97% accuracy
Month 3: 97.5% accuracy
As system sees more examples, accuracy improves. Self-improving.
UNEXPECTED BENEFIT:
Classification logs show trends. Invoices spike Tuesdays. Contracts cluster month-end. Adjusted staffing accordingly.
THE LESSON:
Don't force humans to sort documents. Build intelligent routing. Humans only review genuinely ambiguous cases.
What document types are you manually sorting every day?