Built document extraction system. 99.4% accuracy on test data.
Client lost $42,000 in first month. The 0.6% errors hit their highest-value transactions.
WHAT WENT WRONG:
System accuracy breakdown:
- 99.9% accurate on standard invoices
- 96% accurate on invoices $5,000-$10,000
- 91% accurate on invoices $10,000-$50,000
- 87% accurate on invoices over $50,000
The bigger the transaction, the worse the accuracy. And that's where the money was.
THE $42,000 INCIDENT:
Invoice: $420,000 construction vendor payment
Extracted amount: $42,000 (missed a zero)
Payment processed: $42,000
Vendor called: "Where's the other $378,000?"
Bank reversal fees: $850
Staff time to fix: 12 hours
Client trust damage: Severe
WHY HIGH-VALUE INVOICES FAIL:
- Custom contract formats
- Handwritten notes and exceptions
- Multi-page complex layouts
- Non-standard payment terms
- Executive approval signatures
All the edge cases. All on the expensive invoices.
THE FIX - THREE VALIDATION LAYERS:
LAYER 1: Format Validation
- Extracted amount matches currency format?
- Dates within valid business ranges?
- Line items actually sum to total?
- All required fields present?
LAYER 2: Business Rule Validation
- Amount within vendor's 6-month history range?
- Payment terms match vendor profile?
- Missing PO number for amounts over $5,000?
- Unusual discounts or adjustments flagged?
LAYER 3: Confidence-Based Routing
- High confidence (>95%) + Low value (<$5,000) = Auto-process
- Medium confidence (85-95%) OR Medium value ($5-20K) = Review queue
- Low confidence (<85%) OR High value (>$20K) = Always human approval
RESULTS AFTER IMPLEMENTING LAYERS:
- Zero high-value errors: 8 months and counting
- Human review required: Only 8% of documents (was 100%)
- Processing speed: Still 82% faster than full manual
- Errors caught before processing: 23 incidents
- Estimated damage prevented: $187,000
THE LESSON:
Accuracy percentage means nothing without context.
What matters: What happens when it fails? Does it fail safely?
Always assume the system will fail. Route intelligently. Save the expensive mistakes.
What validation layers protect your high-value workflows from catastrophic errors?