The challenge
What was at stake
A global media rights firm managing 28,000+ music acquisition agreements was losing an estimated $8.2M annually to undetected clause mismatches and manual extraction errors.
A legal team of 28 attorneys spent 6.5 hours per contract file on OCR cleanup. High-value contracts averaged 55 pages each, with complex royalty matrices and multi-party split schedules that standard cloud AI tools misclassified at a 15% error rate—each percentage point of miscalculation represented $420,000 in misallocated royalties.
The approach
What we built
We built a 3-stage AI-Human-in-the-Loop (HITL) contract intelligence platform. 9% accuracy.
Stage 2 ran a multi-agent semantic diff engine that compared amendments against a master rights registry, flagging 22 categories of legal deviation. 5 hours to 18 minutes.
Inside the system
Key capabilities
Layout-Aware Deep OCR Pipeline
Trained on 2,100 real contract samples; extracts royalty matrices and multi-party tables with 99.9% field-level accuracy.
PII-Aware Metadata Masking
Automatically identifies and redacts 45 categories of sensitive data — artist legal names, banking details, and proprietary split ratios — before AI processing.
Multi-Agent Semantic Diff Engine
Runs 22 deviation checks against amendments, comparing territorial rights, exclusivity, and royalty splits against baseline in under 6 seconds per document.
Human-in-the-Loop Review Console
Routes extractions below 94% confidence to legal staff, reducing per-file review time from 6.5 hours to 18 minutes.
Under the hood
Technology
- Custom layout-aware OCR engine
- private LLM-based clause extractor
- Neo4j-backed rights registry
- React HITL console
- semantic diff agent cluster
The outcome
What changed
Processed 28,000 contracts in the first 8 months. Per-file review time reduced from 6.5 hours to 18 minutes (95% reduction). Royalty extraction error rate dropped from 15% to 0.01%. $7.4M in misallocated royalties recovered.
- 99.9%
- Extraction accuracy
- 18 min
- Review time
- $7.4M
- Royalties recovered