The challenge
A legal-tech SaaS processing contracts wanted to cut manual review time by 70% using LLM-assisted extraction and summarisation, while maintaining zero data-retention compliance and a human-in-the-loop approval flow.
Our approach
- LLM architecture review comparing OpenAI, Anthropic and fine-tuned open-source options for the workload.
- RAG pipeline with embeddings indexed on customer-owned vector store (no third-party data retention).
- Prompt library with versioning, A/B testing and cost monitoring dashboards.
- Human-in-the-loop approval UI with inline citations from source documents.
- SOC 2-aligned audit logs, role-based access and customer-side data isolation.
The outcome
- Contract review time dropped from 42 minutes to 11 minutes average.
- Extraction accuracy exceeded 96% against the legacy manual process baseline.
- LLM cost per review optimised to USD 0.21 via model routing and prompt caching.
- Enterprise customer onboarding timelines compressed by 40% due to faster reviews.
Technology stack
Anthropic ClaudeOpenAINext.jsPostgres + pgvectorLangGraphVercel
