Mike's team began by mapping the contract value chain to identify where human data entry introduced errors and where risk queries consumed the most analyst time. The architecture challenge was significant: reinsurance contracts contained both precisely structured data — dollar amounts, dates, policy regions — and highly subjective legal language around inclusions, exclusions, and cyber clauses that resisted standard pattern matching. To handle both query types, the team designed a hybrid retrieval architecture combining SQL-style structured queries for precise data extraction with semantic document store retrieval for open-ended risk questions. Semantic chunking was tuned to preserve legal clause boundaries. A custom LLM dictionary of reinsurance-specific legal terminology was built to ensure the model interpreted domain language consistently and accurately. QA was conducted iteratively — 10 contracts first, then 100, then the full 10,000 — allowing the team to identify edge cases and tune accuracy before scaling. The system reached 97%+ accuracy at production scale, using make.com and relevance.ai alongside the core RAG layer.