SAGE Uses a Governed LLM Pipeline to Turn Enterprise Guidelines into Work Artifacts

Summary

Enterprise guideline documents combine narrative text, complex tables, and embedded images, making conversion into structured work artifacts a process that can require two to three days of manual effort per document. The authors introduce SAGE, a governed multi-stage large language model pipeline designed to add controls beyond extraction. It uses a shared, versioned rule store with stable identifiers, schema-validated contracts between stages, and end-to-end provenance tracking. Extracted rules undergo deterministic structural validation and LLM-based semantic scoring. A consistency module removes duplicates, flags contradictions, and identifies gaps in the specification. High-confidence results can be auto-approved, while uncertain or flagged items are routed to human reviewers. In an evaluation covering 120 documents, SAGE reduced turnaround time to 20-100 minutes and achieved a 96% document-level success rate. It extracted 3,896 rules and generated 812 artifacts ready for human review. The reported hallucination rate was 3.2%, compared with 15.7% for a version without governance. The results describe a controlled document-to-artifact workflow, although the abstract does not establish performance beyond this evaluation setting.