Researchers have built a system that turns dense enterprise guideline documents into structured, reviewable rules in under two hours instead of days.
The system, called SAGE, is a multi-stage pipeline that extracts rules from documents mixing narrative text, tables, and images, then checks them for consistency, duplicates, and contradictions before routing only uncertain items to a human reviewer. Everything else gets auto-approved. Tested on 120 documents, the pipeline cut turnaround from two to three days down to 20 to 100 minutes, pulling 3,896 rules and producing 812 artifacts ready for review. Researchers report a 96% document-level success rate and a 3.2% hallucination rate; that figure jumps to 15.7% when the governance layer is switched off.
That governance layer is the actual story here, not the speed. Plenty of tools can extract text from a document; the harder problem is knowing which extracted rule to trust, and SAGE's answer is a versioned rule store with provenance tracking rather than a bigger model. For compliance teams currently paying humans to manually convert policy documents into checklists, a nearly fivefold drop in hallucination matters more than the time savings.
It's a research paper, not a shipped product, so the real test is whether 'auto-approved' holds up outside a 120-document benchmark.