Document intelligence demos look great: ask a question, contract clauses and SOPs appear. After launch, incidents look great too—models cite last year's voided drawings, terminated contracts, another division's pricing. Root cause: ingest too eager, classification too lazy. Search does not know "valid"—only "similar." Similar turns expired knowledge into current answers.
Classify contracts, drawings, and SOPs before they enter retrieval. An unclassified knowledge base is a talking archive accident.
How Dangerous Unclassified Retrieval Is
Commercial terms answered from expired versions—legal takes blame. Process answered from void SOPs—quality takes blame. Classified files mixed into company-wide Q&A—compliance shuts the assistant down. The smarter it sounds, the more authoritative—and one wrong authority and nobody dares use it again.
Folder names are not classification. People mis-file. Classification belongs in metadata and follows permissions.
Ingest Gates
- Every document requires: type, security level, applicable org, effective/expiry dates, owner.
- Expired and void documents auto-leave search; archive only—Q&A must not cite them.
- Contracts and drawings anchor on version; answers must show version number—no paragraph-only replies.
- Above the asker's clearance: refuse outright—do not summarize and leak.
The XYN digital intelligence system can make archive metadata and permissions a configurable knowledge entry. Document intelligence starts with classification. Classified retrieval is an assistant; unclassified retrieval is a faster rumor mill.
