Geek Out Time: Why RAG Alone Fails Enterprise Governance (And How Ontology Fixes It)

In the earlier geek out ( https://medium.com/the-constellar-digital-technology-blog/geek-out-why-enterprise-ai-quietly-depends-on-ontology-485bdecc733c ), I focused on the separation of responsibilities: letting LLMs deal with language, while ontology enforces rules. That experiment assumed something important but often unspoken that the facts being checked were already correct. Now, I m going to remove that safety net.

Instead of asking whether an ontology can enforce a rule, I asked a more uncomfortable question: what happens when the model is confidently shown the wrong information?

If you’ve worked in an enterprise long enough, you already know the backdrop. Policies change. Old documents don’t disappear. FAQs, quick references, email threads and meeting notes all survive long after the official rule has moved on. The interesting question is how an AI system behaves inside that mess.

I kept the same approval rule as before: Finance Managers can approve invoices up to $50,000, CFOs have no limit, everyone else has none. What changed completely was the retrieval layer.

This is an excerpt — the full article continues on Medium.