Some VP, in a meeting near you, is asking: "We already have a data warehouse, a semantic layer, and a knowledge graph. Do we really need an ontology?" This 30-minute, demo-heavy talk answers yes — and shows it under a closed-loop LLM constraint that prevents hand-waving. We walk the four-layer enterprise data stack : Lakehouse, Semantic Layer, Knowledge Graph, Ontology, and run one live LLM demo per layer against a small, fully readable fictional dataset (NorthWind Outdoor: ~9 customers, ~21 orders). Each demo poses one question that the layer in question cannot answer cleanly, and the audience watches the next layer turn out to be exactly what removes the gap: 1. Lakehouse + revenue. The LLM lands on the right number, but only by stacking three labeled guesses (what column is money, what "revenue" is, whether to include returns). Motivates the semantic layer. 2. Semantic + co-purchase. The LLM hand-rolls a multi-CTE self-join. Works, but the LLM is the query author. Motivates the knowledge graph. 3. KG + churn risk. Every fact is on every node, but no class for ChurnRiskCustomer. The LLM invents a rule or declines. Motivates the ontology. 4. Ontology + VIP promo email. The LLM cites nw:VIPCustomer, nw:RestrictedMarketingRegion, and a SHACL Forbid rule by name; returns an auditable answer with citations and exemptions explained. You leave with a portable narrative you can run against your own data, a 4-cell rubric for evaluating whether your ontology layer is earning its keep, and a defensible 30-day starter plan.
Git Hub repository Containing Code samples and the deck.