To understand how to answer world questions from data models, we don't need semantic layers, we need ontologies

dltHub

Modern data stacks suffer from theoretical blindness by over-focusing on dimensional models and semantic layers, losing the ability to represent real-world systems. The core argument is that data models are lossy compressions of reality, and when LLMs operate on them without an ontology, they hallucinate missing context and apply generic common sense instead of business-specific logic. An ontology—defining entities, relationships, and attributes as a 'blueprint of truth'—bridges the gap between raw data and real-world meaning. A practical A/B test scenario illustrates how an LLM with only a data model catastrophically misreads a 'dashboard win' (falling support tickets) as success, while an LLM equipped with a business ontology correctly identifies it as enterprise client disengagement. The conclusion: semantic layers make LLMs better reporters; ontologies make them strategic analysts capable of understanding causality.

Ontology driven Dimensional Modeling

The Map (ontology) vs. The Territory (data model) Link icon

Implications: Upgrading the AI from Reporter to Strategist Link icon

The Tale of the Two Analysts and Experimental Outcomes Link icon