Ontologie

Tony Seale calls ontology the interface through which agents understand data: it defines what exists, its properties, and its relations. In "There is a growing disconnect in the way people think about building AI agents" (2026-04-17) he sets that against the model side of the same architecture. Frontier models are rented and anyone can build a harness, so both commoditize. What does not commoditize is the domain model: the structured, linked knowledge capturing how one organization understands the world. Seale's formula for the Semantic Agent is (Model + Harness) + (Ontology + Data); everything else he treats as scaffolding destined to fall away.

The earlier "Philosophy Eats AI" (2025-05-30) supplies the reasoning behind that claim. Seale, drawing on Michael Schrage and David Kiron in MIT Sloan, argues most organizations never formalize how they create value into a machine-readable structure, so their models learn from noise rather than meaningful patterns. His ontological core emerges from the use cases producing real value and from competency questions that test each concept.

McKinsey's Padhi and co-authors (2026-10-01) put ontologies and knowledge graphs in the semantic layer of their AI operating system, and recommend fixing semantic governance as an early step. Rakesh Gohel's Semantic Spine (2026-10-25) assigns the ontology the first of five links: it says what things are, before the knowledge graph says what is true, the context graph what matters now, and the procedural graph what to do next.