Your Data Warehouse Is Not An Ontology
Why ten years of dbt models won't make your AI reliable — and what to build on top of them instead.
Abstract
Every enterprise we have studied that struggles with AI deployment has the same diagnosis: they confused the data warehouse for the operational layer. The warehouse is a passive substrate. The ontology is the active layer above it where AI actually lives — and where the real work of the next five years will be done.
The warehouse-as-ontology fallacy
When an enterprise decides to invest seriously in AI, the instinct is almost always the same: point the model at the warehouse. Years of investment in dbt, Snowflake, BigQuery, and well-modelled fact tables feels like the right substrate. It is not — and the pilots that result tend to stall in exactly the same place.
A data warehouse is a record of what has happened. An ontology is a working model of what can happen next. The two are related but not equivalent, and treating them as equivalent is the single most expensive mistake in enterprise AI.
What the warehouse gives you
- Clean tables, joins, lineage, and history — a normalised record of past state.
- A common query language across the organisation.
- Reasonable performance for analytical reads.
What the warehouse does not give you
- Object identity that survives across source systems and remains stable over time.
- The meaning of a relationship — what 'fulfilled by' actually entails in policy and process.
- The operational consequences of an action, including downstream effects and rollback paths.
- The permissions, constraints, and approval flows that govern who is allowed to do what.
What the ontology adds on top
The ontology layer takes the warehouse's rows and promotes them to first-class objects with identity. It takes the warehouse's joins and gives them named, directional meaning. And — most importantly — it adds a thing the warehouse has never had: a vocabulary of actions an AI agent can take with bounded, well-understood consequences.
Why this is the bottleneck for enterprise AI
Frontier models can read a warehouse. They cannot operate on one. The transition from 'AI that reads tables' to 'AI that runs the business' happens inside the ontology, not inside the warehouse — which is why every enterprise that has shipped useful agentic AI has built one, even if they called it something else.
The warehouse tells you what happened. The ontology lets the AI decide what to do next.
Hiring implication
Ontology engineers think differently from data engineers. They are less interested in pipelines and more interested in the operational meaning of a single object across the business. On the StarPlan marketplace they tend to filter to skills like agent orchestration, tool design, and evaluation — and to industries where they have already done the modelling work.