Data · AI · Law · Ethics

The magazine for decision-makers

·

Ontologies for AI: Databricks and Microsoft Fabric rediscover data semantics

Databricks with Genie Ontology and Microsoft with the Fabric IQ ontology are betting on the same concept: giving AI agents a shared map of the business. What a data ontology is and why it matters.

by

·

·

4–6 minutes
Svincolo autostradale visto dall'alto, con strade che si intrecciano e collegano quartieri diversi

Key points

  • What is happening: Databricks (Genie Ontology) and Microsoft (the Fabric IQ ontology) have both introduced, in preview, a semantic layer designed for AI agents.
  • Why: an agent that sees only tables and columns does not know what an “active customer” is. The ontology tells it, once and for all.
  • What to do: start with a few business concepts, with clear definitions and owners, before connecting them to agents.

For years, data ontology seemed like a topic for academics and semantic web projects. Today it is back in the spotlight for a very practical reason: AI agents need to understand the business, not just read databases. And the two major data platform vendors, Databricks and Microsoft, have chosen precisely this word for their new context layers.

What an ontology is, in simple terms

An ontology is a shared, machine-readable description of the concepts that matter to an organization: customers, products, orders, plants, stores. For each one, it defines the properties, the relationships with other concepts, the official metrics and the business rules. It then links these definitions to the actual data, wherever it resides.

The difference from a traditional data catalog is the point of view. The catalog describes how the data is structured (tables, columns, types). The ontology describes what it means for the business.

WHO USES THE DATA BI dashboards AI agents Applications Excel and sheets Ontology · shared semantic layer Entities Relationships Metrics Rules Permissions Lakehouse Warehouse Databases Events and streams BI models DATA SOURCES (they stay where they are) QrmServizi illustrative diagram. The ontology links definitions to data without copying it.
The ontology sits in the middle: a single business vocabulary for all consumers.

Why AI agents need it

Ask an agent “how many active customers do we have in Lombardy?”. Without context, it has to guess: which table contains customers, what “active” means, how a customer is linked to a region. Every agent, every prompt and every dashboard risks giving a different answer.

With an ontology, the definition exists only once and is governed. The agent does not invent the revenue formula: it uses the official metric. It does not infer the relationship between orders and stores: it finds it written down.

places contains takes place at Customer Order Product Store linked to: crm.customers linked to: sales.orders linked to: catalog.products Metric · Revenue sum of order amounts Rule · Active customer at least one order in the last 90 days
A fragment of an ontology: concepts, relationships, an official metric and a business rule (illustrative example).

Databricks: Genie Ontology

In the Databricks documentation, Genie Ontology, currently in Public Preview, is described as the unified context layer that gives Genie a map of the organization’s business. It combines two sources:

  • modeled and governed context in Unity Catalog semantics: metric views (metrics defined once and queried at runtime), domains, Pages to describe business concepts (in Beta since August 2026), and certification and deprecation of assets;
  • context inferred automatically by Genie from assets and how they are used.

The same ontology is consulted by Genie One, the interface for business users, and by Genie Code, the assistant for developers. At the end of October 2026, Databricks added relationships between metric views, in Beta: an important building block for moving from a list of metrics to a true network of concepts.

Microsoft: the Fabric IQ ontology

Microsoft presented Fabric IQ and its ontology item at Ignite in November 2025. Here too, it is in preview. The Fabric ontology defines entity types, properties, relationships, metrics (derived from the DAX measures of Power BI semantic models) and rules written in natural language. It supports namespaces and inheritance, and connects to lakehouses, warehouses, eventhouses and databases without copying their data.

Two interesting details for those who work with standards: the ontology can import RDF, Turtle and OWL definitions and export them to RDF or Turtle. And it respects the security of the connected data, including row-level and column-level filters. The old ontology experience will be retired on January 31, 2027.

Two paths to the same goal

Databricks · Genie OntologyMicrosoft · Fabric IQ ontology
StatusPublic PreviewPreview
FoundationUnity Catalog semantics (metric views, domains, Pages)Dedicated item in Fabric, can be generated from Power BI models
How context is createdModeled by people + inferred from usageModeled by people, with the help of an agent that proposes changes for approval
Semantic standardsNot stated in the documentation reviewedImport RDF, Turtle, OWL; export RDF and Turtle
Who uses itGenie One, Genie Code, agentsData agents, real-time dashboards, agents
Comparison based on the official documentation reviewed in November 2026. Both features are in preview and may change.

The point is not the technology

An ontology cannot be bought: it has to be built. The tool helps you write it and connect it to the data, but the hard part remains organizational. Someone has to decide what “active customer” means, who is responsible for the definition and what happens when two departments disagree.

  • Start small: ten key concepts are worth more than an encyclopedic model that is never finished.
  • Give every definition an owner: without ownership, the ontology ages quickly.
  • Reuse what already exists: BI semantic models and already certified metrics are the natural starting point.
  • Measure consistency: the same question, asked through different channels, must produce the same answer.

The compass

AI agents expose an age-old problem: companies often lack a shared vocabulary. A data ontology is the way to write it down, and those who do so first will have agents that are more reliable, not just faster.

Sources

Leave a Reply

Your email address will not be published. Required fields are marked *