What is a data ontology and how does it help AI?

Summary

• A data ontology bridges the gap between how business stakeholders think about their domain (in concepts like customer, order, revenue) and how data is physically structured in tables and schemas.
• Genie Ontology, Databricks' implementation, automatically extracts knowledge from existing assets (dashboards, queries, tables, notebooks, pipelines) and organizes it into a living graph of how a company works and what its data actually means.
• Ontologies enable AI systems to ground reasoning in business context, dramatically improving accuracy and preventing hallucinations by providing explicit definitions instead of guesses from raw data.
• The OntoRank algorithm ranks conflicting definitions by authority, source credibility, usage patterns, and freshness, so AI agents always use the most trustworthy source when multiple definitions exist.
• Permission-aware context delivery ensures AI agents only access definitions from assets a user has permission to see, preventing information leakage while enabling secure cross-organizational reasoning.

What is a data ontology and how does it help AI?
A data ontology is a structured representation of business concepts, their relationships, and their meanings within an organization. For any company, there's a gap between how business stakeholders think about their world and how that world is represented in data. A business user thinks in terms of "customer", "order", "revenue", and "churn". The data system thinks in tables, primary keys, foreign keys, and join conditions. An ontology bridges that gap by capturing the conceptual layer (what these terms mean, how they relate, which definitions are authoritative) and making that knowledge queryable and actionable for AI systems.

Why Ontologies Matter for AI

AI systems are only as good as the context they're given. When an AI agent receives a raw SQL query or a database schema, it has no understanding of business semantics. If the schema has a table called users with a status column, the agent doesn't know if status values are "active", "churned", "suspended", or something else. It doesn't know whether "revenue" in a column labeled rev_usd is net or gross, or whether it includes refunds. Without this context, AI systems make guesses, and guesses compound into hallucinations and incorrect answers.
An ontology provides explicit business context so AI systems can reason accurately. Instead of guessing, an agent searches the ontology, finds that "revenue" is defined as "gross transaction value excluding refunds" and is located in the order_summary.gross_rev column with source authority from the Chief Financial Officer, and answers with confidence grounded in that authority.

Genie Ontology: Modeled and Inferred Context

Genie Ontology, Databricks' implementation, follows a "model the head and let Genie infer the tail" philosophy. Organizations deliberately define their most critical business concepts in Unity Catalog through metric views, semantic definitions, and the Unity Catalog Business Glossary (the modeled layer). Genie then extends that foundation by automatically inferring broader context from existing assets (dashboards, SQL queries, notebooks, and usage patterns) to build a richer understanding without requiring manual curation of every definition.
This hybrid approach scales. Teams curate what matters most; Genie learns the rest.

OntoRank: Authority-Based Ranking

Business organizations often have conflicting or overlapping definitions. Finance defines "customer" one way (accounting basis). Sales defines it another (prospect, customer, renewal, at-risk). Product uses yet another definition (users who logged in this month). Without a way to rank definitions, AI agents face ambiguity and return inconsistent results.
Genie Ontology solves this with OntoRank, an authority-ranking algorithm similar to PageRank. OntoRank weighs:

Factor	Example
Source authority	Definition from Finance VP ranked higher than definition from a sales rep
Usage prevalence	Definitions that appear in certified metric views and dashboards rank higher
Asset tier	Definitions tied to gold-tier, certified assets outrank inferred definitions
Recency	Freshly updated definitions rank higher than stale ones
Organizational domain	Finance domain's "revenue" is authoritative for Finance questions; Sales domain's "prospect" is authoritative for Sales

Using these signals, Genie Ontology ensures that when an AI agent searches for a definition, it consistently returns the highest-authority source. This makes AI answers predictable and trustworthy.

How Ontologies Improve AI Accuracy

Ontology grounding significantly improves how well AI systems answer financial and operational questions on the first attempt. The difference is context. When a VP asks "Why did we miss forecast on that account?", Genie using ontology understands what "account" means in the CRM system, what "forecast" means in the revenue model, and where these concepts map to tables. Without ontology, a general agent searches for clues in raw schemas and often returns incorrect joins or missing filters.

Permission-Aware Context

Ontologies are fully permission-aware. Genie Ontology only extracts and surfaces context from assets the requesting user has permission to access. If a user cannot see the Finance schema, definitions sourced from Finance data are not available to agents acting on that user's behalf. This prevents information leakage while still enabling broad AI reasoning. Users see consistent, grounded AI answers from contexts they're authorized to explore.

Token Efficiency and Cost

By injecting only relevant, ranked context into AI reasoning loops, ontologies reduce the amount of data agents need to crawl and process. Genie answers faster and more cost-efficiently because it has better context upfront, generates more efficient queries, and avoids exploratory data work that wastes tokens. In high-volume agent deployments, this efficiency compounds.

Unified Context Across All Surfaces

Genie Ontology serves as a single context layer for all Genie experiences: Genie One (natural language chat), Genie Agents (autonomous agents), and Genie Code (AI coding). Context curated once applies across all surfaces, preventing the terminology drift and inconsistency that occurs when each agent or tool maintains its own definitions.

How to Build an Ontology

Start with your top business metrics. Which metrics do executives ask about most often? Start there. Define them in Unity Catalog as metric views with clear, stable definitions.
Document domain-specific terms. Use the Unity Catalog Business Glossary to document key concepts: Customer, Revenue, Order, Churn, and so on. Include authoritative sources and usage context.
Let Genie infer the rest. Once your core concepts are defined, Genie Ontology automatically extracts and ranks related definitions from existing assets, building out a rich context graph without manual effort.
Evolve based on usage. Monitor which definitions AI agents rely on most. If an inferred definition is frequently used, promote it to a metric view or glossary entry so it becomes part of the curated core.