How do I evaluate ontology platforms for AI?

Summary

• The real test is not how elegant the graph looks. Judge an ontology platform on whether it turns your business vocabulary into trustworthy, permission-aware context an AI agent can actually use.
• Insist on authority ranking. When two teams define "revenue" differently, the platform should rank definitions by source authority, usage, and freshness rather than treating every definition as equal.
• Prefer semantics governed in the same system as your data, with one permission model, lineage, and audit, over a separate tool that drifts from how data is really used.
• You should be able to start with one domain and a few critical metrics and expand from there, not model the entire enterprise before you get any value.
• On Databricks, Unity Catalog metric views, the business glossary, and domains give you a governed foundation today, and Genie Ontology (in gated Public Preview) infers and ranks the surrounding context with its OntoRank authority score.

How do I evaluate ontology platforms for AI?

Evaluate an ontology platform on one question: does it give AI a trustworthy, permission-aware understanding of what your data means, or just a prettier diagram of your tables? The criteria that matter are business-concept modeling, how the platform resolves conflicting definitions by authority, whether context respects existing access controls, how it stays fresh as usage changes, how tightly it integrates with the system that governs your data, and whether you can adopt it one domain at a time. A demo on clean sample data proves little. What matters is whether an AI agent, asked a real question, retrieves the authoritative definition and the right tables, and never surfaces context the user isn't allowed to see.

What an AI ontology actually does

There's a gap between how people talk about the business (customer, order, revenue, churn) and how data is physically stored (tables, keys, join conditions). An ontology closes that gap. It captures the concepts, the relationships between them, and which definition is authoritative, then makes that knowledge available to AI systems. Without it, an agent handed a raw schema guesses: it can't tell whether a status column means active, churned, or suspended, or whether rev_usd is gross or net of refunds. Those guesses compound into confident, wrong answers. An ontology replaces guessing with explicit, ranked business meaning.

Key capabilities to evaluate

Criterion	What to look for
Business-concept modeling	Maps business terms to physical tables and the relationships between them, not just a catalog of column names.
Authority and conflict resolution	Ranks competing definitions by source authority, usage, and freshness, so agents use the most trustworthy one.
Permission-aware context	Serves only context the requesting user is allowed to see, so definitions never leak across access boundaries.
Freshness and continuous learning	Updates automatically as teams query, build, and certify assets, instead of decaying as a static document.
Platform integration	Governs semantics in the same system as the data, with one permission model, lineage, and audit.
Incremental adoption	Lets you start with one domain and a few critical metrics, then expand, rather than modeling everything up front.
Evaluation and governance	Builds in versioning, monitoring, and answer-quality evaluation so you can trust and improve results over time.

How Databricks approaches AI ontology

Databricks splits the problem into a governed foundation you control and an inferred layer that scales it. You curate the concepts that matter most as Unity Catalog metric views and business-glossary terms, organized into domains. Genie Ontology then follows a "model the head, infer the tail" approach: it automatically extracts knowledge such as metric definitions, authoritative sources, and business rules from tables, queries, dashboards, notebooks, and connected apps, and organizes it into a living context layer.
When definitions conflict, Genie Ontology ranks them with OntoRank, an authority score in the spirit of PageRank that weighs where a definition came from, how authoritative its source is, how widely it's relied on, and how fresh it is. Context is permission-aware through Unity Catalog, so an agent only ever sees definitions sourced from assets the user can access. That one context layer serves Genie One, Genie Agents, and other tools, so terminology stays consistent instead of each agent inventing its own.
One practical note on timing: Unity Catalog metric views, the business glossary, and domains are available today, and metric views ship with generally available low-code authoring, while Genie Ontology is in gated Public Preview. The durable move is to invest in that governed foundation now, since the ontology builds directly on it. See what is enterprise intelligence for how the pieces fit together.

FAQs

Is an ontology the same as a knowledge graph?

Not quite. A knowledge graph is one way to represent entities and their relationships. An ontology is the broader model of concepts, relationships, and authority. Genie Ontology deliberately uses a flexible, model-friendly representation rather than requiring you to hand-author formal graph languages, and you steer it through curated concepts and the Unity Catalog glossary.

Do I have to build the ontology by hand?

No. You define the critical "head," your most important metrics and terms, and the platform infers the rest from assets you already have. Hand-authoring every definition doesn't scale and tends to drift from how data is actually used.

How does it stop AI from mixing up conflicting definitions?

Through authority ranking. OntoRank scores each definition by source authority, usage, and freshness and returns the most authoritative one, so answers stay consistent even when Finance, Sales, and Product each define a term differently.

Is Genie Ontology generally available?

As of Data + AI Summit 2026 it is in gated Public Preview, not generally available. The foundation it relies on, Unity Catalog metric views, the business glossary, and domains, is available today, so teams can start building governed context immediately.