Databricks Genie Ontology became generally available in November 2026, and it starts by reading what your account already holds: certified and uncertified tables, dashboards nobody maintains, SQL in query history, Genie Agents, metric views, and the grants on all of them. The ontology ranks context with signals such as certification, popularity and recency, so a Unity Catalog with three competing revenue tables and broad read grants gives agents three competing answers, each with an authority score attached. Before agents use it, fix the small set of objects the ontology treats as authoritative and limit what it learns from. Below: what shipped, what is still Beta, and what to clean up so Databricks Genie Ontology starts from a governed base.
What Databricks Genie Ontology Is: Learned and Modeled Knowledge
The Genie Ontology documentation calls it "the unified context layer that gives AI agents a business-aware map of your organization." It holds two kinds of knowledge.
Learned knowledge is a set of snippets that Genie extracts from metric views, dashboards, SQL queries and Genie Agents, including queries that external tools such as dbt run on Databricks. The GA announcement calls these units Ontology Snippets: calculations, business rules and data relationships that refresh as the source assets change. You cannot edit a snippet directly; you change its source asset or model the concept explicitly.
Modeled knowledge is what you define through Unity Catalog semantics: metric views, domains, Pages and certification. According to the docs, Genie treats modeled knowledge as more authoritative, so it takes precedence when the two conflict.
Between the two sits OntoRank, which the announcement describes as an "authority-aware ranking algorithm." It weighs source authority, certification, the expertise of the people behind the knowledge, popularity and recency. Read that list as a to-do list: every signal comes from your catalog. Databricks Genie Ontology does not reconcile inconsistent definitions for you. It ranks them.
Agents outside Databricks reach the same context through the Genie One MCP server, under the same Unity Catalog permissions. That makes the ontology a candidate for a shared enterprise ontology for AI agents, not a feature of one chat window.
What Is GA, What Is Beta, and What Is Only Announced
Databricks published a blog post, a press release, release notes and a what's-coming page, and they do not all agree. The table follows the November 2026 release notes where an entry exists and names the source otherwise.
| Capability | Status | Source |
|---|---|---|
| Genie Ontology | GA | Release notes, November 9, 2026 |
| Ontology Snippets, organizational directories (Workday) | GA | GA blog |
| Domains and subdomains | GA | Release notes, October 8, 2026 |
| Pages | GA | Release notes, November 11, 2026 |
| Genie Search | GA in blog and release notes; "coming soon" in press release | See below |
| External knowledge sources (Power BI) | Beta; Tableau and Sigma "coming soon" | Release notes, GA blog |
| Ontology Manager | "coming soon in Beta" | GA blog |
| Always-on Genie | Coming to preview and beta programs | Press release |
| Unity Gateway | Scheduled for general availability "soon" | What's-coming page |
Genie Search is the one real conflict. The blog and the release notes (entry dated November 9, 2026) call it generally available (see the setup page). The press release says "Genie Search, Ontology Manager and the new always-on Genie are coming soon to preview and beta programs." We read the release notes as the operational source, but the vendor's materials disagree, and the notes warn that an account "might not be updated until a week or more after the initial release date." Check your account console.
One date needs action now. The what's-coming page says the Genie One MCP Beta endpoint, /api/2.0/mcp/genie, reaches end of life on November 30, 2026. Clients must move to system.ai.genie_one_mcp in Unity Gateway and request the ai-gateway OAuth scope instead of genie. If your agents already call Genie over MCP, that switch is the first Databricks Genie Ontology task with a deadline.
What to Fix in Unity Catalog Before Genie Reads It
This is not a guide to building metrics from scratch; for that, read our walkthrough of a semantic layer built on metric views. The question here is narrower: which parts of an existing catalog does Databricks Genie Ontology treat as authority, and are they ready.
Unity Catalog metric views for KPIs that must be exact
The ontology docs say to "define your critical KPIs as metric views," and the GA blog describes modeled knowledge as "the comparatively small set of definitions that must be exact." That means a short list: revenue, margin, active customer. If two Unity Catalog metric views define the same KPI differently, retire one before launch. Metric views also accept agent metadata such as synonyms and display names, which is where business vocabulary belongs.
Certification and deprecation
Certification is a governed system tag, system.certification_status, with two values: certified and deprecated. The certification docs say it steers Genie One "toward the assets your organization vouches for." Applying it requires ASSIGN on the governed tag plus APPLY TAG on the object. The same page describes an automation, in Beta, that deprecates tables with no reads or writes over a set period. Run something like it before launch: usage and recency feed the ranking, and an old table that a forgotten job still queries can keep looking relevant.
Domains on governed tags
Domains are account-level and built on governed tags. Subdomains use a Finance/Tax naming pattern and nest up to five levels, and tagging an asset Finance/Tax does not tag it Finance. Genie uses the domain to choose between definitions when one term means different things to different teams, so an improvised tag taxonomy shows up in answers. The docs also warn that tag and domain metadata replicate globally, so keep sensitive terms out of them.
Pages and who may publish them
Pages are governed definitions of business terms, each with an owner and a home in one domain, and Genie prefers a Page over learned context. Check one default before anyone writes them: according to the Pages documentation, MANAGE PAGES "is granted to All account users on every domain." Revoke it per domain and grant it to the people who own the definitions. Pages do not support customer-managed keys either, which matters for regulated teams.
Permissions as the filter on what Genie learns
Snippets inherit the permissions of their source asset, and the ontology docs name the lever: "To keep stale or experimental assets out of answers, restrict access to them with Unity Catalog permissions." Sandboxes and scratch schemas that everyone can read become context. The exception runs the other way. Snippets from external knowledge sources "are available to all users in your account," so turning on the Power BI connector is an account-wide decision, not a team experiment.
Governing Agent Access Through Unity Gateway
Unity Catalog governs the knowledge, and Unity Gateway governs what agents do with it. The gateway docs list the controls: AI services registered as Unity Catalog securables, grants on MCP services, service policies as guardrails, rate limits, and request and response logging to Unity Catalog Delta tables for audit.
For Genie, check three settings. Access: calling the Genie One MCP server requires EXECUTE on system.ai.genie_one_mcp plus USE CATALOG on system and USE SCHEMA on system.ai, per the Genie One MCP docs, so grant it to groups for known agents. Budgets: Unity Gateway budgets cover Genie, and since November 12, 2026 per-user thresholds can be daily as well as monthly, with alert or block actions. Approvals: the press release says Genie actions stay within "Unity Gateway approval rules and budgets," but we did not find approval rules described in the gateway pages we checked, so treat that part as announced.
On pricing, Databricks says customers "pay for what they use, not per seat," with spending limits by team or user set through Unity Gateway. The ontology docs add that curating snippets carries no charge at this time and that Databricks may change this. If several agents will share this layer, our agentic AI architecture work starts from the same split: governed context in the catalog, runtime limits at the gateway.
A 30-Day Readiness Checklist for Databricks Genie Ontology
This is the order we use. The day ranges are a planning frame, not a promise.
- Days 1 to 5, inventory: list the KPIs leadership uses, the tables and dashboards behind each, and every MCP client that calls Genie. Move those clients off the Beta endpoint before November 30.
- Days 6 to 10, permissions: remove broad read grants on sandbox, staging and personal schemas, and check what All account users can see.
- Days 11 to 15, certification: deprecate duplicates and unused tables, then certify one source per KPI, including the dashboards and Genie Agents you keep.
- Days 16 to 20, domains: agree on the top-level taxonomy with business owners, create the governed tags, and assign curators with MANAGE DOMAIN.
- Days 21 to 25, modeled knowledge: put the exact KPIs in metric views and write Pages for the terms that cause arguments, with named Page managers only.
- Days 26 to 30, test and govern: ask Genie the questions people really ask, read the cited snippets, fix sources rather than answers, and set budgets before a wider rollout.
The day-30 check is simple. For each critical KPI, Genie should cite a certified metric view or a Page, not a learned snippet from someone's notebook.
When It Is Not Worth Doing Yet
Some teams should wait. If most tables still sit in the legacy Hive metastore, the ontology has little to read, because snippets follow Unity Catalog permissions and modeled knowledge lives in Unity Catalog semantics; the catalog move comes first. If nobody owns definitions, Pages and certification have no maintainers, and the ontology turns into another stale wiki. If compliance requires customer-managed keys, note that snippets and Pages do not support them. And if your agents answer narrow questions over a few curated tables, one well-built Genie Agent may be enough. Databricks Genie Ontology pays off when many agents need one shared reading of the business.
About the headline number: the GA blog reports that Genie with the ontology passed 91% of enterprise task runs, against 45% and 58% for two frontier agents. That is the vendor's own benchmark, built with Turing on a simulated workspace for a fictional energy supplier, and the blog says the results "reflect an established deployment where teams have accumulated that knowledge." Treat it as a claim to reproduce on your own questions.
The groundwork is ownership and permissions, not features. For help, see our Databricks platform engagements, or send us your critical KPIs and the tables behind them.
FAQ
Is Databricks Genie Ontology generally available?
Yes. The release notes list Databricks Genie Ontology as generally available on November 9, 2026, and the announcement followed on November 11. External knowledge sources are in Beta; Ontology Manager and always-on Genie are announced for preview or Beta.
Is Genie Search GA or still in preview?
The sources disagree. The release notes and the GA blog call Genie Search generally available, while the press release from the same week places it among features "coming soon to preview and beta programs." Check your account console before planning around it.
Do I need metric views before turning the ontology on?
No. The docs say Genie works from learned knowledge as soon as you start and that answer quality improves as you add modeled knowledge. Model the few KPIs that must be exact first.
How is Databricks Genie Ontology priced?
Databricks describes Genie pricing as usage-based, "not per seat," with limits set per team or user through Unity Gateway budgets. The ontology docs say curating snippets currently carries no charge and that pricing may change.
