Databricks Genie One Explained: The Data-Smart AI Coworker (and the Ontology It Runs On)

TL;DR

• What it is: Genie One is a "data-smart AI coworker," now generally available, that answers business questions, builds documents and reports, and triggers actions, in natural language, across web, Slack, Teams, and mobile.
• Genie Agents (GA): save any Genie conversation as a reusable agent that inherits its sources, instructions, and behaviour, and can reason over unstructured data, not just tables and views.
• Genie Ontology (Public Preview): an automatic context layer that extracts knowledge from tables, queries, dashboards, pipelines, and connected apps, and ranks data authority so Genie knows what to trust.
• Governed by default: permissions are enforced on every answer through source-native access controls or Unity Catalog. Genie only surfaces what the user is allowed to see.
• The catch: the Ontology is automatic, but it can only map meaning that exists. Undefined metrics, ungoverned tables, and no semantic layer give you a confident coworker built on sand.

What Genie One actually is

Genie One is Databricks' answer to the "AI coworker" category, and the framing it uses is "the data-smart AI coworker for every team." The important word is data-smart. A general AI assistant guesses from whatever it can scrape together. Genie One is designed to answer from your actual, governed enterprise data, with permissions enforced.

In practice that means an employee can ask a business question in plain language, and Genie does not just return a number, it produces the document, the report, or the artifact. It can schedule alerts and monitoring, and it can trigger actions across the business tools you connect to it. Databricks reports connectivity to more than fifty popular apps and data systems, including Google Drive, Jira, Slack, Confluence, and SharePoint, and Genie is reachable from the web, Slack, Microsoft Teams, iOS, and Android.

This is genuinely useful, and the accuracy claim is not marketing hand-waving. On a 28-question real-world data-analysis suite, Databricks reports Genie answered 84.5% of questions correctly on the first attempt while the strongest general-purpose coding agent managed 52.4%, and did it roughly twice as fast. We would still tell any client to benchmark it on their own questions before betting a workflow on it, but the direction is clear.

Genie Agents: the reusable layer

The piece that turns a clever chat box into something operational is Genie Agents, and these are now generally available.

A Genie Agent is what you get when you save a Genie conversation as a reusable agent. It inherits that conversation's memory, sources, instructions, and behaviour, so a question you worked out once becomes a repeatable, shareable capability rather than a one-off. Databricks describes these as curated, domain-specific agents that can take autonomous action and, importantly, reason over unstructured data, not just tables and views. You can spin one up from a single prompt.

For an enterprise, this is the difference between "someone in finance asked the AI a good question once" and "finance has an agent that answers that class of question consistently, for everyone, with the same governed sources behind it."

The Genie Ontology: the part that decides whether any of this works

This is the piece we would pay the most attention to, and it is the one most likely to be misread, because it is in Public Preview, not GA.

The Genie Ontology is an automatic context layer. Databricks describes it as automatically extracting snippets of knowledge from your tables, queries, dashboards, pipelines, and connected apps, and building a self-improving map of how your organisation defines and uses its data. Concretely, the verified description says it captures:

• Metric definitions and the business terms and unique calculations behind them.
• Relationships between concepts, metrics, tables, and teams.
• Data authority, computed with a PageRank-like weighting that considers where a definition comes from, the author's authority, how often an asset is used, whether it is tied to certified assets, and how fresh it is.

The point of all this is that Genie gets context about where to look, what to trust, and how to answer in a way that reflects how the company actually uses its data. That is the difference between an assistant that picks the first table it finds and one that picks the table your organisation has agreed is the source of truth.

Notice what this depends on. The Ontology can only map meaning that already exists somewhere. If your metrics are undefined, your tables are ungoverned, and there is no agreed semantic layer, the Ontology will faithfully map the mess. This is exactly the work we covered in our piece on the Unity Catalog semantic layer, the glossary, domains, and governance hub that give the Ontology something real to learn from.

Governed by default

The reason a tool like this can be let near a business at all is the governance model, and Databricks built it in rather than bolting it on.

Permissions are enforced by default on every answer, through source-native access controls or Unity Catalog. The Ontology only shows a user content they already have permission to see, so Genie cannot become a side door around your access policy. Governance of the models, tools, and costs behind Genie runs through Unity AI Gateway, which we broke down in our Unity AI Gateway deep dive. And because Genie Agents are agents, everything we said about running production agents in our Agent Bricks breakdown, evaluation, monitoring, budgets, applies here too.

The catch, and where it gets real for you

Genie One is one of the more impressive things on the Summit floor, and it is also the clearest illustration of the whole Summit 2026 message: the platform is ready, the question is whether your foundation is.

Two honest cautions.

First, the Ontology is automatic, not magic, and it is in Public Preview. Automatic extraction is a real advance, but it maps the meaning your organisation has actually encoded. If two teams define "active customer" three different ways and none of them is certified, Genie does not resolve the disagreement for you, it inherits it. The Ontology makes a good semantic layer pay off enormously; it does not create one.

Second, a coworker that can trigger actions raises the stakes on governance. An assistant that returns a wrong number wastes an hour. An agent that takes an action on a wrong number is a different kind of problem. That is precisely why the permission enforcement, the Gateway governance, and a real evaluation practice matter before you let Genie Agents act, not after.

The European angle

For regulated European enterprises, the attraction of Genie is that it answers from governed data with access controls enforced, rather than from an ungoverned copy sitting in a separate tool. That is the right shape. The work to make it approvable by a risk team is the foundation work: a governed Unity Catalog, a defined semantic layer the Ontology can learn from, identity and access that hold up to audit, and a clear view of what any action-taking agent is allowed to do. Get that right and Genie is a genuine productivity unlock. Skip it and you have given every employee a fast, confident path to the wrong answer.

How to get "Genie-ready"

Adopting Genie well is mostly foundation work, not Genie configuration.

• Define your metrics and business terms. The Ontology ranks data authority, so give it certified, agreed definitions to rank. This is glossary and domains work in Unity Catalog.
• Govern the catalog first. Permissions are enforced through Unity Catalog or source-native controls, so the quality of your access model is the quality of Genie's guardrails.
• Decide what "good" looks like. Benchmark Genie on your own real questions, not a generic suite, and define which answers and actions are acceptable before rollout.
• Treat action-taking as a security decision. For any Genie Agent that can trigger actions, review the tools it can call and the budgets and policies around it, the same way you would any production agent.
• Start where the semantic layer is strongest. Roll Genie out first in the domain where your definitions are cleanest, and expand as the foundation improves.

The Cosmos Thrace perspective

This is the layer we spend our time on, and Genie makes the case for it better than we ever could. We are a Databricks Silver Partner, and the unglamorous work, defining metrics, governing the catalog, building a semantic layer the business actually agrees on, is exactly what decides whether a tool like Genie is a productivity unlock or a confident liability. We have delivered dozens of data platform implementations across Europe, many on Databricks, with more than $50M saved for clients in 2025, a 100% client retention rate, and 106 million data points moved daily.

Our honest read on Genie One: the accuracy is real, the GA of Genie One and Genie Agents is significant, and the Ontology is the most strategically interesting piece, precisely because it makes your semantic and governance investment compound. But it is Public Preview, and it maps the meaning you have already created. The enterprises that win with Genie will be the ones who defined their metrics, governed their catalog, and treated action-taking as a security decision. Do that and Genie is a coworker worth having. Skip it and it is a very articulate way to be wrong at scale.

Sources

Databricks blog: Introducing Genie One, Genie Ontology, and Genie Agents

Databricks newsroom: Databricks launches Genie One, an all-new agentic coworker for every team