Databricks Genie becomes an autonomous AI coworker for enterprise work

Databricks unveiled on November 11, 2026, during its Data + AI Live event, the new Databricks Genie: an autonomous AI coworker for continuous, long-running enterprise work. The presentation featured Ali Ghodsi, Databricks CEO and co-founder, and Greg Brockman, OpenAI co-founder and president.

Unlike an assistant that only answers when asked, the new Genie takes responsibility for work in progress: it tracks changes, follows up on missing inputs, runs scheduled or event-triggered jobs, and manages several assignments at once. According to the company, agents can even delegate parts of the work to subagents. Users set the boundaries for what Genie can do on its own and what needs approval.

Examples cited by Databricks include monitoring insurance claims for fraud patterns, assembling the case file and routing it to underwriters when a segment moves outside guidelines, as well as managing marketing campaigns and proposing weekly budget changes for approval.

Genie Ontology and Genie Search reach general availability

At the center of the announcement is Genie Ontology, now generally available. It is the context layer that maps how a business works — definitions, metrics, relationships and organizational knowledge — from dashboards, SQL queries, notebooks and connected apps. The OntoRank system weighs source authority, certification, relevance, popularity and recency to resolve conflicting definitions.

Also generally available, Genie Search is a hybrid search spanning Google Drive, Jira, Confluence and Workday, with SharePoint in beta, respecting each source system's permissions. Genie Agents can now analyze PDFs and slide decks alongside tables, with page-level citations, in public preview, and native MCP connectors are generally available. OpenSharing and Genie Clean Rooms remain in beta.

Controls and benchmark

The autonomy comes with controls in Unity Gateway: audit traces record the question, answer, cited objects, time and actions; administrators can restrict MCP tools, require approvals, and set budgets and spending caps. Genie acts under each user's identity, and access approvals stay with administrators and data owners.

In a benchmark developed with Turing, Databricks says Databricks Genie answered 91% of task runs correctly, versus 58% for the strongest frontier agent evaluated, averaging 431,829 tokens and 100.3 seconds per run. The company itself cautions that the test reflects an established deployment with accumulated knowledge — results may be lower in new environments. The figures come from Databricks' own evaluation, not an independent measurement.

General Motors, according to Databricks, has deployed more than 1,000 Genie agents across sales, engineering, supply chain and HR. Genie One, the interface for this work, also creates slide decks, dashboards and interactive pages, and remains free for users through January 31, 2027.

Not to be confused: Databricks Genie is Databricks' enterprise assistant. Google DeepMind has another product with the same name, Genie 4, a world model for interactive simulations — a different topic, already covered by Geekiki.

Image credit: Genie Ontology architecture diagram, published on the official Databricks blog.

Fontes

Transparency: this article was produced with the assistance of artificial intelligence, based on public posts on X and sources from the internet, and reviewed against the original sources cited. We recommend consulting the original sources to confirm the details.

Muse