Databricks launches Genie, an autonomous AI coworker for long-running enterprise work, answering 91% of benchmark task runs correctly
Synopsis
Databricks released a new version of Genie that turns it from an on-demand data assistant into an autonomous AI coworker that continuously carries long-running enterprise work: it tracks progress, follows up on missing inputs, and acts as new information arrives, while users set guardrails and approve key actions. The autonomy rests on Genie Ontology, which organizes business definitions, metrics, relationships and organizational knowledge into context agents can use. On a new benchmark built with Turing, in the same Databricks workspace, Genie powered by Genie Ontology answered correctly in 91% of task runs, 1.6 times the rate of the stronger frontier agent, while using less than half as many tokens.
Interpretation
Genie moves from an assistant that answers single requests to an autonomous AI coworker that stays on the job, tracking progress, following up on missing inputs and acting as new information arrives. Earlier AI assistants could run analyses and summarize meetings, but employees still managed the work around them, because most enterprise work is continuous, complex and data-intensive. Databricks says thousands of customers use Genie in production, including Zillow, Unilever, adidas, Atlassian and AstraZeneca.
The autonomy depends on business context and clear controls: Genie Ontology supplies a map of the business, Genie Search finds information across the lakehouse and connected business applications, and Unity Gateway and Unity Catalog govern its actions, usage and cost within existing permissions and policies. Genie Ontology is built automatically by agents from the dashboards, queries and notebooks people already use, and keeps learning with every query and AI interaction, while teams establish authoritative definitions in Unity Catalog. The OntoRank algorithm gives those authoritative definitions precedence, so every agent works from the same governed understanding of the business.
On a new benchmark built with Turing, Genie powered by Genie Ontology answered correctly in 91% of task runs, 1.6 times the rate of the stronger frontier agent, while using less than half as many tokens. The comparison ran Genie against two leading frontier agents on the same tasks in the same Databricks workspace. The tasks required connecting knowledge spread across tables, dashboards and notebooks, not simply accessing data.
Perspective
The report covers Databricks' new Genie release and its companion components, including Genie Ontology, Genie Search, Genie Agents, Genie plugins and Unity Gateway; the quantitative comparison is limited to the benchmark built with Turing, whose tasks require connecting knowledge spread across tables, dashboards and notebooks.
The 91% accuracy, the 1.6 times rate and the under-half token figure come from a new benchmark built with Turing; the text does not state task scale, the names of the frontier agents or evaluation details, and customer usage is given as Databricks' own account.