A coworker, sold on context

The pitch is not a smarter model. Databricks argues that enterprise agents fail because they do not know what a company means: which revenue figure is authoritative, which dashboard is stale, which notebook quietly overrides the table it builds. The new Genie is an autonomous layer on top of that problem. Give it a job and it runs continuously, on a schedule or in response to triggers, checking in when it needs input or approval. Databricks' examples are insurance claims routed to underwriters when a segment drifts outside guidelines, and marketing campaigns monitored around the clock with weekly budget changes proposed for approval.

Underneath sits Genie Ontology, generally available as of November 11. Agents build most of it automatically, extracting reusable snippets from the dashboards, queries and notebooks a company already has, while people govern the definitions that must be exact in Unity Catalog. OntoRank weighs source authority, certification, popularity and recency to pick which definition wins when sources conflict. Genie Search, also generally available, extends the reach to Google Drive, Jira, Confluence and Workday, with SharePoint still in beta, and carries each source's permissions with its content. Unity Gateway is the control layer: traces of every action, MCP tool policies, budgets and spending caps.

Not everything in the announcement carries the same status. Document analysis in Genie Agents is in Public Preview. OpenSharing and Genie Clean Rooms are in beta. External BI knowledge sources are in beta and support Power BI only, with Tableau and Sigma promised. Ontology Manager, the tool for reviewing what the ontology has learned, is coming soon in beta. The launch is real, and parts of it are still a roadmap.

The 91%, with its setup attached

Databricks built the supporting benchmark with Turing, and published the setup in the same post as the score. The workspace belongs to Calderwick Energy, a fictional UK electricity and gas supplier with hundreds of tables, the largest holding about 120 million rows of half-hourly meter data, plus notebooks, dashboards, saved queries, documents and deliberately messy assets: copied notebooks, abandoned analyses, frozen reporting tables. Each agent ran each task five times.

Genie passed 91.0% of runs, averaging 100.3 seconds and 431,829 tokens per run. Agent A passed 45.0% at 889,019 tokens. Agent B passed 58.0% at 1,051,070 tokens. Databricks does not name A or B, describing them only as leading frontier agents from frontier labs, so there is no way to check whether the comparison used each rival's current best configuration.

Two setup choices matter more than the scores. First, Genie ran natively in its own workspace, with an ontology that had already extracted the workspace's knowledge. The rival agents connected through a toolbelt that excluded Genie MCP, explicitly because that access "would give those other agents access to all the Genie capabilities we were testing." The exclusion lets Databricks compare its integrated system with outside agents lacking Genie's capabilities. It does not isolate the contribution of the ontology from the other differences between the systems, and it should not be read as a measure of the rival models alone.

Second, Databricks states the boundary itself. The benchmark tests agents in an established Databricks deployment where business knowledge has accumulated in workspace assets. In a new deployment, or on questions answerable from tables and schemas alone, Databricks says the difference may be smaller. Read plainly: the result concerns an established workspace containing accumulated business knowledge, not a general ranking of agent intelligence. A company starting from an empty workspace should not expect that gap on day one.

The worked example illustrates one task where the systems differed. The question asked for a historical planning rate. The obvious table column held 0.31 pence per kWh; the correct answer, 0.52, sat in a notebook explaining that the table had been rebuilt on a later year's assumptions. Genie surfaced that notebook through the ontology and answered correctly in all five runs. One rival returned 0.31 throughout; the other returned 0.31 or questioned whether a year-specific planning rate was documented. Neither produced the correct 0.52 planning rate. That is the risk of answering from a table without finding the historical business definition preserved elsewhere, and it is a genuine reason context layers matter. It does not need the anonymized scoreboard to make the point.

What it costs, and what is missing

Genie One is free for users through January 31, 2027, a date worth writing down, because the Genie One post does not specify what it costs after that date. The announcement says more detailed cost controls will be available before then, which Databricks says will provide more detailed controls for admins planning spend. The other missing number is adoption depth. Databricks says General Motors has deployed more than 1,000 Genie Agents; the announcement provides no associated outcome measurements for that deployment.

For a Databricks shop with years of dashboards and notebooks, the offer is coherent: the context already exists, the ontology harvests it, and the governance tools match how those companies already control data. For everyone else, this launch is a benchmark of a home-field advantage, published by the home team, with the conditions printed on the same page. Credit for printing them. The score should still travel with its conditions attached.

Sources

  1. [1] Databricks, "The new Genie: autonomous AI for complex enterprise work," November 11, 2026Read source
  2. [2] Databricks, "Genie with Genie Ontology Delivers 1.6x the Accuracy with Less than Half the Tokens," November 11, 2026 (benchmark methodology, Calderwick Energy setup, MCP exclusion, established-deployment caveat, full results table)Read source
  3. [3] Databricks, "Announcing GA of Genie Ontology," November 11, 2026 (GA/beta status of Ontology, Snippets, Search, directories, external sources)Read source
  4. [4] Databricks, "Genie One: The AI cowork home for everyone," November 11, 2026 (Genie One free through January 31, 2027; Unity Gateway traces, budgets and MCP policies)Read source
  5. [5] Unite.AI, "Databricks Turns Genie Into an Autonomous Enterprise AI Coworker," November 11, 2026 (secondary summary used for cross-checking status labels and the GM figure)Read source