Databricks Genie stopped waiting to be asked

Databricks introduced Genie in June 2024 as a conversational data assistant: users asked a question in plain language, and it wrote and ran SQL on governed company data, then returned tables and charts. In June 2026 it added Genie One, which Databricks later called the enterprise home for AI cowork. Those earlier versions answered after a person asked.

Databricks says the remaining limit was that employees still had to manage the work around those answers — tracking changes, deciding what matters, and taking action.

On November 11, Databricks announced what it calls a second-generation, always-on Genie. The company says employees can give Genie a goal and it will take ownership of the work: run jobs on a schedule or in response to triggers, delegate to subagents when needed, follow up on missing inputs, and reach out when a person needs to make a decision. Users choose which actions require approval; Databricks says the rest runs in the background on Genie's own compute. Databricks' press release states that the new always-on Genie is coming soon to preview and beta programs.

Databricks says that work is grounded in Genie Ontology, its context layer, which the company announced as generally available (Databricks release notes date that GA to November 9). The ontology combines governed metric definitions with context the system infers from dashboards, queries, notebooks, and documents, and a ranking method called OntoRank weighs which sources to trust. Genie Search is also generally available now, indexing knowledge in Google Drive, Jira, Confluence, and Workday — SharePoint is in beta — so Genie can answer from documents and tickets as well as tables. Genie One turns answers into finished work: slide decks, dashboards, and interactive pages built in the conversation, with scheduled refreshes and shared skills so a workflow one person builds can be reused by a team.

Native Genie plugins bring it into ChatGPT, Claude, Microsoft Copilot Cowork, and Grok Bot, with the same governed context and permissions. Genie Agents can now read PDFs and slide decks alongside governed tables, in public preview, returning page-level citations.

Databricks released a benchmark built with Turing; the company says Genie's edge comes from business context already in the workspace, not from table schemas alone. In the company's evaluation, Genie passed 91% of task runs, against 58% for the stronger of two anonymized frontier agents. Each agent ran every task five times in the same workspace, a fictional UK energy supplier stocked with tables, notebooks, dashboards, saved queries, and documents — including outdated and duplicated assets. Genie averaged 431,829 tokens and 100.3 seconds per run.

Databricks itself cautions that the test reflects an established deployment with accumulated business knowledge, and that the gap may be smaller in a new deployment, or where table schemas alone are enough. The 91% is from Databricks' own evaluation, not an independent one — no outside lab has published one yet. Databricks also says General Motors has deployed more than 1,000 Genie agents across sales, engineering, supply chain, and HR, and that thousands of customers already use Genie in production.

Unity Gateway traces the question, answer, cited objects, response time, and actions taken, and admins can restrict which tools Genie uses, require human approval for actions, and set budgets and spending caps. Databricks says Genie acts under each user's own identity and permissions, so it sees only the data that person is allowed to see. Databricks CEO Ali Ghodsi, who said he uses Genie himself, put it this way: "Genie has made me a 10x more productive CEO, and now it doesn't wait for me to ask." In a walkthrough with TechCrunch he used it to examine low scores on an employee satisfaction survey, rate fixes executives had suggested, summarize a sales pitch that went poorly, and prep for a podcast. TechCrunch described this as a source of unvarnished truth; Ghodsi cited Dunbar's number and said anyone with more than 150 employees can get value out of this.

Databricks frames the new Genie as a move from answers to follow-through for organizations that already have governed data and accumulated workspace knowledge. The company itself says the benchmark gap may be smaller in a new deployment or where table schemas alone are enough. Databricks says Genie One is free for users through January 31, 2027, with more detailed cost controls to follow before then. The press release says the new always-on Genie is coming soon to preview and beta programs.