Databricks on Wednesday turned its Genie assistant into an always-on AI coworker that takes on long-running enterprise jobs by itself, in a launch aimed at moving workplace AI from answering questions to finishing work.

Key Points

  • Databricks unveiled the second generation of Genie on 11 November 2026, pitching it as an autonomous AI coworker rather than a question-answering assistant.
  • Genie can now take a goal, work on it in the background, hand parts of it to sub-agents, chase missing information and come back to a person only when a decision is needed.
  • Genie Ontology, the knowledge layer that teaches the agent how a company works, is now generally available, alongside a new governed search tool, Genie Search.
  • Databricks says General Motors runs more than 1,000 Genie agents, and its own benchmark put Genie ahead of two unnamed frontier agents — figures the company itself produced.
  • Independent analysts praised the knowledge layer but warned that firms still need controls over what autonomous agents do, not only over what they can read.

What happened

Databricks announced the update on Wednesday from San Francisco. Genie, first introduced in June 2024 as a natural-language assistant for data analysis, now runs in autonomous mode on its own computing power. A worker gives it a goal and sets boundaries, and the agent manages the task, splits work between sub-agents, follows up when information is missing and alerts its owner when a human choice is required.

The agent opens to a priority inbox instead of a blank chat window, so the items needing a decision stand out. Each new alert also arrives as its own unread email, with later updates kept in the same thread so inboxes do not fill up.

At the centre of the release sits Genie Ontology, a knowledge layer first shown in June and now generally available. It captures a company’s working context — what its definitions and metrics mean, which figures are official, who can see what — and keeps that picture current as the business changes. Databricks says the layer builds reusable knowledge from dashboards, SQL queries and notebooks, and weighs source authority, certification, relevance, popularity and recency when definitions conflict. A new Ontology Manager lets data teams review and refine that knowledge, with suggestions flagging the concepts people ask about most.

Alongside it comes Genie Search, a governed way for people and agents to find information across a company’s data and the documents, tickets and chats in its business apps. It is generally available for Google Drive, Jira, Confluence and Workday, with SharePoint in beta, and works through more than 100 connectors. Genie Agents can now also read PDFs, slide decks and contracts alongside governed tables, returning page-level citations, while new sharing tools let partners work with an agent without seeing the underlying data.

Teams can reach Genie in a browser, a mobile app, email or inside other tools, including ChatGPT, Claude, Microsoft Copilot and Grok. A Chrome extension entered beta on 10 November and a Windows desktop app for Genie One entered beta on 11 November. Genie One, the interface that also builds slide decks, dashboards and interactive pages, is free for users through 31 January 2027, Databricks says. Pricing for the wider product stays usage-based, with no per-seat fees and up to $10 in free credits per user each month.

Why it matters

The launch is the clearest sign yet of where enterprise AI is heading: away from assistants that wait for a prompt and towards agents that own a piece of work. Databricks is not alone — rival data platforms and the big cloud providers are building similar tools — but analysts say its knowledge layer sets it apart for now.

“Genie stops waiting to be asked,” Mike Leone, an analyst at Moor Insights & Strategy, told TechTarget. “Up to now, getting value out of Genie meant knowing the right question and when to ask it. Handing it a goal instead puts answers in front of folks who would never open a dashboard in the first place.”

Sanjeev Mohan, founder of analyst firm SanjMo, said the ontology is what makes the rest possible. “Always-on Genie will get the headlines, but it only works if the agent understands the business,” he told TechTarget. “By now, we understand that agents fail less from bad models than from missing context.” An agent that does not know what “active customer” means at one company, or which revenue figure the finance team trusts, cannot give a reliable answer, he said.

Databricks chief executive Ali Ghodsi offered his own use as the proof. He showed TechCrunch his live agent analysing the parts of an employee satisfaction survey where the company scored badly, rating — and discarding — fixes suggested by executives, summarising a sales pitch that went poorly, and preparing him for a podcast interview by reviewing his earlier appearances. With 12,000 employees, he said, a chief executive rarely hears bad news directly, and the agent had become a source of unvarnished truth. “Genie has made me a 10x more productive CEO, and now it doesn’t wait for me to ask,” he said.

To back its claims, Databricks released a benchmark built with Turing. In the company’s own test, Genie passed 91% of task runs, against 58% for the stronger of two unnamed frontier agents, averaging about 431,829 tokens and 100.3 seconds per run. Each agent ran every task five times in a workspace modelled on a fictional British energy supplier, filled with tables, notebooks, dashboards and documents, including outdated and duplicated files. Databricks itself cautions that the test reflects a deployment with accumulated business knowledge, and that the gap may shrink in a new setup. The company also says General Motors has deployed more than 1,000 Genie agents across sales, engineering, supply chain and human resources. By mid-2026, more than 1.5 million Genie Spaces — the workspaces where agents live — had been created.

The same shift that makes the tool attractive also carries a warning. Mohan said companies need to govern what agents do in production, not only what data they can read — a clear identity for each agent, limits on its actions, human approval for consequential steps and a full audit trail. Leone added that always-on Genie should be able to say when it is unsure of an answer, because “a wrong answer that shows up unprompted can reach a lot of people before anyone thinks to question it.” Databricks says every action stays within Unity Catalog permissions and Unity Gateway approval rules and budgets, with a trace recording the question, answer, cited objects, response time and actions taken.

The launch lands in a crowded week for workplace AI. Google’s Gemini Robotics 2.5 is putting humanoid robots to work on Mercedes-Benz factory lines, Google’s Gemini 4 Carbon is now available to paid developers, and IBM and Microsoft have just launched their own enterprise AI offering, Imagine Enterprise — a sign that every large platform now wants to own the agent that does the work, not just the model that answers the question.

What happens next

Databricks will be watched on two fronts: whether customers hand Genie real, recurring work rather than one-off questions, and whether its governance controls hold up once agents start acting without being asked. The free Genie One period runs until the end of January 2027, which gives companies a window to test the agent before paying. Rivals are moving on the same ground, so the contest is now over whose agent understands a business best — and whose stays inside the lines.

Source: Databricks announcement, SD Times, TechTarget, TechCrunch, Unite.AI, CryptoBriefing — 11 November 2026. Time: 03:33 +08, 12 November 2026.