Genie Just Became a Coworker, Not a Chatbot

Databricks announced today that Genie — the thing most data teams knew as "ask your data a question, get a chart back" — is now an autonomous AI coworker. It runs scheduled and event-triggered jobs, chases down missing information on its own, and juggles several assignments at once. You still decide which actions need your approval first.

I've been watching the "AI coworker" label get thrown around all year. Most of the time it means a chatbot with a calendar integration. This one is different in a way that actually matters to anyone running a data team: Genie already lives inside the warehouse, the governance layer, and the dashboards. The question was never whether it could answer questions. The question was whether it could keep working after you close the tab.

Full disclosure up front: I haven't used any of this. This is an announcement-day write-up, built from Databricks' announcement as reported by Unite.AI on November 11, 2026. A lot of what's below is generally available, but several pieces are still preview or beta — and the headline benchmark number comes from Databricks' own testing. I'll flag every one of those as we go.

What actually changed

Genie works while you're away. The new Genie takes on continuous, long-running assignments. It can run on a schedule or when an event fires, follow up when information is missing, and manage multiple jobs at the same time. The control stays with you: you decide which actions require approval before Genie takes them. That's the line between a coworker and a runaway script, and it's the right line to draw.

Thousands of customers already run Genie in production. Databricks gave two concrete examples of the kind of work it's doing: routing insurance claims that fall outside underwriting guidelines, and proposing marketing budget changes for a human to approve. Those aren't demos. They're the boring, repetitive, genuinely expensive kind of work that eats a data team's week.

Genie Ontology is now generally available. This is the piece I think matters most, and it's the least flashy. Ontology automatically builds reusable knowledge snippets from your dashboards, SQL queries, and notebooks — the definitions your team has already argued about and settled, captured instead of re-explained every time someone asks a question. Unity Catalog Pages and Metric Views preserve the definitions that were explicitly governed. Underneath, the OntoRank system weighs source authority, certification, user relevance, popularity, and recency to pick the right context and resolve conflicting definitions. If you've ever watched two dashboards disagree about what "revenue" means, you know why this exists.

Genie Search is now generally available for Google Drive, Jira, Confluence, and Workday, with SharePoint in beta. It indexes content from those systems, enforces their permissions, and can combine that index with SQL and MCP connectors. Organizational directories are also generally available, starting with Workday. In plain terms: Genie can now answer questions that span your warehouse and the documents where the actual context lives.

External BI knowledge is in beta, starting with Power BI. And an Ontology Manager for reviewing and curating the ontology is coming soon in beta. Both worth watching — the ontology is only as good as the curation behind it.

Genie Agents can now read documents, not just tables — in Public Preview. PDFs and slide decks sit alongside governed tables, with page-level citations coming back in the answer. Native MCP connectors are generally available, so agents can pull context and take actions: creating tickets, updating records, posting messages. OpenSharing and Genie Clean Rooms are in beta. The scale proof point Databricks offered: General Motors has deployed more than 1,000 Genie agents across sales, engineering, supply chain, and HR.

Genie One is the interface for all of this. It also creates slide decks, dashboards, and interactive pages, with scheduled refreshes and shared skills for repeatable workflows. Databricks says Genie One is free for users through January 31, 2027, and that more detailed cost controls will arrive before then. If you're evaluating this, that free window is your real trial period — use it to measure what the thing costs you in tokens and attention before the bill starts.

The controls, because they matter more than the features

Unity Gateway gives admins the levers: traces capture the question, the answer, the cited objects, response time, and the actions taken. Admins can restrict which MCP tools Genie can use, require approvals, set budgets and spending caps, and review usage. Genie acts under each user's identity, and admins and data owners keep approval authority.

This is the part I'd actually interrogate before rolling it out. An agent that can create tickets, update records, and post messages — while running on a schedule, unsupervised — is exactly the kind of thing that goes wrong quietly. The traces and spending caps are the right answers. Whether your team actually reads the traces is a different question, and it's yours to answer honestly.

The 91% benchmark — read the fine print

Databricks built a benchmark with Turing and reports that Genie passed 91% of task runs, against 58% for the stronger of two anonymized frontier agents. Genie averaged 431,829 tokens and 100.3 seconds per run. Each agent ran every task five times in the same workspace, working on a fictional UK energy supplier.

Two things to take from that, and one thing not to:

• The gap is large, and the setup was at least a fair fight — same workspace, same tasks, five runs each.
• Databricks itself cautions that the test reflects an established deployment with accumulated business knowledge, and that differences may be smaller in new deployments. That's an honest caveat, and it's the most important sentence in the whole benchmark story.
• Don't take 91% as "Genie is better than frontier models at everything." Take it as "Genie is very good at the kind of multi-step work that lives inside a governed Databricks deployment where the business knowledge is already there." Those are different claims, and only the second one is supported.

What I'd actually do

If your team already runs on Databricks, this is worth a real look — not because of the benchmark, but because the pieces that are generally available now (Ontology, Search, MCP connectors) solve problems you already have: conflicting definitions, context trapped in docs, agents that can't act on what they find.

Start with one narrow job. Pick a recurring task that currently eats a few hours a week — the kind with a clear trigger and a clear approval step, like the insurance-claims routing example. Give Genie the job, keep approval on for every action, and read the traces for a week. If the traces look clean and the follow-ups are sensible, widen the scope. If the traces show it guessing at missing information instead of asking, you've learned something important for the price of one experiment.

And if you're not on Databricks? This announcement doesn't change your stack. It does tell you where the market is going: the next wave of data agents isn't about answering questions faster. It's about staying on the job after the question is asked.

What would you actually hand to an agent like this first — the repetitive reporting, or something with real decisions attached? I'd love to hear where you'd draw the line.

Source: Unite.AI, "Databricks Turns Genie Into an Autonomous Enterprise AI Coworker" (November 11, 2026). Nothing here is hands-on tested; the 91% benchmark is vendor-run with Turing, and several features noted above are in preview or beta.