Databricks launches Genie One, a data-smart AI coworker for business users

Databricks launched Genie One on Tuesday, an "agentic AI coworker" that the company says is designed to put a data-smart assistant in the hands of every business user, not just analysts. The release, the centerpiece of the company's Data + AI Summit keynote, ships with a new context layer called Genie Ontology and a way to turn one prompt into a shareable agent the whole team can use.

What Genie One actually does for a business user

Genie started in 2023 as a chat-with-your-data product inside Databricks AI/BI. The point was to let a non-technical user ask a question of a governed dataset and get an answer. Genie One is the same idea, except it is meant to be a coworker, not a chat window. Users @mention Genie in Slack or Microsoft Teams and get an answer in the same thread, with citations back to the source data and the permissions of the user who asked. The product also ships as a native iOS and Android app so a sales lead can ask a revenue question from the airport, and as an MCP (Model Context Protocol) app so organizations that already run an agent in Claude, ChatGPT, or a homegrown system can route questions through Genie without changing the user surface.

The new piece under the hood is Genie Ontology. Databricks describes it as a living context graph that automatically extracts snippets of knowledge from tables, queries, dashboards, pipelines, and connected apps and figures out which ones are authoritative. The model is the same one Google Search uses to rank pages: a PageRank-style approach that weighs who wrote a definition, how often people use it, how it ties to certified assets, and how fresh it is. When a user asks a question, Genie pulls from the sources with the most weight and enforces the permissions on each one. In internal benchmarks, Databricks says Genie answered 84.5 percent of complex enterprise data questions correctly on the first attempt, while the strongest general-purpose coding agent it tested managed 52.4 percent, and the weakest only 25 percent. Genie also returned answers about twice as fast as the strongest coding agent, which the company attributes to the same context work.

The second new piece is the path from a single prompt to a shareable agent. Customers have built more than a million "Genie Spaces," curated chat experiences scoped to a specific topic. Genie Spaces now becomes Genie Agents. A user describes what they want in plain language inside Genie One or Genie Code, scopes the agent to a specific data set and set of tools, runs an internal benchmark against it, and shares the agent for teammates to use or customize. Databricks' framing is that domain experts can now scale their own expertise by turning trusted rules, data, and workflows into coworkers the whole team relies on, instead of having to write a custom application for every recurring business process.

The third piece is the rest of the stack that the launch leans on. The new Reyden compute engine, named for Databricks co-founder Reynold Xin, is the first major compute release since the original Lakehouse and is built to deliver millisecond query latency for tens of thousands of concurrent agents. Databricks also extended Unity AI Gateway, the governance layer it launched in March, to cover MCP, tools, and cost controls, so an admin can see every model call, every tool invocation, and the running bill, and shut any of them off. The whole package is positioned as an answer to the same pain point that has dogged enterprise AI rollouts all year: the answer the agent gives is only as good as the context it has, and the context is scattered across dashboards, queries, pipelines, wikis, tickets, documents, and chat threads.

The shape of the agent market after Tuesday

The launch lands on a day that is otherwise dominated by agent stories. It is the third major "enterprise AI agent" product launch of the last week, and the most ambitious of the three. Beyond Identity's Ceros brought device-bound passkeys to agent identity on Monday. AppViewX brought PKI-backed agent credentialing and audit on Tuesday morning. Genie One brings a complete business-user agent, a context graph, and a new compute engine in the same announcement. Read in order, the three launches describe the layers of the stack that the enterprise agent market is coalescing around: who is the agent, what can it touch, and what does it actually do. None of them solve the trust problem alone. Together, they describe what a 2026 enterprise agent rollout looks like when the buyers are serious. If you are still sorting out where agent identity fits into your stack, our enterprise AI governance checklist is a useful starting point. For a broader look at who is actually adopting agents and what they are using them for, our adoption survey from earlier this month has the numbers.

The competitive context matters too. Snowflake, Databricks' largest direct rival, has been pushing Cortex Agents and its own MCP integrations for the last six months. The Snowflake pitch is that the data, the warehouse, and the agent should be one platform. Databricks' pitch on Tuesday is similar but with two distinguishing claims. The first is that Genie Ontology is doing the same kind of authority ranking for the enterprise that PageRank did for the public web, and the second is that the new compute engine is fast enough that a single tenant can run thousands of agents against the same data without falling over. Whether those claims hold up under independent benchmark scrutiny is a question for the analyst community in the next month. The most credible test will be the independent Customer Experience reports from the large Databricks customers running the new product in production.

The keynote also gave the market an unusually candid read on the state of the agent industry. Ali Ghodsi, Databricks' co-founder and CEO, said during the keynote that "AGI is already here" but that the problem is not intelligence, it is permeation. "How do we enable this at work?" Ghodsi asked. He was blunt about cost. "Every organization is super worried about costs going through the roof. It's the No. 1 question that we get asked." The same cost theme showed up in the product: Unity AI Gateway now lets an admin set a per-agent budget and an alert before the bill arrives, and the new MCP app makes it possible to route a non-Databricks agent through Databricks governance without rewriting the agent. The framing is that the next phase of the agent market is going to be won on cost controls and audit logs, not on raw model capability. Greg Brockman, OpenAI's president, was on the same stage later in the day and offered a more measured read. "It's almost like AGI is a feeling, not a defined thing," Brockman said. "It's never been a better time to be a builder." The two reads are not actually far apart. The capability question is no longer the bottleneck. The bottleneck is the boring infrastructure underneath.

Three near-term questions to track

The launch opens three near-term questions. The first is whether Genie Ontology's authority ranking is actually portable. The current product is tightly bound to Databricks' own data catalog, and the open question is whether the same approach can be applied to non-Databricks systems, like a Salesforce instance or an internal wiki, without losing the authority signal. The second is whether the Genie MCP app becomes the default way for non-Databricks agents to reach governed data, or whether the major agent platforms build their own connectors and route around it. The third is what the new compute engine means for the Snowflake rivalry, especially for customers running large agent fleets on tight latency budgets. The answers will come quickly. Genie One is generally available to all Databricks customers today, and the new agents, the MCP app, and the mobile apps are listed as available now. The mobile apps in particular are a leading indicator worth flagging: the willingness of a serious data platform to ship a consumer-style app for the data product is a sign of how much the audience has expanded beyond analysts. The full Databricks launch post has the product walkthrough, benchmark methodology, and the rollout dates for each component.