Databricks announces an always-on Genie and makes Genie Ontology generally available

In what the press release calls autonomous mode, Genie runs in the background on its own compute. Databricks says it can take daily work such as email, Slack and meeting preparation, or long-running work such as keeping a quarterly forecast on track. The user sets a goal and clear boundaries. Genie manages the task and reaches out when it needs input. The examples in the release are a marketer asking Genie to monitor GEO performance and suggest content, a customer success team asking it to watch usage patterns and notify the account team of a meaningful shift, and an executive asking it to track revenue trends and the feature requests of top customers, then deliver a weekly brief on what changed and why.[1]

Three controls are named. Genie delegates tasks to multiple subagents so work can proceed in parallel. The app opens on a priority inbox of items that need a decision, rather than a history of old conversations. Each new alert arrives as its own unread email. Updates on that alert go into the same thread, so the inbox does not fill with a new message for every change, and each email links back into the Genie app.[1]

TechTarget, reporting the same day, describes the change as turning Genie from an assistant people call when they need it into a worker that can be given a goal and can assign tasks to subagents. It says Genie dates from June 2024, with later products named Genie One, Genie Agents, Genie Code and Genie Ontology.[4]

TechCrunch's Julie Bort reported a demonstration in which Ghodsi, instead of giving a product demo, opened the agent he actually uses and walked her through recent chats and charts. She wrote that it analysed the parts of an employee satisfaction survey where Databricks scored poorly, and that it helped him rate and discard fixes some executives had suggested. It summarised a sales pitch to a prospect that went badly, including the features the prospect said it needed before buying. It prepared him for a podcast by reviewing earlier appearances and developing likely questions. During the interview he had it suggest what he should say to her.[5]

Bort wrote that he was using Genie to find what he is now unlikely to hear as the boss. "Have you heard of Dunbar's number? The magic number is 150," he said. TechCrunch reports that the company has 12,000 employees, and that at that size he cannot know everyone. "Anyone who has more than 150 employees, which is a lot of companies, a lot of people can get value out of this," he said, of AI agents in general and of Genie in particular. He told Bort that spending can be limited through Unity Gateway, which lets a company set usage limits and switch in less expensive open-weight models.[5]

TechTarget says Genie Ontology, a knowledge layer first introduced in June, is now generally available, and that it captures definitions and metrics across structured and unstructured data and makes that context available to agents. The ontology blog says Databricks introduced it at Data + AI Summit and is now making it generally available, with new sources for agents to learn from and new tools for people to govern what they learn.[4][2]

The blog says thousands of customers already run Genie Agents in production, including 7-Eleven, Virgin Atlantic and General Motors. It says the context an enterprise needs is larger than a team can write down by hand. It lives in dashboards, SQL queries, notebooks, BI semantic models, documents and applications, and it changes as products, campaigns and teams change. Most of Genie Ontology is learned from that work and stored as Ontology Snippets, which stay tied to the asset they came from and inherit its permissions. People govern the smaller set of definitions that must be exact, such as recognised revenue or ARR, as modelled knowledge in Unity Catalog. Where those definitions exist, the blog says the ontology treats them as the source of truth.[2]

OntoRank is the name the blog gives the ranking that chooses which context to trust. It weighs source authority, certification, the expertise and relevance of the people behind the knowledge, popularity and recency. That context is passed to the agent. Through Genie MCP, the blog says, the same context is available to agents outside Databricks, including ChatGPT, Claude and Microsoft Copilot.[2]

This release, the blog says, also makes Ontology Snippets, organisational directories, and Unity Catalog Domains, subdomains and Pages generally available. Directories start with Workday. They bring people, teams and reporting lines into the ontology, and into a people browser in Genie. External knowledge sources, which read semantic models from BI tools, are in beta. Power BI is supported today. Tableau and Sigma are described as coming soon. Ontology Manager, a visual place for stewards to see gaps and promote learned knowledge into Pages, is coming soon in beta. The press release describes that manager as a way to govern business context, with recommendations that flag frequently asked concepts which still need a tighter definition.[2][1]

Marlissa Wong, director of data engineering at Headspace, is quoted in the press release. "AI is only as good as the definitions behind it, so we built our Genie Agents on agreed, governed definitions, and now every answer matches."[1]

Databricks says it built a benchmark with Turing because existing data-agent tests mostly cover tables, and frontier agents now score similarly on those, while real questions depend on context scattered across notebooks, dashboards and documents. The test simulates a live Databricks workspace for a fictional UK energy supplier. Genie, powered by Genie Ontology, was compared with two leading frontier agents, labelled Agent A and Agent B. Every agent answered the same questions five times per task. Each answer was graded against a rubric defined before the runs.[2]

Databricks reports that Genie passed 91% of runs, at an average of 100 seconds and 432K tokens. It reports Agent A at 45%, 435 seconds and 889K tokens, and Agent B at 58%, 307 seconds and 1.05M tokens. From those results it says Genie passed at 1.6 times the rate of the stronger frontier agent, used less than half the tokens, and finished three to four times faster. The blog says the gap is context. All three agents write SQL well, it says, but Genie used business rules the ontology had already learned, while the other agents had to find that knowledge again on every run and often missed the rules. Databricks says the figures reflect an established deployment in which teams had already accumulated that knowledge.[2]

The press release says Genie Search gives people and agents one governed way to find information across the lakehouse and across documents, tickets and chats in business apps. Because the search goes through Genie Ontology, Databricks says it computes and interprets the material in business context, inside the permissions the person already has, and with sources a team can inspect. Agents can search structured and unstructured data in one pass. The release says that cuts unnecessary tool calls and token use. There is no separate search infrastructure, and access follows rules already set in Unity Catalog.[1]

The release says Databricks now offers 100+ connectors. A team can ingest data into the lakehouse through Lakeflow Connect, index content in place, or query an application live through MCP. It names Google Drive, Jira and Confluence as content that can be indexed where it already lives, without moving or copying it. Genie MCP can also let other AI tools use Genie Ontology and Genie Search.[1]

Ken Wong, senior director of product management at Databricks, told TechTarget, "Many of these features were always in the plan, but customer feedback definitely influenced our thinking." He said, "Most significantly, customers increasingly asked Genie questions that depended on information and context outside Databricks and traditional data systems, like context captured in documents, tickets and chats. And while we had federated connectors at launch, we wanted to boost performance, so we built Genie Search to index those systems as well." TechTarget reports that Ontology Manager, and pulling business logic from external BI platforms, were also responses to customer requests.[4]

The Genie One blog, by Sydney Sundell and Elise Georis, calls Genie One the home for this work: search, organisational directories, and artefacts that stay current. It says Genie Search indexes documents, tickets and wikis in tools such as Google Drive, Jira, Confluence, Workday and SharePoint, and that it supports over 100 MCP connectors for live lookups and actions, such as checking a Jira ticket or creating one. Genie chooses among the index, MCP and SQL for each task. Indexed content also feeds OntoRank. From the same chat bar, a person can search for a file, and results from Google Drive, Jira and Confluence appear beside Databricks dashboards, tables and Genie Agents. The post tells readers to check the documentation for which of these features are available.[3]

The same post says Genie One can create slide decks, dashboards and interactive pages in the conversation, on the same governed data, and that a person can inspect the source behind a number before sharing. Scheduled tasks keep the data in an artefact up to date. Skills, published to Unity Catalog, rerun a workflow with the permissions of the person who runs them. Genie can update connected tools such as Jira or Asana through MCP. The user chooses which actions need approval, and admins choose which tools a team may connect.[3]

Employees can open Genie One from a Chrome extension that can take over the new-tab page, from desktop apps for macOS and Windows, from iOS and Android apps, and from Slack and Microsoft Teams, where a person can message Genie or mention it in a channel. The post says Genie keeps that person's identity and permissions in those chats. Plugins bring the same ontology, and the same Unity Catalog permissions, into ChatGPT, Claude, Microsoft Copilot Cowork and Grok Bot. The press release lists those same outside tools, and adds email. It also says Genie Agents can read PDFs, contracts and slides in Unity Catalog volumes next to governed tables, without a separate vector index. It names new MCP connectors for Atlassian, Google Drive and GitHub, and OpenSharing, which lets teams work with partners without exposing the underlying data.[3][1]

On price, the two Databricks pages speak differently. The press release says Genie is part of the Databricks Data + AI Platform, that customers pay for what they use rather than per seat, and that admins can set spending limits by team or by user through Unity Gateway. The Genie One post says Genie One is free for users through 31 January 2027, and that more detailed cost controls will be available before then. It says Unity Gateway traces store the question, the answer, cited objects, response time and the actions taken, in a table admins query with their own compute.[1][3]

The press release marks Genie and Genie Ontology as generally available to all customers, and marks the new always-on Genie, Genie Search and Ontology Manager as coming soon to preview and beta. The always-on behaviour described above, including work on its own compute, the priority inbox and email, sits in that second group. The ontology post, beside it, calls Genie Search generally available for Google Drive, Jira, Confluence and Workday, with SharePoint in beta and more connectors on the way. The two pages, published together, do not assign Genie Search the same status.[1][2]

Analysts who spoke to TechTarget treated the ontology as the part that makes an always-on agent usable. Sanjeev Mohan, founder and principal of SanjMo, said always-on Genie will get the headlines, "but it only works if the agent understands the business," and that agents fail less from bad models than from missing context. Ontology, he said, learns definitions, metrics and relationships from the data and from usage. He singled out business logic from external BI tools, because much of that logic sits outside Databricks today.[4]

Mike Leone, an analyst at Moor Insights & Strategy, told TechTarget that handing Genie a goal, rather than waiting for someone to ask the right question, puts answers in front of people who would never open a dashboard. TechTarget reports that he sees rival Snowflake, and the hyperscalers AWS, Google Cloud and Microsoft, developing similar tools to make agents into autonomous co-workers. "Databricks is in step with the market on the overall idea, but it pulls ahead on the ontology," he said. "It picks up a company's definitions from how people actually query the data, so a team doesn't have to write them all out first." "Genie Search stands out because it changes how agents reach information spread across a company," he said. "Right now, an agent hopping from app to app needs its own access to each one, and every extra stop adds to the bill. One governed search gives you one place to set permissions and a lot fewer calls to pay for."[4]

Both analysts told TechTarget that governing what an agent does is still open. Mohan said the new capabilities make real progress on governed access, and that enterprises need to govern what agents do, not only what they read: an identity for each agent, limits on which actions it can take, a person approving consequential steps, and a full audit trail of its reasoning and actions. Leone said he would add a way for always-on Genie to say when it is not sure. "A wrong answer that shows up unprompted can reach a lot of people before anyone thinks to question it." The press release says Unity Gateway already applies approval rules and budgets, and that actions are auditable. The Genie One post says admins choose which MCP tools are allowed and which actions need a person to approve them, and that traces record what Genie did. Those pages describe approvals, budgets and traces. They do not describe a way for the agent to say when it is unsure, which is the gap Leone named.[4][1][3]

The 91% figure is Databricks' own claim, from a benchmark it says it built with Turing on a simulated workspace for a fictional UK energy supplier. The blog says the score reflects an established deployment, where teams had already accumulated that knowledge. Databricks offers that result as evidence that shared context beats a frontier agent that has to rediscover the business on every run.[2]