When the CEO Stopped Getting the Truth
Ali Ghodsi runs a company with 12,000 employees. Every morning, the chief executive of Databricks sits down to the same problem every large-company boss quietly knows: the bigger the company gets, the less honest the information that reaches the top.
A salesperson will not walk into the CEO’s office to confess that a major pitch just collapsed. An HR manager will not volunteer that the new morale initiative is probably going to fail. Good news travels upward fast. Bad news gets filtered, softened, or simply lost.
So Ghodsi tried something different. He asked Databricks Genie, the company’s own AI agent, to tell him what nobody else would. What it showed him — an employee satisfaction survey’s weak spots, a sales pitch that went wrong, the fixes that would not work — changed how he thinks about his own company.
Now Databricks is giving that same experience to everyone else. On November 11, the company unveiled a new version of Genie that works as an always-on AI coworker, taking on multi-step jobs and finishing them without supervision.
What Databricks Genie Does Now
Databricks Genie has existed since June 2024 as a natural-language assistant for data analysis. Users asked it questions, and it answered them using the company’s governed data. It was helpful, but it waited to be asked.
The new version flips that relationship. Instead of answering one question at a time, Genie can be handed an assignment or a goal and take ownership of the work. It runs scheduled or event-triggered jobs, follows up when information is missing, and juggles several assignments at once.
According to Databricks, the agent can even assign tasks to subagents when a job gets complicated. Users decide in advance which actions need human approval and which Genie can take on its own.
The examples Databricks gives are deliberately ordinary. Genie can route an insurance claim that falls outside underwriting guidelines to the right person. It can propose changes to a marketing budget and wait for approval. The point is not drama — it is the slow, continuous work that fills an enterprise day.
The Ontology Behind the Autonomy
An autonomous agent is only as good as the context it understands. That is why the centrepiece of this launch is not a new model, but a knowledge layer called Genie Ontology, which Databricks first introduced in June and has now made generally available.
Genie Ontology automatically builds reusable knowledge snippets from dashboards, SQL queries and notebooks. It combines deliberately defined metrics and business relationships with context inferred from governed tables, queries and documents.
Unity Catalog Pages and Metric Views preserve the explicitly governed definitions, so a number means the same thing every time it is used. Databricks’ OntoRank system then weighs source authority, certification, user relevance, popularity and recency to pick the right context and resolve conflicting definitions.
The practical result is that Genie does not just know where the data lives. It knows what the business means when it says “active user” or “revenue”, and which definition is the trusted one. That shared understanding is what lets it work unsupervised without constantly asking for clarification.
Genie Search Brings the Whole Company In
Alongside the ontology, Databricks launched Genie Search, a governed way for people and agents to search across the entire data estate. It is generally available for Google Drive, Jira, Confluence and Workday, with SharePoint in beta.
The search indexes content from those systems and enforces their permissions before returning results. It can combine that index with SQL queries and MCP connectors, which retrieve live information and take actions in external tools.
Organisational directories, also generally available and beginning with Workday, add reporting relationships and collaboration signals to help rank relevant knowledge. External business-intelligence sources are in beta, initially supporting Power BI, while an Ontology Manager for reviewing and curating the ontology is coming soon in beta.
The message is clear: the agent’s world is no longer just the database. Emails, documents, tickets and org charts are all part of the context it can draw on.
Genie Agents Read Documents and Share Work
The launch also expands what Genie Agents can do. In public preview, they can now analyse PDFs, slide decks and other documents alongside governed tables, returning page-level citations for what they find.
Native MCP connectors are generally available, letting agents retrieve context and take actions such as creating tickets, updating records and posting messages. That turns Genie from a question-answerer into something that can actually get work done in the tools a company already uses.
Two sharing features are in beta. OpenSharing provides a governed agent preloaded with data, instructions and knowledge for use across Databricks accounts. Genie Clean Rooms lets recipients ask questions without seeing the underlying data or proprietary logic.
Databricks says General Motors has already deployed more than 1,000 Genie agents across sales, engineering, supply chain and HR operations — a real-world example of the multi-department model the company is selling.
Genie One: The New Home for AI Cowork
All of this comes together in Genie One, the interface where every conversation, artifact and workflow draws on enterprise context from Genie Ontology and is governed through Unity Gateway. Genie One 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, with more detailed cost controls arriving before then. The company also launched native Genie plugins for ChatGPT, Claude, Microsoft Copilot Cowork and Grok Bot on the same day, giving those assistants governed access to business context held in Databricks.
The same announcement matters for a wider trend. Anthropic launched its Claude office suite on November 10, and IBM and Microsoft unveiled their Imagine Enterprise agent platform the same week. Every major AI vendor is now racing to turn its assistant into a coworker that does the job, not just answers the question.
The Benchmark — and Its Honest Caveat
To back up its claims, Databricks released a benchmark developed with Turing. In the company’s evaluation, Genie passed 91% of task runs, compared with 58% for the stronger of two anonymised frontier agents, while averaging 431,829 tokens and 100.3 seconds per run.
Each agent ran every task five times in the same workspace for a fictional UK energy supplier, working with tables, notebooks, dashboards, saved queries and documents — including outdated and duplicated assets. The setup was deliberately messy, the way real enterprise data is messy.
Databricks itself adds an important caveat: the test reflects an established deployment with accumulated business knowledge. In a brand-new deployment, the differences may be smaller. Broader autonomy depends on governed, accessible institutional knowledge.
It is a refreshingly honest note for a product launch, and it fits the story of this release. Genie is not being sold as magic. It is being sold as the colleague who remembers everything, checks its own work, and tells the boss the truth.
What Happens Next
The controls matter as much as the autonomy. Unity Gateway gives administrators a queryable record of Genie’s work: the question, the answer, the cited objects and the actions taken. Administrators can restrict MCP tools, require approvals, and set budgets and spending caps.
Genie acts under each user’s identity and can prepare an access request, but administrators and data owners retain approval authority. That is the guardrail that makes always-on work acceptable inside a large company: the agent can move fast, but it cannot move beyond the permissions its humans set.
Cost is the other obvious question. Ghodsi told TechCrunch that costs can be controlled through Unity Gateway, which lets companies set usage limits and swap in less expensive open-weight models.
The economics of always-on agents will decide how fast they spread, and every vendor knows it. Google’s strongest new model, Gemini 4 Carbon, is being pitched on similar grounds — more capability per dollar.
For now, the most telling detail of the launch is not a feature at all. It is that Databricks’ own CEO uses Genie to find out what is going wrong in his company, before anyone else tells him.
If the agent works well enough for the person at the top, it may soon be working for everyone below him too.
Source: reporting by TechCrunch and Databricks’ own announcement, via Unite.AI.
