Databricks strengthens Genie for continuous enterprise tasks
Databricks has introduced a new Genie capable of chaining long tasks, requesting validations, and acting on multiple business tools. The version relies on a broader governance foundation, with search, context, and access control.
Databricks unveiled a new Genie on November 11, 2026, an AI assistant designed as an autonomous teammate for continuous enterprise work. The platform aims for it to handle long tasks, track changes, follow up when information is missing, and execute multiple missions in parallel, while allowing users to decide on actions requiring validation. The goal is simple: to elevate AI from co-pilot status to an agent capable of going the distance, without losing control over permissions.
From chat to monitored execution
Genie is no longer limited to answering a one-off question. In the logic described by Databricks, it can monitor a folder, wait for a new item, then act when the context changes, targeting use cases such as routing out-of-policy claims or proposing budget adjustments for approval. More than a conversational interface, the tool aims to become an orchestration layer between data, decisions, and business tools.
This increase in autonomy relies on Genie Ontology, now generally available. The system automatically builds reusable knowledge blocks from dashboards, SQL queries, and notebooks, while Unity Catalog Pages and Metric Views maintain governed and explicit definitions. Databricks also explains that its OntoRank system weighs source authority, certification, user relevance, popularity, and recency to differentiate competing definitions. Without this filtering, enterprise assistants quickly get lost in their own archives.
Research, documents, and external tools: the stack thickens
Genie Search is now generally available for Google Drive, Jira, Confluence, and Workday, with SharePoint still in beta. The engine indexes this content, checks access rights before any response, and can combine this index with SQL and MCP connectors, the latter used to retrieve live information or trigger actions in external tools. Organizational directories, also generally available and initially compatible with Workday, add hierarchical relationships and collaboration signals to better classify useful knowledge.
Another added component: Genie Agents can analyze PDFs, presentations, and other documents, in addition to governed tables, in a public preview. Responses include page-level citations, facilitating human review. Native MCP connectors are also generally available; they allow creating tickets, updating records, or posting messages. Meanwhile, OpenSharing and Genie Clean Rooms are in beta for partner and client use cases, with a logic of governed sharing or querying without exposing underlying data.
Governance and control remain the real issue
Databricks emphasizes safeguards. Unity Gateway keeps records of every question, answer, cited object, response time, and action taken, providing administrators with a searchable log of Genie's work. They can restrict MCP tools, enforce validations, set budgets and spending caps, then review usage. The agent acts under each user's identity and can prepare an access request, but final approval remains with administrators and data owners.
The Genie One product, the interface for this system, can also create presentations, dashboards, and interactive pages, with scheduled refreshes and shared skills for reusable flows. Databricks announces it is free for users until January 31, 2027, before more detailed cost controls arrive. This kind of free access feels less like a gift and more like a carefully monitored trial period.
Les chiffres qui disent tout sur l’ambition
To support this bet, Databricks published a benchmark designed with Turing. In this exercise, Genie succeeded in 91% of tested tasks, compared to 58% for the best of two anonymized border agents compared by the company. In the same evaluation, Genie used on average 431,829 tokens and 100.3 seconds per execution.
The test was based on a fictitious British energy company with tables, notebooks, dashboards, saved queries, and documents, including duplicates and obsolete content. Each agent executed each task five times in the same workspace. Databricks warns, however, that this result reflects a deployment already rich in business knowledge; in a new environment, or when table schemas suffice, the gap may narrow. This is where the launch makes perfect sense: autonomy truly advances only if the knowledge base is governed, accessible, and sufficiently clean.
What the market should take away
Databricks' strategy is clear: make Genie an agent capable of working over time, but without bypassing administration or access rules. The company targets cross-functional use cases, from finance to insurance to operations, with agents deployed across multiple departments and tools. According to Databricks, General Motors has already deployed more than 1,000 Genie agents in sales, engineering, supply chain, and human resources.
This approach puts competition on concrete ground: who will be able to link business context, enterprise search, external actions, and fine-grained rights control. The agent market is full of promise; it still often lacks organized memory. Databricks wants precisely to sell that memory.