AI Assistants Are Still Confusing—Databricks Has Already Put One on the Long-Term Payroll
Databricks launches a new version of Genie, positioned as a persistent, autonomous AI coworker. It can take on long-running tasks, find information across systems, act according to permissions, and comes with budget and audit dashboards. The competitive focus in enterprise AI is shifting from whether answers are accurate to whether it can consistently get work done.
Databricks released the new version of Genie on November 11, and the official account promoted it again on November 18. The positioning is straightforward: a persistent, autonomous AI coworker that takes on complex, long-running enterprise work and operates within the guardrails set by the organization.
Enterprise work rarely ends after a single answer. Monitoring pipeline performance, managing budgets, and watching for claims fraud all require people to continuously track changes, fill in missing information, and act when something happens. Databricks says Genie can now take over this kind of work. Users give it tasks and tool permissions through Genie One, email, Slack, and Microsoft Teams; it handles the follow-through, comes back when input or approval is needed, and adjusts based on feedback. It retains personal memory, so preferences and instructions don’t have to be repeated every time. It notices repetitive work, such as weekly updates, and proactively offers to take it over; it can also flag issues that need attention before a task is set up. Tasks can run on a schedule or be triggered by events, such as a metric dropping or more than 10 customers requesting the same feature. Multiple tasks can run at once, and users can supervise, adjust, and stop them in a single conversation, with a control panel showing active tasks and outputs.

For Genie to work autonomously, it first has to know how the company operates. Genie Ontology is now generally available, giving AI a business map: definitions, metrics, relationships, organizational knowledge, and links across structured and unstructured data. It is built automatically by agents from the dashboards, queries, and notebooks people commonly use, with humans responsible for governance. External knowledge sources can connect business logic from BI tools such as Power BI. Teams define authoritative definitions in Unity Catalog, and the OntoRank algorithm prioritizes those definitions so every agent works from the same governed business understanding. The organizational directory starts with Workday, bringing in people, teams, and reporting relationships; OntoRank then combines collaboration activity signals to determine what is most relevant to the person asking. Ontology Manager gives teams a shared place to manage context, with admins monitoring context health and domain experts maintaining it. Domains, Pages, and Ontology Snippets in Unity Catalog support expert definitions and automatic learning. Gap alerts flag high-frequency concepts that lack governed definitions, and owners can promote learned knowledge into authoritative Pages. Databricks also built a benchmark with Turing comparing Genie and two frontier agents doing the same tasks in the same Databricks workspace. Genie answered correctly in 91% of task runs—1.6x the stronger frontier agent—using less than half the tokens. The tasks required connecting knowledge scattered across tables, dashboards, and notebooks, not just accessing data.


Genie Search is a new hybrid AI search that brings enterprise search into Databricks. There’s no need to buy separate search infrastructure or build a separate permissions model. It indexes content from core business applications such as Google Drive, Jira, Confluence, Workday, and SharePoint, and combines that with real-time search through MCP. The index provides fast, accurate access, while MCP pulls the latest updates and long-tail applications that can’t be indexed. For each task, Genie retrieves data from the index, MCP, or SQL, then synthesizes a cited answer with fewer tool calls and tokens. Search results also feed back into Genie Ontology, with tickets and documents becoming evidence for OntoRank to judge credibility. Through Unity Catalog, Genie Search maps access controls together with content for each source and enforces permission checks before returning results, so users won’t see documents in Genie that they can’t open in the original application.

Genie Agents give teams specialized capabilities. Departments configure agents with data, files, tools, and instructions, and manage access and usage. Teams can use them in Genie One or through APIs, custom applications, and automated workflows. They can now handle structured tables and also read PDFs and slide decks. Content search in Unity Catalog Volumes makes this possible without building a separate vector index. When a sales leader asks why deals are being lost in a region, they can get an analysis that combines pipeline data with win/loss reports, with page-level citations so the evidence is easy to check. Connectors let agents read from and act in Slack, Atlassian, GitHub, Gmail, Google Drive, and Microsoft 365. Agents can investigate issues, create tickets, and prepare summaries for team channels. OpenSharing for Genie Agents lets teams share governed agents with partners and customers, preloaded with data, instructions, and knowledge. Genie Clean Rooms let recipients ask questions and get answers without exposing underlying data and proprietary logic, with support for conversation and output restrictions. Priyanka John, Vice President at S&P Global Energy, said: “Genie Agents let our domain experts directly productize data knowledge. What used to take a full development cycle now takes days, and every answer stays within the governance boundary.”

Genie can also generate finished outputs in a conversation: describe the slides, dashboard, or interactive HTML page you need, and it builds it directly, lets you edit it in chat, then presents, shares, and reuses it. Scheduled tasks keep artifacts up to date as data changes, and shared skills apply company brand guidelines. Guilherme Vivian, Data Manager at Agibank, said: “With artifacts in Genie, we have full control over how data is presented. We get the look we want while everything remains connected to governed data—we get both.” MCP connectors cover Slack, Zoom, Microsoft 365, Jira, GitHub, and calendars, and custom MCP can connect to ServiceNow and PagerDuty. Genie can check GitHub and Jira changes to locate incidents, update tickets, and keep the team in sync until resolution. Repeatable workflows can be published as skills and shared across teams; skills live in Unity Catalog and run with the user’s permissions. You can @Genie in Slack and Teams, and use it in the browser, on desktop, iOS, and Android. Plugins bring Genie’s business context into ChatGPT, Claude, Copilot, and Grok Bot. Genie MCP can also connect directly to other AI tools. External agents can hand entire business problems to Genie or request the context needed for larger tasks; Genie applies governed definitions, respects user permissions, and cites sources.


The more work Genie takes on, the more admins need to control who can use it, what it can access, and how much it costs. Unity Gateway brings these controls together, letting admins govern coding agents, models, MCP tools, and skills alongside it. For data access, Genie acts as the user and only accesses information the user is permitted to see, following Unity Catalog permissions; content indexed by Genie Search also follows each source system’s permissions. When users need additional permissions, Genie can prepare an access request for the user to review and send, while admins and data owners retain approval authority. For tools and actions, admins control connector access, MCP tools, and approval requirements; business users choose what Genie can do autonomously within those limits. Unity Catalog governs reusable skills, and external agents such as Claude are equally controlled when calling Genie via MCP. For usage, cost, and audit, every Genie action is traceable and auditable. Admins can set workspace and per-user budgets, spending caps, and alerts, and can have Genie investigate wasted tokens and expensive tool calls in usage and agent traces.
The term AI coworker has already been overused. What’s worth paying attention to in Databricks’ launch is that it has turned three things—context, permissions, and cost—into product: Ontology provides business understanding, Search provides cross-system information, and Unity Gateway provides controls and budgets. As enterprise AI moves from Q&A to continuous execution, model capability is only the entry ticket. What really determines whether it can “show up for work” is whether it knows how the company operates, what it can touch, and how much it spends. Otherwise, the “coworker” could turn into a budget black hole on day one.
Original article: The new Genie: autonomous AI for complex enterprise work
Tweet: Databricks on X
发布时间: 2026-11-19 04:28