For Windows and Microsoft 365 administrators, there are concrete developments beneath the “AI coworker” branding: a Windows desktop app, a generally available Teams integration, and connections to Microsoft 365 information. But the deployment details matter. Indexed search is not the same as live retrieval, generated artifacts have their own sharing boundaries, and diagnostic traces can expose sensitive content to their readers.
Windows gets a desktop app; Teams reaches GA
Databricks’ November release notes identify November 11, 2026, as the Beta release date for the Genie One Windows desktop app. The Databricks Genie app in Microsoft Teams became generally available on November 9, supporting direct messages and @Databricks Genie mentions in channels or group chats. These are separate ways to reach Genie, not interchangeable installations. Databricks also warns that staged releases can take a week or more to reach an account.
The desktop application offers the same Genie One capabilities as the web experience, including work with conversations, documents, Genie Agents, Unity Catalog tables, and Databricks Apps. It adds navigation controls and automatic background updates. One small but useful correction to the broader announcement: the desktop documentation specifies that the unread-activity badge is macOS-only, not a Windows feature.
For an authorized Windows pilot, Databricks documents this installation and sign-in sequence:
- Download the Windows installer from Databricks’ desktop documentation.
- Run
GenieOne-latest-Setup.exe; the application opens after installation. - Select Log in, then choose the account and workspace.
- If the normal flow fails because of account SSO or email-verification issues, use Enter workspace link to sign in directly.
Administrators should also note that Databricks provides no dedicated organization-wide desktop-app enable/disable control. Its documentation points to endpoint management, such as MDM, for controlling the client. Signing out clears the application session, but the workspace session remains valid until expiry.
Search combines an index with live connections
Genie Search brings external knowledge into the same discovery experience as Databricks assets. The announcement also adds organizational directories, company-wide and personal homepage pins, and source-authority ranking through OntoRank. Databricks says Genie can select search, MCP, or SQL according to the task, with support for more than 100 MCP connectors.
The important architectural distinction is between indexed content and live application access:
| Access method | What administrators should expect |
|---|---|
| Genie Search | An account-admin-connected, read-only index whose content and permissions refresh over time |
| MCP connectors | Per-user authenticated connections for live lookups and supported actions |
Databricks documents these as complementary mechanisms, not two names for the same service.
There is also a Microsoft-specific documentation boundary. Although the announcement names SharePoint among external knowledge sources, the inspected Genie Search setup page lists Google Drive, Jira Cloud, Confluence Cloud, and Workday as indexed sources. SharePoint is separately documented through the Microsoft 365 MCP connector. Administrators should not assume SharePoint has the same indexing path.
Search setup has practical consequences: initial indexing can take hours or days, and every connected source in an account uses the same selected data-residency geo. Document access is enforced using source permissions, with additional query-time checks for non-public documents. Organizational directories are an exception: everyone using Genie One in the account sees the same directory profile details.
Finished work has a separate sharing boundary
Databricks says Genie One can create decks, dashboards, and interactive pages inside conversations, with scheduled tasks maintaining artifact data and skills applying branding. Skills can be shared through Unity Catalog and run with the permissions of the person using them.
The artifact documentation adds an essential qualification: artifacts store snapshots of query results. They do not automatically change whenever their source tables change. Scheduled maintenance should therefore be understood as a workflow, not an inherent live-data connection. Artifacts are in Beta and require a workspace administrator to enable the feature.
More importantly, anyone permitted to view an artifact can see its embedded snapshot data—even without permission to access the original tables. Sharing the conversation does not share the artifact; its permissions are managed separately. Before distributing a deck, use Data mode to inspect the datasets behind its elements and verify that the intended recipients should receive them.
That makes artifact sharing a publication decision, not merely a convenience. “Governed source data” does not mean every downstream presentation inherits the original table-access boundary.
Microsoft 365 actions have real limits
The Microsoft 365 MCP connector can search SharePoint, Teams, Outlook, and Calendar, and draft Outlook messages. It cannot send Outlook email. Supported write actions depend on the tool, OAuth scopes, source-system permissions, and administrative policies—not simply the wording of a prompt.
Consent deserves particular attention. Databricks documents a single Microsoft Graph consent covering several Microsoft 365 services; administrators cannot approve only the Outlook portion, for example. Removing write tools in Unity Gateway restricts callable functionality but does not reduce the permissions requested on the consent screen.
The announcement also describes mobile apps, a Chrome extension, Slack access, and plugins for external assistants, including Microsoft Copilot Cowork. Those additional entry points broaden access, but they should not be mistaken for identical deployment or permission configurations.
For reliability, include authentication expiry in pilot testing. Databricks says chat may not warn when a connector’s credentials expire, and permission changes require users to authenticate again. Scheduled tasks using that connection can fail until they do.
Governance includes protecting the monitoring data
Unity Gateway’s tracing capability can record prompts, responses, cited assets, outcomes, and actions. It is useful for investigating incorrect answers and failed tool calls—but monitoring creates another sensitive dataset.
Databricks explicitly warns that trace readers can see full prompts, responses, and tool arguments or results, including material from conversations and data they otherwise could not access. Review trace-table permissions and retention before enabling collection. Without row filtering, table access can expose traces across workspaces attached to the metastore.
There are further limits:
- Tracing is in Beta and delivery is best-effort; audit logs remain the compliance system of record.
- Genie Agent conversations are not fully captured.
- Deleting a conversation does not delete its traces; retention policy governs removal.
The sensible rollout starts with boundaries
Databricks says Genie One is free for users through January 31, 2027, with more detailed cost controls promised before then. That statement should not be stretched into a guarantee that every associated service is cost-free.
The strongest reason to evaluate this release is the connection between discovery, analysis, deliverables, and actions. The strongest reason to proceed carefully is equally clear: source permissions, artifact sharing, connector consent, and trace access are different controls. A useful pilot should test all four—not just whether Genie produces a convincing answer.