Gemini Enterprise for Legal Arrives—Permissions Travel Through MCP
Google Cloud launched Gemini Enterprise for Legal in preview on August 25, 2026, combining legal skills, agents, connectors and centralized governance; launch-day coverage from Legal IT Insider corroborates the date, preview status, initial components and participating law firms. The product is aimed at law firms and corporate legal departments, with Cleary Gottlieb, Freshfields, Weil and Williams & Connolly identified as early customers.
The business-critical distinction is architectural: Gemini agents can reach work held in existing legal platforms through Model Context Protocol connections, while those platforms continue to determine which matters, documents and operations a user may access. Firms are therefore not being asked to flatten ethical walls and repository permissions into one shared access layer, although the exact identity, processing and audit model still depends on each connector.
The permission decision remains with the connected platform
Gemini Enterprise governs the agent environment, but an MCP connection does not automatically replace the authorization boundary of a document-management or e-discovery system. In the announced design, the source platform evaluates its established roles, matter restrictions and document-level controls before returning context or performing an action.
The permission flow can be mapped across five boundaries:
- A user starts an agent workflow. Gemini Enterprise applies the organization’s agent configuration and platform-level governance.
- The agent invokes an MCP connector. The request passes through the authentication arrangement supported by that particular integration.
- The legal system evaluates access. The connected platform applies its existing user, matter and document permissions.
- Permitted context or an allowed action passes back. Where data is stored and where substantive analysis occurs remain connector-specific questions.
- The agent produces or continues the workflow. Gemini Enterprise can govern the agent and its output without becoming the sole authority for repository access.
NetDocuments provides the clearest implementation detail so far. In NetDocuments’ account of its MCP connection, every interaction runs under the user’s own credentials, existing ethical walls and matter restrictions continue to apply, documents remain in NetDocuments, and access is available to ndMAX Enterprise customers.
That example illustrates why “permission-aware” is more precise than “one permission system.” Gemini Enterprise governs the requesting agent, while NetDocuments governs access to its repository; a complete assessment must cover both layers.
The confirmed connector roster spans the legal technology stack
Google Cloud’s official product account lists Google Workspace, Microsoft 365, iManage, NetDocuments, Docusign, Everlaw, RelativityOne, Thomson Reuters HighQ, Free Law Project’s CourtListener, Courtroom5, Harvey, Solve Intelligence and Legora; it also describes inherited access controls, a governed control plane, preview availability, and private customer data and outputs that are not used to train or fine-tune Google’s foundation models.
The roster covers productivity tools, document management, contract execution, e-discovery, legal research, public court records and specialist legal AI. Inclusion on the list does not establish a uniform release state or technical model: authentication, supported actions, licensing, audit records, regional processing and the location of analysis may differ between integrations.
The product’s reusable skills cover tasks including contract review and redlining, playbook creation, regulatory monitoring, legal research and data-subject access requests. Those skills provide task instructions and organizational context, but they do not independently expand a user’s rights inside a connected repository.
The control plane is intended to give legal IT and risk teams centralized oversight of agents, including platform security and audit controls. It sits above the connectors rather than erasing their boundaries, so buyers still need connector-level documentation to determine whether a connection is read-only, which identity it uses and which system records the resulting activity.
RelativityOne keeps orchestration, data and analysis separate
The RelativityOne implementation makes the boundary map more concrete. In Relativity’s integration announcement, the MCP connection lets administrators create matters, align workspaces and manage access and workflows through natural language, while sensitive data remains in RelativityOne and substantive analysis stays within Relativity aiR.
For that connection, Gemini Enterprise provides the conversational orchestration layer; RelativityOne retains the case information, and Relativity aiR performs the substantive analysis. Describing the arrangement simply as Gemini analyzing all connected legal data would collapse three distinct functions into one.
Other connectors may divide retrieval, action and analysis differently. MCP provides a common way for an agent and an external system to communicate, but the protocol’s presence alone does not prove where every processing step occurs or which party enforces every control.
Preview status limits what buyers can conclude
The launch establishes a product architecture, a connector catalogue and several partner-specific availability statements. It does not establish that every connector is generally available, supports the same operations or has undergone the same production testing.
The privacy and governance assurances are vendor commitments. No independent end-to-end security assessment, permission-leakage test, comparative legal-research benchmark or uniform connector matrix accompanies the launch, and no published evidence yet shows consistent enforcement across the full catalogue under production conditions.
As of August 26, Gemini Enterprise for Legal remains a preview, while individual partners can offer more specific access—for example, the NetDocuments connection for ndMAX Enterprise customers and the active RelativityOne integration. The next material evidence for buyers will be connector-level specifications, broader availability terms and independently demonstrated outcomes showing that inherited permissions remain effective across real deployments.
Also read:
Subscribe to our newsletter
Get the latest Web3, AI, and crypto news delivered straight to your inbox.