SoftTech AI Registry
Microsoft · Microsoft 365

Researcher

A Microsoft-documented first-party reasoning agent in Microsoft 365 Copilot (Microsoft Copilot) that performs complex, multi-step research. The user submits a prompt; Researcher may ask clarifying questions, plans and iteratively researches across the user's work content (files, emails, meetings and chats the user has access to) via Microsoft Graph, Microsoft 365 Copilot/Graph connectors, and — when tenant web search is enabled — the Bing index, then synthesises a structured, source-cited report inside the Copilot experience. Microsoft documents the output as a report the user can then edit, share or carry into Pages/documents; persistence beyond the Copilot conversation is described as a subsequent user action. A separately admin-gated extension, Researcher with Computer Use, adds sandboxed virtual-machine web interaction and is recorded here only as a boundary, not as base Researcher authority.

Recorded characteristics

Function
Microsoft documents Researcher as an agent for deep, multi-step research that deliberately spends longer retrieving and analysing than standard Copilot chat. The documented workflow is: user prompt -> optional clarifying questions from the agent -> retrieval and analysis across the selected scope (work content, the web, or both) -> synthesis -> a structured report containing organised sections, visuals/charts/graphs, key insights, cited sources and suggested next steps. Microsoft states users can then edit the report, share it, or use it to start a presentation or document. In Microsoft 365 Copilot Notebooks, Microsoft documents a user-invoked "Edit in Pages" control that creates a page from the response. Microsoft states Researcher is a first-party agent whose predefined logic and content cannot be modified or extended (for example via Copilot Studio). Microsoft states Deep Research has been retired and Researcher is now the in-depth research experience, with previous Deep Research reports still accessible.
Data access
Microsoft documents that Researcher draws on the web and on the user's work content, specifically naming files, emails, meetings and chats that the user has access to, and states it uses Microsoft Graph, connectors, and the Bing index for recent web data. Microsoft states archived mailbox/folder email is included as backfill when the primary inbox lacks sufficient data, and that Researcher can fetch data from Graph connectors (with connector configuration described as a future update). Microsoft states that at work Researcher "respects the same permissions, policies, and compliance you already rely on" and draws only on content the user has access to. Users can scope a request to workplace resources, the web, or both. Microsoft does not enumerate access to every Microsoft Graph workload beyond the named sources, and the registry does not infer access to calendars, people, transcripts or other workloads that Microsoft does not name for Researcher.
Actions
Generative only
External actions
Unknown
Human confirmation
Not established
Permission basis
User permissions
Administrative control
Microsoft documents that Researcher is part of the core default experience for Microsoft Copilot licensed users and remains accessible in Microsoft Copilot Chat under Tools even when Copilot agents are disabled for some or all users in the Microsoft 365 admin center, while separately stating that admins can block Researcher and Analyst agents and pointing to agent management/agent settings in the Microsoft 365 admin center. Microsoft states users cannot disable, remove or unpin Researcher. A tenant-level web search toggle is the only admin control over web content: Microsoft states there is no granular allow/block list for specific websites in base Researcher, and that if web search is disabled at tenant level Researcher will not use any web data. Microsoft states Researcher adheres to DLP policies and to the same data protection, retention, Responsible AI and Microsoft 365 commercial data boundary commitments as Microsoft Copilot. Anthropic model access in Researcher requires the admin to allow Anthropic AI models in the Microsoft 365 admin center. Researcher with Computer Use has its own admin settings (per-user/group availability, work-content access, and allowed/excluded websites). Microsoft states there is no existing reporting tool or dashboard for admins to track Researcher/Analyst agent usage; no Researcher-specific audit or logging surface is established beyond the general Microsoft Copilot commitments.
Default state
Enabled
Availability
Microsoft documents Researcher as available inside the Microsoft Copilot app under Agents, including the iOS and Android Copilot mobile apps, and as usable inside Microsoft 365 Copilot Notebooks and as @Researcher inside a Copilot chat. Microsoft states it supports over 30 languages, the same as Microsoft Copilot, and that availability in Sovereign Clouds is "to be released soon". Microsoft states Researcher adheres to the European Union and Microsoft 365 data boundaries. Researcher with Computer Use is an admin-gated extension requiring admin allowance before users can select it.
Licensing
Microsoft states Researcher is available to Microsoft 365 Premium and Pro subscribers and to Microsoft 365 business and enterprise users with a Microsoft 365 Copilot add-on license. Microsoft documents a usage limit of a maximum of 25 queries per user per month.
External model or provider
Microsoft's current documentation does not commit to a single fixed model for Researcher. Microsoft states a model picker is available in the top-right of the Researcher agent allowing the user to pick the model and mode, that when Researcher is added to a Copilot chat outside the agent the picker is disabled and defaults to auto, and that Anthropic's Claude models can be used in Researcher when the admin allows Anthropic AI models. Microsoft's March 2025 launch material describes Researcher as combining OpenAI's deep research model with Microsoft 365 Copilot orchestration; the current support documentation for Researcher in Copilot Notebooks still refers to running research "with OpenAI". Microsoft states Researcher does not send user data or file attachments to OpenAI and is not trained on tenant data. Provenance is recorded; no single current model commitment is established.
Limitations and uncertainty
Microsoft does not document Researcher writing into Microsoft 365, creating documents autonomously, sending email, posting Teams messages, changing source data, or modifying external systems; equally Microsoft does not publish an affirmative statement that base Researcher cannot perform external actions, so external action capability is recorded as unknown rather than no. Persistence of output beyond the Copilot conversation is documented only through user-initiated actions (edit, share, "Edit in Pages", use in a document/presentation) and is not attributed to the agent. Connector access is documented as retrieval of external enterprise data only; no connector write authority is documented. Human confirmation is recorded as not_established because Microsoft documents no consequential execution step requiring approval; clarifying questions and the user's own prompt are not confirmation of an action. Microsoft does not document a second runtime identity for Researcher, including for Graph connector access, so permission basis is user_permissions; admin ability to block Researcher is a feature-availability control, not a runtime permission mechanism. Boundary: Researcher with Computer Use is a separately admin-gated extension in which the agent interacts with public, gated and interactive web content through an ephemeral Windows 365 for Agents sandbox with browser/network policy, screenshots and user-in-control steps for authentication/CAPTCHA; Microsoft describes it as enabling users to "take action". Its authority is deliberately NOT merged into this base Researcher record and is not used to classify base action or external action. Other uncertainties: no Researcher usage reporting/dashboard exists; report length/detail control and agent-to-agent conversations are marked "to be released soon"; connector configuration is a future update; Sovereign Cloud availability is pending; because the model was trained on web data up to 2024, responses reflect that period when web search is not used.

Evidence