Data, privacy, and security for Interactive insights (Frontier)

Important

This feature is part of the Frontier preview program. Frontier connects you directly with Microsoft's latest AI innovations. Frontier previews are subject to the existing preview terms of your customer agreements. As these features are still in development, their availability and capabilities might change over time.

Interactive insights helps authorized users investigate how AI is used across their organization, interpret patterns, and turn evidence into decision-ready answers. With it, you can access insights about Microsoft 365 Copilot, agents, work performed with AI, AI consumption, and other organizational AI signals.

Unlike a dashboard that retrieves predefined views like the Copilot Dashboard, Interactive insights can plan and perform custom analysis against data available to the user.

Interactive insights operates within the broader Microsoft Copilot Analytics, Viva Insights, and Microsoft 365 governance, privacy, security, access-control, and compliance framework. This article covers the AI capability itself. For Copilot in Viva Insights — the query-assistance feature — see Data, privacy, and security for Microsoft Copilot in Viva Insights.

In this article, "Interactive insights" refers to three experiences that use the same Viva Insights data, metric definitions, permissions, and privacy controls:

  • Interactive insights in Viva Insights (the in-app experience).

  • The Microsoft Agent 365 Business Analytics plugin for Copilot Cowork (A365 Analytics plugin).

  • The Microsoft Agent 365 Business Analytics MCP (A365 Analytics MCP).

Unless stated otherwise, statements in this article apply to all three experiences. The A365 Analytics MCP can be used in supported Microsoft clients, such as GitHub Copilot CLI and Visual Studio Code.

How does it work on the backend?

A typical analysis follows four steps.

  1. Interactive insights identifies the user's question, relevant information, intended outcome, and applicable constraints.

  2. It selects an approach and uses available skills and tools to analyze the information.

  3. It combines findings, explanations, caveats, recommendations, and visualizations as appropriate to the request.

  4. It returns the output for review, and the user can ask follow-up questions, refine the scope, or edit supported outputs before sharing them.

Complex requests can take longer than a typical chat response because the agent performs custom analysis rather than retrieving only a predefined answer.

What data does Interactive insights use to generate its responses?

Interactive insights works only with data and context made available through the product or provided by the user. Access to product data follows applicable permissions, organizational settings, and governance controls. User-provided data is available only when the capability is enabled and permitted by the organization.

Does Microsoft use my organization's data to train its models?

No. Interactive insights does not train or fine-tune foundation models on a customer's organizational data, prompts, user-provided content, or generated outputs. These inputs are used to operate the product and produce the requested analysis according to applicable Microsoft product terms, privacy commitments, and organizational controls. The model also doesn't determine which organizational data a user can access — data availability is established by the product, organizational configuration, user authorization, and applicable policies. The in-app experience and the A365 Analytics plugin use Microsoft-hosted AI models through Microsoft 365 Copilot infrastructure, and the A365 Analytics MCP uses the AI model provided by the supported client; models can change as Microsoft evaluates quality, safety, reliability, and other measures.

How does Interactive insights interact with Microsoft Graph and other Microsoft 365 services?

The in-app experience and the A365 Analytics plugin use Microsoft-hosted AI models through Microsoft 365 Copilot infrastructure and rely on Microsoft 365 Copilot content-safety systems. The A365 Analytics MCP uses the AI model provided by the supported client, along with its own safety filters and the client's content-safety systems. Interactive insights' autonomy is limited to analysis and content generation within the user's authorized, sandboxed product environment.

Interactive insights does not:

  • Browse the open web.
  • Operate the user's device.
  • Call arbitrary external services.
  • Write back to user assets, identity, licensing, or other source systems.
  • Take automated employment or organizational actions.
  • Expose a public API for unattended use of its outputs.

The A365 Analytics plugin and MCP provide data and analysis to the client they're used in (Copilot Cowork for the A365 Analytics plugin, or a supported client for the A365 Analytics MCP) and don't take actions on their own. These clients can take other actions using the returned Viva Insights data only when the user asks, such as sending email, creating presentations, and more. The client governs those actions, not Interactive insights. Responses returned to a client are stored and handled according to that client's terms, retention, and compliance settings. Users can also combine these responses with content from other sources in the client. Interactive insights applies its privacy controls to the data it returns but doesn't control how the client combines it.

What is Interactive insights designed to do?

Interactive insights is designed to:

  • Answer focused questions about supported organizational AI data.
  • Plan and perform custom analysis for broader transformation questions.
  • Compare permitted organizational groups and trends.
  • Explain supported metrics, concepts, and product information.
  • Generate findings, caveats, interpretations, and recommendations.
  • Create visualizations and shareable outputs where those capabilities are available.
  • Refine an analysis through conversation and available workspace context.
  • Use additional context supplied by the user when the capability is enabled and permitted by the organization.

The exact scenarios, data sources, skills, tools, and output types available to a user depend on product availability, organizational configuration, permissions, policies, and data readiness. Interactive insights can recommend analytical or program-level next steps, but those recommendations are advisory and should be reviewed in light of organizational context, applicable law, company policy, and other evidence.

What is Interactive insights not designed to do?

Interactive insights is not designed or evaluated to:

  • Evaluate or rank an individual's performance, productivity, suitability, or value.

  • Recommend or make decisions about hiring, termination, promotion, compensation, discipline, or performance management.

  • Monitor or score individual employees.

  • Infer sensitive personal attributes or conditions that are not explicitly present as governed data.

  • Reconstruct an identity mapping between identified and de-identified data.

  • Provide legal, medical, mental health, or individualized human resources advice.

  • Establish causation solely from observed correlations.

  • Produce answers requiring data that's unavailable or outside the user's authorized scope.

  • Serve as the sole basis for a decision with legal or similarly significant effects on a person.

Interactive insights is designed to refuse prohibited requests. However, customers remain responsible for how they interpret, share, and use generated outputs. It's designed to support human decision-making and doesn't make employment decisions or take automated downstream actions.

What are the known limitations of Interactive insights?

Interactive insights has the following known limitations:

  • It can select an inappropriate measure, population, comparison, or analytical approach.

  • It can misinterpret data, overstate a relationship, or produce a recommendation that's not adequately supported.

  • Results depend on the coverage, quality, freshness, and configuration of available data.

  • Activity and consumption signals don't, on their own, establish employee performance, productivity, business value, or causation.

  • Generated findings, explanations, recommendations, visualizations, and outputs can be inaccurate, incomplete, inconsistent, or misleading.

  • The model can produce confident language even when evidence is weak.

  • Repeated requests can produce different wording, emphasis, or analytical choices.

  • Complex analyses can encounter context, tool, computation, or output-generation limits and can return partial results.

  • Available workspace context and generated outputs are bounded by product availability, configuration, and retention behavior.

  • Interactive insights can't account for all legal, cultural, operational, or business context known to the customer.

  • A polished output isn't proof that the underlying analysis is correct.

Generated outputs require human review before consequential use or broad distribution.

When does Interactive insights perform best?

Interactive insights performs best when:

  • The request concerns a supported organizational AI scenario.

  • Required data is available, current, and sufficiently complete.

  • You provide clear context for broad or strategic questions.

  • Comparisons use sufficiently large and meaningful groups.

  • The data you provide has clear definitions and appropriate coverage.

  • You review important assumptions, definitions, populations, time periods, and caveats.

  • High-stakes or bespoke analysis receives additional review from an appropriate expert.

How does Microsoft evaluate Interactive insights?

Microsoft regularly performs evaluations across response accuracy, quality, and consistency. Evaluation methods can include structured test sets, repeated-run comparisons, automated evaluation, output review, and human review. Microsoft also evaluates responsible AI and safety risks relevant to the experience, including harmful content, sensitive inferences, and inappropriate employment use. Adversarial testing and ongoing product feedback help identify failures and prioritize improvements. Evaluation is ongoing because models, skills, supported data, and product capabilities can change.

What safeguards does Interactive insights provide?

Interactive insights uses various safeguards such as:

  • Permission and policy boundaries - The product limits data and capabilities according to user authorization and organizational configuration.

  • Grounding and analytical controls - The agent is instructed to use available data and approved product information, disclose material uncertainty, and avoid unsupported claims.

  • Responsible AI instructions - Product-specific rules restrict inappropriate employment uses, sensitive inference, re-identification, harmful content, and attempts to override system safeguards.

  • Content-safety protections - Microsoft 365 Copilot safety systems help detect and block harmful content and attempts to bypass safeguards.

  • Sandboxed operation - Analytical execution occurs within a restricted product environment and doesn't provide unrestricted access to external systems.

  • Human oversight - Outputs are returned to the user for review and aren't automatically applied to downstream organizational systems.

  • Transparency - The experience identifies itself as an AI agent and informs users that AI-generated content can be incorrect.

  • Feedback and monitoring - In-product feedback, evaluation, monitoring, and incident-response processes support investigation and improvement.

  • Administrative controls - Organizations can govern access and available capabilities through applicable Microsoft 365 and Viva Insights controls.

No single mitigation guarantees that the agent will always behave correctly. Safety depends on the combined operation of product architecture, access controls, data governance, model and content-safety systems, product instructions, evaluation, monitoring, and human oversight.

How does Interactive insights handle untrusted content and prompts?

User-provided and retrieved content can be incomplete, inaccurate, or malicious. Interactive insights treats that content as information to analyze, not as a source of new system instructions. Microsoft evaluates this content as part of its ongoing responsible AI and safety evaluation.

What should organizations and administrators do to support responsible use?

Organizations and administrators can:

  • Review applicable Viva Insights, Microsoft 365 Copilot, data, access, and compliance settings.

  • Enable Interactive insights and optional capabilities only for users with an appropriate business need and authorization.

  • Maintain accurate organizational data and understand how incomplete or stale data can affect analysis.

  • Apply appropriate retention, audit, and compliance policies.

  • Educate users about intended uses, prohibited employment uses, human-review expectations, and internal approval requirements.

  • Investigate reported misuse and adjust access when necessary.

What should users do to support responsible use?

Users can:

  • Use Interactive insights for supported organizational analysis, not individual employee evaluation.

  • Verify important definitions, populations, time periods, comparisons, and assumptions.

  • Treat recommendations as proposals for human consideration, not automated decisions.

  • Review generated content before sharing or acting on it.

  • Provide only data that's necessary, authorized, and appropriate for the analysis.

  • Avoid attempting to re-identify people in de-identified data.

  • Seek appropriate expert review when the intended use is sensitive or consequential.

  • Report outputs that appear unsafe, incorrect, biased, or inconsistent with the available evidence.

How do I give feedback or report a problem?

You can provide in-product feedback on Interactive insights responses. Feedback and reported issues can be used to investigate defects and improve the product. Microsoft takes potential safety, privacy, security, or compliance issues seriously. Depending on severity, Microsoft can investigate and pause or roll back the feature while an issue is addressed.

How does Interactive insights handle data residency, data subject rights, and privacy regulations?

Interactive insights operates within the broader Microsoft Viva Insights and Microsoft 365 privacy, security, and compliance framework. Learn more.

When Interactive insights is used through the Business Analytics plugin or MCP, responses returned to the client are also subject to that client's data residency, data subject rights, and privacy commitments.

How to use Interactive insights