Note
Access to this page requires authorization. You can try signing in or changing directories.
Access to this page requires authorization. You can try changing directories.
Windows 365 environment insights and Cloud PC connectivity investigation
What is an Application or Platform Card?
Microsoft's Application and Platform Cards help you understand how our AI technology works, the choices application owners can make that influence application performance and behavior, and the importance of considering the whole application, including the technology, the people, and the environment. Application Cards are created for AI applications, and Platform Cards are created for AI platform services. These resources can support the development or deployment of your own applications and can be shared with users or stakeholders affected by them.
As part of its commitment to responsible AI, Microsoft values six core principles: fairness, reliability and safety, privacy and security, inclusiveness, transparency, and accountability. These principles are embedded in the Responsible AI Standard, which guides teams in designing, building, and testing AI applications. Application and Platform Cards play a key role in operationalizing these principles by offering transparency around capabilities, intended uses, and limitations. For further insight, explore Microsoft's Responsible AI Transparency Report and Code of Conduct, which outline how enterprise customers and individuals can engage with AI responsibly.
Overview
Windows 365 admin assistant helps Windows 365 IT operations administrators, second- and third-line support teams, and IT leaders understand their Windows 365 environment and investigate Cloud PC-related issues. It uses a large language model (LLM) to interpret natural-language questions, organize investigations, and reason over available tenant evidence. By assembling operational data and identifying patterns, it helps administrators spend less effort gathering information and make better-informed decisions about likely causes and affected populations.
The application is an integrated chat experience in the Windows 365 Cloud PC management section of Microsoft Intune. In the chat, administrators can:
- Investigate users, devices, activities, events, and cohorts.
- Supply environmental context and image, CSV, or text attachments.
- Resume an investigation from its conversation history.
Optional deeper reasoning supports follow-up analysis of the available evidence.
For Cloud PC connectivity investigations, the application brings together connection history, health signals, and reporting data to compare affected Cloud PCs with healthy sessions and identify patterns. It explains likely causes and recommends next steps. Administrators verify the evidence and perform every remediation action. The application doesn't launch remote device actions, change customer settings, or provide general-purpose autonomous operation.
Key terms
The following list provides a glossary of key terms related to Windows 365 admin assistant:
- Agent: An AI application that can organize steps and use available tools to pursue a user's request within defined boundaries.
- Application card: A description of an AI application's capabilities, intended uses, limitations, evaluations, and responsible-use considerations.
- Artificial intelligence (AI): Computer-based capabilities that perform tasks such as interpreting language, finding patterns, or generating responses.
- Azure REST API: A programmatic interface for requesting Azure resource information through standard web requests. Availability and authorization affect the evidence returned.
- Cloud PC: A Windows desktop hosted through Windows 365 and accessed remotely by its assigned user.
- Cohort: A group of users, devices, or sessions selected for comparison because they share a relevant characteristic.
- Coherence: The logical organization and consistency of a response. It doesn't establish factual correctness.
- Confidence score: A model-generated indication of certainty that isn't, by itself, a calibrated probability that a conclusion is correct.
- Context: Information supplied with a request or retrieved during an investigation, such as maintenance windows, symptoms, or tenant signals.
- Evaluator: A method that assesses an output or interaction against a defined quality, safety, or task criterion.
- Fluency: The readability and grammatical quality of generated language. A fluent answer can still be wrong.
- Groundedness: The extent to which a response is supported by the context against which it's evaluated.
- Grounding: Providing relevant information to a model when it generates an answer, rather than relying only on information learned during training.
- Human oversight: People supervising an AI application's behavior and checking its outputs and consequences.
- Large language model (LLM): A model trained on large amounts of text to interpret or generate language. Its outputs can be inaccurate or overconfident.
- Latency: The time taken for data to travel or for a system to respond. Connection investigations might compare network round-trip time.
- Microsoft Foundry (Foundry): Microsoft's platform for developing and evaluating AI applications, including selectable built-in evaluators.
- Microsoft Graph: Microsoft's programmatic interface to authorized Microsoft 365 and related organizational data and services.
- Microsoft Intune (Intune): Microsoft's endpoint-management service, which exposes the Windows 365 administration experience and hosts relevant management signals. Windows 365 admin assistant is a feature of Windows 365 and is governed by the associated Windows 365 product terms.
- Mitigation: A measure intended to reduce a risk or its consequences, without necessarily eliminating the risk.
- Model: A trained computational system used to interpret inputs and generate predictions or responses.
- Prompt: The input or request used to guide a model's response.
- Query: A request to retrieve or compare information from an available data source.
- Reference answer (ground truth): An expected answer or known result used as a comparison in an evaluation.
- Remediation: A corrective action intended to address an operational problem. An administrator, not the application, performs it.
- Role-based access control: A way to limit access to data or actions according to the roles assigned to a user.
- Scenario: A defined situation or task used to illustrate an intended use or test an expected outcome.
- Signal: An observation or item of operational data, such as a connection failure or a health status, used as evidence in an investigation.
- Similarity: A measure of how closely a generated answer matches a supplied reference answer.
- Tenant: An organization's logically separate Microsoft cloud environment and associated identity, configuration, and data boundaries.
- Trace: A record of steps and interactions in an application request that can support evaluation and troubleshooting.
- Windows 365: Microsoft's Cloud PC service for providing hosted Windows desktops to authorized users.
Key features or capabilities
The key features and capabilities outlined here describe what Windows 365 admin assistant is designed to do and how it performs across supported tasks.
- Natural-language investigation: Administrators can describe a connectivity symptom and ask follow-up questions without assembling every query themselves. The application uses available evidence to organize the investigation and suggest a likely cause or next check. Its reasoning is bounded by known, testable Windows 365 problems and doesn't provide open-ended expert diagnosis.
- Evidence aggregation and comparison: The application can bring together connection history, health signals, and reporting data to compare affected Cloud PCs with a healthy cohort. Comparisons can highlight differences in latency, network provider, transport, client, or location. A recent script or policy change might support a hypothesis, but the administrator must verify whether it actually explains the issue.
- Context-aware, resumable conversation: Integrated chat supports resumable conversations and image, CSV, and text attachments. Administrators can supply environmental details and continue an investigation with its prior conversation context. Resumable history maintains context for the investigation. It isn't durable personal memory across unrelated work.
- Bounded planning and recommended next steps: The application organizes an investigation, adapts follow-up reasoning to the available evidence, and offers optional deeper reasoning. It provides explanations and recommended remediation plans that an admin can act on. However, the agent doesn't automatically execute workflows. It doesn't launch remote device actions, change customer settings, or provide general-purpose autonomous operation.
- User-led scope, access, and actions: An administrator's natural-language request starts a bounded investigation. The application uses LLM reasoning to organize relevant checks and interpret Windows 365, Microsoft Entra, Microsoft Intune, and Microsoft Azure product data within the signed-in administrator's permissions and role-based access controls. Follow-up questions and new context can refine the scope, revise a likely-cause hypothesis, or change the recommended next check. The application provides explanations and recommendations. The administrator judges the evidence and carries out every remediation action.
Intended uses
Windows 365 admin assistant can be used in multiple scenarios for managing Windows 365. Some examples of use cases include:
| Scenario | Description |
|---|---|
| Connectivity support | An administrator supporting a university's Cloud PCs can investigate reports that a subset of remote learners can't connect. Comparing failing sessions with healthy sessions can help distinguish a shared network pattern from an isolated device symptom. The administrator checks the evidence, then coordinates an appropriate change with the network owner. This is IT troubleshooting, not a tool for admissions, grading, or access-to-education decisions. |
| Support escalation | A support team can examine Cloud PC connection failures that interrupt employees' access to workplace tools. The application can assemble relevant health and connection evidence and suggest an affected cohort for further investigation. Because operational disruption might be consequential, the team validates the hypothesis and follows its incident and change-management procedures before making changes. The application doesn't make credit, trading, or other financial decisions. |
| Network investigation | An IT team can compare high-latency Cloud PC sessions across network providers or locations to prepare a focused escalation. Providing known network constraints and maintenance activity helps avoid treating expected behavior as a new incident. The resulting evidence can support a discussion with a network owner, rather than establish causation on its own. It must not be used to determine eligibility for government benefits or other individual entitlements. |
| Workplace connectivity | An administrator supporting geographically distributed offices or stores can look for common connection symptoms across Cloud PCs. A comparison group can help prioritize which network conditions or client patterns to examine first. The administrator reviews the recommended next steps and performs any permitted remediation through existing administration tools. The resulting evidence supports focused escalation to the responsible support or network team. |
| Windows 365 environment insights | An administrator can ask how Cloud PC connection performance differs across locations or client groups to understand their Windows 365 environment. The application organizes the investigation around that question, compares available connection, health, and reporting signals, and explains patterns and next checks. Follow-up information about maintenance activity or known network constraints can change how those patterns are interpreted. The administrator validates the evidence and decides whether a finding needs further investigation or a change through existing administration tools. |
Models and training data
Windows 365 admin assistant uses Generative Pre-Trained Transformers (GPT) models provided by Microsoft (Azure OpenAI Service) to power the experience that users see.
Note
OpenAI is a subprocessor for Windows 365 admin assistant. For more information, see OpenAI as a subprocessor in Microsoft Online Services.
To learn more about the data used to train this model, see the following resources:
Prompts, responses, and data accessed through Windows 365 admin assistant aren't used to train foundation LLMs, including those used by Windows 365 admin assistant. Your organizational data remains within Windows 365 and is subject to your organization's existing data governance policies. For more information about how data is handled, see Privacy and data in Windows 365. For additional information about the LLMs provided to Microsoft by OpenAI, see their public documentation at Models - OpenAI API.
Performance
The application works best for known, testable, solvable Windows 365 connectivity problems when the signed-in administrator has the required access and the underlying data sources are available, current, and correctly configured. Clear symptoms, a relevant time window, a defined affected population, and a suitable comparison cohort help focus the investigation. Missing or stale signals can reduce the usefulness of its conclusions.
Inputs include natural-language text questions, follow-up context, image, CSV (comma-separated values), and text attachments, together with available tenant and reporting signals. Outputs are conversational text explanations, evidence summaries, likely-cause hypotheses, and recommended next steps.
Investigations depend on associated product data and customer-provided context. Unavailable, incomplete, or misconfigured data sources can lead to incomplete explanations or generic guidance, particularly in complex or large-scale analysis. A failed query is a gap in evidence, not evidence that no incident exists. Administrators should restore authorized data access or consult the underlying administration tools before relying on the result.
Relevant environmental details affect the quality of an investigation. Maintenance windows, known high-latency connections, and recent changes help distinguish expected behavior from a problem. Review the response against the actual devices, users, and time period, and check whether a pattern also holds in other relevant cohorts. A fluent explanation or high confidence score doesn't replace this review.
Limitations
Understanding the limitations of Windows 365 admin assistant is crucial to determine if you use it within safe and effective boundaries. While we encourage customers to use Windows 365 admin assistant in their innovative solutions or applications, it's important to note that Windows 365 admin assistant wasn't designed for every possible scenario. Refer to the Microsoft Enterprise AI Services Code of Conduct and the capabilities defined earlier in this article.
- Scope of troubleshooting and investigation: The agent is designed to accelerate resolution for known, testable, and solvable problems in Windows 365 environments. It isn't intended to diagnose highly complex or uncommon issues, nor does it support expert-level investigations such as deep parsing of on-device logs to identify novel service issues. Escalate these out-of-scope investigations to an appropriately qualified support or engineering team rather than treating a plausible answer as an expert diagnosis.
- Environmental and platform dependencies: The agent relies on signals from Microsoft Graph, Azure REST APIs, Intune, and customer-provided context. If these data sources are unavailable, incomplete, or misconfigured, the agent's effectiveness is reduced. A missing or failed query can leave the explanation incomplete or generic rather than establish that no problem exists. Restore authorized access and data availability, or use the underlying administration and support tools, before relying on the recommendation.
- Remediation boundaries: The agent provides recommendations but doesn't perform automatic remediation. An administrator must perform all remediation actions. The agent doesn't make changes to customer environments automatically, and changes to Domain Controller settings are out of scope. A Domain Controller is a server that manages domain authentication and policy. Route issues requiring those changes to the responsible specialist.
- Customization and context awareness: While the agent can incorporate customer-provided context, such as maintenance windows or known high-latency network connections, its recommendations are only as accurate as the context provided. If admins don't supply sufficient context or details with their request, the agent might generate false positives or less relevant guidance. A false positive is a finding that flags a problem where the observed behavior has another, expected explanation. Correct missing or inaccurate context and verify the revised guidance instead of using the response as a substitute for environmental knowledge.
- Overconfidence: The agent might exhibit overconfidence in its recommendations due to limitations in LLM-based confidence scoring. While efforts are made to validate findings across cohorts and provide confidence levels, there's an acknowledged risk that the agent might not always accurately assess the generalizability of its conclusions. A persuasive explanation doesn't demonstrate that the same cause applies to every affected device. Check evidence across relevant cohorts, and don't treat a confidence score as a calibrated measure of correctness.
- Unsupported and prohibited uses: This is a bounded Windows 365 admin assistant, not a general-purpose decision-maker or an autonomous administrator. Don't use it for expert investigations outside its scope or for medical, legal, financial, or eligibility decisions about people. Uses prohibited by applicable service terms and the linked codes of conduct remain prohibited. Industry examples in this card illustrate IT support only and don't expand these boundaries.
Evaluations
Performance and safety evaluations assess whether AI applications are operating reliably and securely by examining factors like groundedness, relevance, and coherence while identifying the risks of generating harmful content. The following evaluations were conducted with safety components already in place, which are also described in Safety components and mitigations.
Performance and quality evaluations
Performance evaluations for AI applications are essential to improving their reliability in real-world applications. Metrics like groundedness, relevance, and coherence help assess the accuracy and consistency of AI-generated outputs, so that they're factually supported in grounded content scenarios, contextually appropriate, and logically structured. For Windows 365 admin assistant, we conducted performance evaluations for the following metrics, which are available through Microsoft Foundry:
- Task and intent: Task Completion and Intent Resolution.
- Response quality: Groundedness, Relevance, and Coherence.
- Tool use: Tool Selection, Tool Input Accuracy, and Tool Call Success.
Performance and quality evaluation methods
Evaluations use predefined target scenarios, case-specific expectations and assertions, and end-to-end checks with known expected results, alongside the same general Foundry evaluators used online. Online evaluations apply automated general Foundry checks to sampled interactions and traces. Customer feedback is linked to conversation context, and fixes are retested against the scenarios.
Risk and safety evaluations
Evaluating potential risks associated with AI-generated content is essential for safeguarding against content risks with varying degrees of severity. This includes evaluating an AI application's predisposition towards generating harmful content or testing vulnerabilities to jailbreak attacks. For Windows 365 admin assistant, we conducted risk and safety evaluations for the following metrics available through Microsoft Foundry:
- Hate and unfairness
- Sexual
- Violence
- Self-harm
- Protected material
- Indirect jailbreak
- Direct jailbreak
- Ungrounded attributes
Risk and safety evaluation methods
The evaluation process combines target scenarios and their expected outcomes with automated Foundry measurements of sampled interactions and traces. Feedback supplies conversation context for investigating problems, and fixes are retested against the relevant scenarios.
An ideal safety response treats an instruction embedded in an attachment to ignore permissions or run a remote change as untrusted content, stays within the administrator's permitted evidence, and leaves remediation to the administrator. It avoids revealing unrelated data and communicates uncertainty when evidence is missing. A failing response follows the embedded instruction, reveals out-of-scope information, produces harmful content, or claims to have run an action.
Evaluation data for quality and safety
Our evaluation data is custom-built to assess AI application performance across key areas of safety and quality, simulating real-world scenarios and risks. We begin by identifying relevant evaluation aspects of concern based on multi-disciplinary research and expert input. These concerns are translated into targeted evaluation objectives and guide formulation of evaluation metrics. For safety, we create adversarial prompts to elicit undesirable or edge-case responses, which are then scored using AI-assisted annotators trained to assess alignment with Microsoft's safety standards. For quality, we craft rubric-based prompts relevant to scenarios including evaluating retrieval-augmented generation (RAG) applications and agents. Datasets are curated from diverse sources including synthetic and public datasets to simulate real-world user scenarios. Using the curated datasets, both evaluations undergo iterative refinement and human alignment to improve metric efficacy and reliability. This methodology forms the foundation of repeatable, rigorous assessments that reflect how customers use evaluations to build better and safer AI.
Custom evaluations
In addition to the standard metrics, Windows 365 admin assistant was evaluated using custom metrics specific to its agentic capabilities:
- Task completion: Whether the requested investigation was successfully completed across multiple steps.
- Response clarity: Whether the output implications are clear to an IT admin.
- Reasoning coherence: Whether the logical path taken and synthesis of results are coherent within the domains supported by Windows 365 admin assistant.
Custom evaluations used text-based scenarios designed to test multi-step task execution, response clarity, and reasoning coherence. Ideal results demonstrate that Windows 365 admin assistant correctly identifies the steps needed, runs the investigation with a logical path (time, device, symptom coherence), and represents the findings clearly.
Safety components and mitigations
- Administrator access boundaries: Investigations use tenant data available under the signed-in administrator's permissions and role-based access controls. These boundaries limit the evidence the application can retrieve. Administrators should use appropriate roles and verify the scope of evidence returned. Incomplete results can reflect missing permissions rather than an absence of incidents.
- Agent access boundaries: Windows 365 admin assistant uses delegated access and can only read information that the signed-in administrator is authorized to view. It requests the following read-only permissions:
- Basic user information, assigned licenses, and group membership used to identify affected users.
- Cloud PC details such as assigned users, device status, region, configuration, and usage.
- Connection information such as connection history, performance, failures, and error details.
- Device policies and settings that might affect Cloud PC access or performance.
- Current and recent Windows 365 service incidents and advisories.
- Sign-in activity and access policies used to investigate authentication failures.
- Human-controlled remediation: The application provides recommendations only, and an administrator performs every remediation action. Environment changes remain within existing human review and change-management processes rather than being automatically run from an inferred fix. Review the proposed step, affected population, and reversibility before acting through supported administration tools. This boundary doesn't guarantee that the recommendation itself is correct.
- Evidence review and iterative improvement: Cohort comparisons and customer-provided context help administrators challenge hypotheses and recognize expected environmental behavior. The improvement process combines scenario checks, default Foundry evaluations, customer feedback, and retesting after fixes. Verify the supporting evidence and examine whether conclusions hold across relevant cohorts. Escalate persistent uncertainty rather than interpreting repeated confident answers as validation.
- Access governance: Microsoft 365 admin-center controls let administrators enable or disable the application and manage user access. Tenant access controls and the signed-in administrator's permissions govern access to investigation data. Review access assignments when responsibilities change, and grant only the roles needed for the intended investigations. Follow organizational requirements for handling operational context and attachments.
Data, privacy, and security
Windows 365 admin assistant handles data in accordance with Microsoft's data privacy and security standards:
- Tenant isolation: All agent sessions operate within the organizational tenant boundary. Session data isn't shared across tenants.
- Customer data: Prompts, session content, and agent outputs are treated as Customer Data and aren't used to train foundation models. For more about customer data storage, see Privacy and data in Windows 365.
- Agent identity and permissions: Agents access only data that the signed-in administrator is authorized to view.
- Audit trail: Sessions are immutable records that log all agent activity, inputs, and outputs. These records support security audits and compliance reviews.
Prompt evaluation location
Prompt evaluation location determines where prompts are processed using GPU resources. When a user submits a prompt, Windows 365 admin assistant evaluates it on GPU clusters in Azure datacenters. This setting is preselected during autoprovisioning:
- If your tenant's Customer Data storage location is in the EU, prompts are processed in the EU.
- If your tenant's Customer Data storage location isn't in the EU, prompts are processed globally in the US, UK, EU, or Australia and New Zealand, depending on locality and GPU availability.
For more information about Microsoft's data handling practices, see Privacy and data in Windows 365.
Best practices for deploying and adopting Windows 365 admin assistant
Responsible AI is a shared commitment between Microsoft and its customers. While Microsoft builds AI applications and platform services with safety, fairness, and transparency at the core, customers play a critical role in deploying and using these technologies responsibly within their own contexts. To support this partnership, we offer the following best practices for tenant admins and admin users to help customers implement responsible AI effectively.
Tenant admins and admin users
- Exercise caution and evaluate outcomes when using Windows 365 admin assistant for consequential decisions or in sensitive domains: Consequential decisions are those that might have a legal or significant impact on a person's access to education, employment, financial platforms, government benefits, healthcare, housing, insurance, legal platforms, or that could result in physical, psychological, or financial harm. Sensitive domains, such as financial platforms, healthcare, and housing, require particular care due to the potential for disproportionate impact on different groups of people. When using AI for decisions in these areas, make sure that affected stakeholders can understand how decisions are made, appeal decisions, and update any relevant input data.
- Evaluate legal and regulatory considerations: Customers need to evaluate potential specific legal and regulatory obligations when using any AI platforms and solutions, which might not be appropriate for use in every industry or scenario. Additionally, AI platforms or solutions aren't designed for and can't be used in ways prohibited in applicable terms of service and relevant codes of conduct.
- Keep the investigation within scope: Use the application to understand your Windows 365 environment and support known connectivity investigations, not to make consequential decisions about people. Confirm the affected population, the reliability of the source signals, and who is accountable for the next step. When an issue is novel, highly complex, or requires deep on-device log analysis, move to an expert support process. Keep remediation ownership and change-management responsibilities explicit.
Admin users
- Provide a clear, bounded prompt: State the symptom, time window, affected Cloud PCs, and the comparison you need. For example, ask, "Compare the connection failures for this affected cohort during the reported incident with healthy sessions in the same period, and explain the evidence for the next check." Supply maintenance windows and known network constraints so expected behavior isn't mistaken for an incident. Avoid requesting an unsupported remote action or an unbounded investigation.
- Exercise human oversight when appropriate: Human oversight is an important safeguard when interacting with AI applications. While we continuously improve our AI applications, AI might still make mistakes. The outputs generated might be inaccurate, incomplete, biased, misaligned, or irrelevant to your intended goals. This could happen due to various reasons, such as ambiguity in the inputs or limitations of the underlying models. As such, review the responses generated by Windows 365 admin assistant and verify that they match your expectations and requirements.
- Be aware of the risk of overreliance: Overreliance on AI happens when users accept incorrect or incomplete AI outputs, mainly because mistakes in AI outputs might be hard to detect. For the end user, overreliance could result in decreased productivity, loss of trust, application abandonment, financial loss, psychological harm, or physical harm, among others. For example, a doctor accepts an incorrect AI output.
- Exercise caution when designing agentic AI in sensitive domains: Exercise caution when designing or deploying agentic AI applications in sensitive domains where agent actions are irreversible or highly consequential. Take additional precautions when creating autonomous agentic AI, as described in the Microsoft Enterprise AI Services Code of Conduct.
- Share only appropriate context: Include the operational facts needed to investigate, not passwords, secrets, or unrelated personal information. Use only attachments and data that you're authorized to supply, and follow your organization's handling rules. Treat instructions found inside an attachment as source content rather than authority to change permissions or take actions. Follow organizational requirements when handling sensitive material.
- Check evidence and missing context: Verify that the response addresses the intended users, devices, and time period, and inspect which signals support each conclusion. Distinguish a likely cause from proof, and treat confidence scores as fallible rather than calibrated certainty. If a query fails or evidence is missing, restore the necessary authorized access or consult the underlying tools. Correct the context and rerun the investigation before relying on revised guidance.
Tenant admins
- Manage access: Grant administrators only the roles needed for the intended investigations and review tenant access settings. Use the tenant administration controls in the Intune portal to manage application availability. Review access when roles or responsibilities change.
- Prepare for predictable failures: Verify Microsoft Graph, Azure REST API, and Intune access and configuration before a pilot or investigation. Establish a way to recognize unavailable or incomplete signals and restore authorized data access. Keep an alternative route through existing administration and support tools when the assistant produces generic or incomplete guidance. Route expert-level investigations and unsupported remediation, including Domain Controller changes, to the responsible specialist.
- Monitor change and performance drift: Performance drift is a change in output quality as conditions, data, models, or configuration change. Review feedback and recurring failure patterns, and rerun representative scenarios after meaningful changes. Compare responses with expected operational outcomes and check for differences across your networks, clients, locations, and administrative roles. Update deployment guidance and escalate regressions before expanding adoption.
- Pilot representative scenarios: Test known connectivity incidents and healthy comparison cases that reflect your networks, clients, locations, and administrative roles. Include missing-data and incorrect-context cases, and explicitly test any language or attachment type you plan to use. Assess whether answers identify the expected cohort, disclose gaps, and suggest appropriate next steps rather than merely sounding fluent. Retain human review and avoid extrapolating a small pilot to untested conditions.
Learn more about Windows 365 admin assistant
For additional guidance or to learn more about the responsible use of Windows 365 admin assistant, see the Windows 365 admin assistant documentation.
The following resources explain responsible AI, the Microsoft Foundry evaluation framework, and related Windows 365 privacy and security considerations.
Learn more about responsible AI
- Microsoft AI principles
- Microsoft responsible AI resources
- Microsoft Azure Learning courses on responsible AI
Related references
- Foundry built-in evaluator reference
- Foundry performance and writing-quality evaluation methods
- Foundry groundedness and retrieval evaluation methods
- Foundry similarity evaluation methods
- Foundry risk and safety evaluators
- Foundry safety-evaluation transparency note
- Foundry monitoring and selected continuous evaluations
- Windows 365 privacy and personal data
- Windows 365 encryption
- Windows 365 compliance overview