A is incorrect: The Copilot Studio Kit provides automated enrichment for Topic Match evaluation. Manual review defeats the purpose of automated testing at scale and does not resolve the pending status that specifically requires the Dataverse enrichment step to complete. Reference: https://learn.microsoft.com/en-us/microsoft-copilot-studio/guidance/kit-analyze-results

B is incorrect: Changing the test type abandons the goal of validating correct topic triggering. The correct approach is to complete the Dataverse enrichment step that provides the triggered topic information needed for Topic Match evaluation. Reference: https://learn.microsoft.com/en-us/microsoft-copilot-studio/guidance/kit-configure-tests

C is incorrect: Disabling Application Insights does not change how Topic Match tests work. These tests require post-run enrichment from Dataverse conversation transcripts to determine the actual triggered topic name, regardless of Application Insights configuration. Reference: https://learn.microsoft.com/en-us/microsoft-copilot-studio/guidance/kit-test-capabilities

D is correct: Topic Match tests in the Copilot Studio Kit need Dataverse enrichment to fully run, as only this post-testing enrichment step provides information on the topic name that was triggered. The enrichment analyzes conversation transcript records including the exact triggered topic name and intent recognition scores. Reference: https://learn.microsoft.com/en-us/microsoft-copilot-studio/guidance/kit-analyze-results

A is incorrect: Upgrading all environments simultaneously is wasteful if only some teams need AI Builder capabilities. Each upgrade consumes tenant Dataverse capacity and may require premium licensing. The recommended approach is to evaluate each upgrade request individually based on business justification. Reference: https://learn.microsoft.com/en-us/power-platform/admin/about-teams-environment

B is correct: Microsoft recommends reactive governance where central IT requires business justification for Dataverse for Teams environments and reviews upgrade requests. Tenant admins should apply data loss prevention policies and monitor capacity usage through the Power Platform admin center to ensure effective resource utilization. Reference: https://learn.microsoft.com/en-us/power-platform/guidance/adoption/teams-environment-strategy

C is incorrect: Allowing unrestricted self-service upgrades without governance review risks uncontrolled capacity consumption, licensing cost overruns, and inconsistent security configurations. Tenant admins should maintain oversight through business justification review and DLP policy enforcement. Reference: https://learn.microsoft.com/en-us/power-platform/guidance/adoption/teams-environment-strategy

D is incorrect: A single centralized environment limits team autonomy, creates cross-team security boundary conflicts, and contradicts the Teams-based collaboration model. The recommended approach is selective upgrades with governance controls and business justification review per environment. Reference: https://learn.microsoft.com/en-us/power-platform/guidance/adoption/teams-environment-strategy

A is incorrect: Content harm category filters detect offensive content across four categories but do not identify adversarial instructions embedded in documents designed to manipulate agent behavior. Prompt injection attacks are a distinct threat category requiring Prompt Shields. Reference: https://learn.microsoft.com/en-us/azure/ai-foundry/foundry-models/concepts/content-filter

B is incorrect: Groundedness detection identifies hallucinated or ungrounded content in model outputs by comparing responses against source materials. It addresses output accuracy rather than adversarial instructions hidden in input documents. Reference: https://learn.microsoft.com/en-us/azure/ai-services/content-safety/concepts/groundedness

C is correct: Prompt Shields for Documents specifically safeguards against indirect prompt injection attacks where adversarial instructions are embedded in external documents, emails, or web content. This feature analyzes documents for hidden malicious instructions before they can influence the LLM session. Reference: https://learn.microsoft.com/en-us/azure/ai-services/content-safety/concepts/jailbreak-detection

D is incorrect: Protected material detection identifies known copyrighted text and code in model outputs to prevent unauthorized reproduction. It does not detect adversarial instructions embedded in input documents that are designed to hijack the LLM session. Reference: https://learn.microsoft.com/en-us/azure/ai-foundry/foundry-models/concepts/content-filter

A is incorrect: Relying on a cloud-hosted large model contradicts the intermittent connectivity constraint and introduces unacceptable latency for in-store interactions. Edge deployment with a small language model addresses both connectivity and latency requirements. Reference: https://learn.microsoft.com/en-us/azure/architecture/ai-ml/guide/choose-ai-model

B is incorrect: Building a custom LLM from scratch is extremely resource-intensive and would not fit on kiosk hardware with limited resources. Small language models are specifically designed for constrained deployment scenarios requiring compact model sizes. Reference: https://learn.microsoft.com/en-us/azure/well-architected/ai/get-started

C is correct: Small language models are designed for resource-constrained environments, on-device deployment, and low-latency scenarios. Phi models can be fine-tuned with domain data and deployed at the edge for offline-capable, fast responses on limited hardware. Reference: https://learn.microsoft.com/en-us/azure/cloud-adoption-framework/scenarios/ai/platform/resource-selection

D is incorrect: Relying on cloud API calls fails when internet connectivity is intermittent, which is explicitly stated as a deployment constraint. On-device inference with a fine-tuned small language model provides reliable offline operation. Reference: https://learn.microsoft.com/en-us/azure/architecture/data-guide/technology-choices/ai-services

A is incorrect: Assigning roles at management group scope to all tenant users violates least privilege and grants access far beyond what is needed. Azure RBAC best practices recommend avoiding broader roles at broader scopes even if it initially seems more convenient to do so. Reference: https://learn.microsoft.com/en-us/azure/role-based-access-control/best-practices

B is correct: The Azure AI User role is the minimum privilege role required to build agents, run traces, and perform core tasks in Foundry. Assigning roles to Microsoft Entra security groups rather than individual users follows Azure RBAC best practices for manageable and scalable access control. Reference: https://learn.microsoft.com/en-us/azure/ai-foundry/concepts/rbac-azure-ai-foundry

C is incorrect: The Owner role at subscription scope grants excessive permissions far beyond what is needed for building agents and running traces. Azure RBAC best practices recommend granting users the least privilege at the narrowest scope to minimize risk if a security principal is compromised. Reference: https://learn.microsoft.com/en-us/azure/role-based-access-control/best-practices

D is incorrect: Key-based authentication bypasses Azure RBAC entirely, granting full access without role restrictions and eliminating granular access control. Microsoft recommends using Entra ID authentication for improved security. The Contributor role also grants broader permissions than necessary. Reference: https://learn.microsoft.com/en-us/azure/ai-foundry/concepts/rbac-foundry

A is incorrect: Generative orchestration does not disable all system topics. System topics such as Conversation Start, Escalate, and Fallback continue to function normally. The specific limitation is that the Conversational boosting system topic is not used for knowledge searches. Reference: https://learn.microsoft.com/en-us/microsoft-copilot-studio/authoring-system-topics

B is incorrect: Power Automate flows continue to function within Copilot Studio when generative orchestration is enabled. The issue is not a flow compatibility limitation but rather that the Conversational boosting system topic itself is not invoked during knowledge searches. Reference: https://learn.microsoft.com/en-us/microsoft-copilot-studio/advanced-generative-actions

C is correct: With generative orchestration turned on, the agent does not use the Conversational boosting system topic when it searches knowledge sources. Any modifications made to this system topic for customizing how the agent searches knowledge are therefore bypassed entirely. Reference: https://learn.microsoft.com/en-us/microsoft-copilot-studio/advanced-generative-actions

D is incorrect: The Conversational boosting system topic does not require manual re-enablement after switching modes. The issue is that generative orchestration fundamentally does not invoke this system topic for knowledge searches, regardless of its enabled state. Reference: https://learn.microsoft.com/en-us/microsoft-copilot-studio/advanced-generative-actions

A is correct: The fairness assessment in the Responsible AI scorecard wizard is currently only available for categorical sensitive attributes such as gender. This is a documented limitation, meaning continuous numerical sensitive attributes are not directly supported for fairness assessment within the scorecard configuration wizard. Reference: https://learn.microsoft.com/en-us/azure/machine-learning/how-to-responsible-ai-insights-ui

B is incorrect: The Responsible AI dashboard supports both regression and classification models, including binary and multi-class classification. The fairness assessment component can evaluate performance disparities across sensitive groups for any of these supported model types when the sensitive attributes are categorical. Reference: https://learn.microsoft.com/en-us/azure/machine-learning/concept-responsible-ai-dashboard

C is incorrect: The opposite is true. The fairness assessment currently supports only categorical sensitive attributes, not continuous numerical ones. Categorical features such as gender or discrete age groups can be used, while raw continuous numerical values like exact income are not directly supported for the fairness assessment step. Reference: https://learn.microsoft.com/en-us/azure/machine-learning/how-to-responsible-ai-insights-ui

D is incorrect: The Responsible AI dashboard and scorecard can evaluate registered models without requiring a live AKS deployment. The model must be registered in Azure Machine Learning in MLflow format with a sklearn flavor, but live endpoint deployment is not a prerequisite for fairness assessment generation. Reference: https://learn.microsoft.com/en-us/azure/machine-learning/how-to-responsible-ai-insights-ui

A is incorrect: Data Lifecycle Management manages retention and deletion, but it doesn't offer the centralized dashboards, risk evaluation, or security guidance features provided by DSPM for AI to actively monitor sensitive data usage.

B is incorrect: Data Map primarily identifies and categorizes static data within registered sources, lacking the real-time monitoring capabilities needed for active AI interactions or user prompt/response tracking.

C is correct: DSPM for AI offers a consolidated platform for managing AI security and compliance, including discovery, protection, risk assessments, and streamlined policy application for sensitive data use in Copilot experiences and other AI applications.

D is incorrect: Information Barriers limit communication paths between user groups. They do not deliver comprehensive monitoring, compliance oversight, or insights into sensitive data movement within enterprise AI interactions.

A is correct: Prompt caching 'allows you to reduce overall request latency and cost for longer prompts that have identical content at the beginning of the prompt.' Rather than reprocessing the same input tokens repeatedly, the service retains a temporary cache of processed computations. Reference: https://learn.microsoft.com/en-us/azure/ai-foundry/openai/how-to/prompt-caching

B is incorrect: Azure OpenAI does not provide an automatic token compression feature. Token encoding is determined by the model's tokenizer and cannot be modified. Prompt caching is the correct mechanism for reducing costs associated with repetitive prompt content. Reference: https://learn.microsoft.com/en-us/azure/ai-foundry/openai/how-to/prompt-caching

C is incorrect: Batch processing is designed for non-real-time, large-scale workloads with a 24-hour turnaround target. It does not reduce latency for interactive agent requests and does not cache identical prompt prefixes between requests. The scenario requires real-time responses. Reference: https://learn.microsoft.com/en-us/azure/ai-foundry/foundry-models/concepts/deployment-types

D is incorrect: Azure OpenAI does not offer a prompt sharding feature that distributes prompts across instances. Prompts are processed as complete units by a single model instance. Prompt caching is the mechanism that addresses repeated identical prompt content. Reference: https://learn.microsoft.com/en-us/azure/ai-foundry/openai/how-to/prompt-caching

A is incorrect: While a Logic App can transform payloads, it adds unnecessary architectural complexity and cost. The common alert schema natively standardizes the JSON payload format across all alert types without requiring an intermediate transformation layer. Reference: https://learn.microsoft.com/en-us/azure/azure-monitor/alerts/alerts-common-schema

B is incorrect: Log search alert rules support KQL queries needed for Application Insights telemetry analysis that metric alert rules cannot perform. Eliminating log search alerts would sacrifice monitoring capabilities. The common alert schema solves the payload format issue directly. Reference: https://learn.microsoft.com/en-us/azure/azure-monitor/alerts/alerts-types

C is correct: The common alert schema provides a single standardized JSON structure for all Azure Monitor alert types. Enabling it on the webhook action ensures that metric alerts and log search alerts produce consistent payloads, eliminating the need for the integration to handle multiple format variations. Reference: https://learn.microsoft.com/en-us/azure/azure-monitor/alerts/alerts-common-schema

D is incorrect: Creating separate endpoints for each alert type increases integration complexity and maintenance burden. The common alert schema is specifically designed to provide a unified payload format across all alert types, avoiding the need for type-specific parsing logic. Reference: https://learn.microsoft.com/en-us/azure/azure-monitor/alerts/alerts-common-schema