A is incorrect: Protected material detection identifies known text content such as lyrics, articles, and code from public repositories in the model output. It prevents copyright infringement in AI-generated content but does not detect or mitigate prompt injection attacks from user inputs or document-based threats. Reference: https://learn.microsoft.com/en-us/azure/ai-foundry/ai-services/content-safety-overview

B is incorrect: The custom categories API allows organizations to define and train filters for domain-specific content. While valuable for customized content moderation, it is not designed to detect adversarial prompt injection patterns across both user prompts and embedded document attacks like Prompt Shields is. Reference: https://learn.microsoft.com/en-us/azure/ai-services/content-safety/overview

C is correct: Prompt Shields is a unified API in Azure AI Content Safety that detects and blocks adversarial user input attacks on LLMs. It addresses both direct attacks (jailbreaks) where users deliberately exploit system vulnerabilities, and indirect attacks where malicious prompts are embedded within documents the AI might process. This covers both identified vulnerability types. Reference: https://learn.microsoft.com/en-us/azure/ai-services/content-safety/concepts/jailbreak-detection

D is incorrect: Groundedness detection identifies whether LLM responses are factually based on provided source materials. While useful for preventing hallucinations, it does not detect or block adversarial prompt injection attacks in user inputs or document-embedded instructions. It addresses output accuracy, not input security. Reference: https://learn.microsoft.com/en-us/azure/ai-foundry/openai/concepts/content-filter-groundedness

A is correct: The Allowed Locations built-in policy uses the deny effect to restrict where resources can be deployed. Assigning it at the parent management group level ensures enforcement propagates to all child subscriptions and resource groups within the hierarchy through a single assignment. Reference: https://learn.microsoft.com/en-us/azure/governance/policy/overview

B is incorrect: Creating duplicate custom definitions per subscription is unnecessary and introduces administrative overhead. The built-in Allowed Locations policy already exists and should be assigned at a higher scope in the management group hierarchy rather than replicated per subscription. Reference: https://learn.microsoft.com/en-us/azure/governance/policy/concepts/definition-structure-basics

C is incorrect: Azure Firewall controls network traffic at the data plane, not resource deployment at the control plane. It cannot prevent the creation of Azure AI resources in non-EU regions through Azure Resource Manager. Azure Policy operates at the control plane to enforce deployment restrictions. Reference: https://learn.microsoft.com/en-us/azure/governance/policy/overview

D is incorrect: Azure RBAC controls user actions based on identity permissions, not resource properties like deployment location. Users in any geographic location can create resources in EU regions. Azure Policy is the correct tool for enforcing resource location constraints. Reference: https://learn.microsoft.com/en-us/azure/governance/policy/overview

A is incorrect: Renaming system status reason values impacts all applications using the opportunity table across Dynamics 365 Sales. This creates regression risks and may break existing business logic, workflows, and reports that depend on the standard value names. Reference: https://learn.microsoft.com/en-us/microsoft-copilot-studio/knowledge-add-dataverse

B is incorrect: Custom Azure Functions add unnecessary infrastructure and maintenance burden. Copilot Studio provides built-in synonyms and glossary features specifically designed to bridge the gap between user language and the Dataverse schema. Reference: https://learn.microsoft.com/en-us/power-apps/maker/data-platform/data-platform-copilot

C is incorrect: Duplicating data into a separate table introduces data synchronization complexity, storage overhead, and consistency risks. Glossary terms solve this mapping problem without duplicating records or modifying the data model. Reference: https://learn.microsoft.com/en-us/microsoft-copilot-studio/knowledge-add-dataverse

D is correct: Glossary terms define domain-specific terminology and acronyms for the agent. Adding a glossary term mapping 'closed deals' to 'opportunities with status reason Won' provides business context that helps the AI model translate the natural language query into the correct Dataverse query. Reference: https://learn.microsoft.com/en-us/power-apps/maker/data-platform/data-platform-copilot

A is incorrect: Cloning the project and rerunning the flow duplicates effort unnecessarily. The built-in capability to start new evaluation rounds on existing batch run outputs avoids rerunning the flow and conserves resources. Reference: https://learn.microsoft.com/en-us/azure/ai-foundry/how-to/flow-bulk-test-evaluation

B is incorrect: While exporting data is possible, this approach is unnecessarily complex. Prompt flow natively supports running additional evaluation rounds against existing batch run outputs directly from the Runs tab interface. Reference: https://learn.microsoft.com/en-us/azure/machine-learning/prompt-flow/how-to-bulk-test-evaluate-flow

C is correct: Prompt flow allows you to run a new evaluation round on completed batch run outputs without rerunning the flow itself. Navigate to the Runs tab, open the batch run detail page, and select Evaluate to apply a different evaluation method to the existing outputs. Reference: https://learn.microsoft.com/en-us/azure/machine-learning/prompt-flow/how-to-bulk-test-evaluate-flow

D is incorrect: Deleting the batch run discards the outputs and wastes the compute resources already consumed. Prompt flow specifically supports starting additional evaluation rounds on completed batch runs to save time and cost. Reference: https://learn.microsoft.com/en-us/azure/ai-foundry/how-to/flow-bulk-test-evaluation

A is incorrect: Manual observation is not scalable, reproducible, or integrable with CI/CD pipelines. The Testing module provides automated structured validation that enhances reliability and integrates into automated deployment pipelines. Reference: https://learn.microsoft.com/en-us/power-automate/desktop-flows/test-desktop-flows

B is correct: The Testing module in Power Automate for desktop provides structured test cases following a Given/When/Then structure. The Assert action validates outputs against expected results using operators like equals, contains, or greater than. Tests can be exported and integrated into automated deployment pipelines. Reference: https://learn.microsoft.com/en-us/power-automate/desktop-flows/test-desktop-flows

C is incorrect: While technically feasible, this approach does not leverage the purpose-built Testing module available in Power Automate for desktop. The Testing module provides native Assert actions and direct integration into automated pipelines without requiring separate cloud flow orchestration. Reference: https://learn.microsoft.com/en-us/power-automate/desktop-flows/test-desktop-flows

D is incorrect: Custom scripts add unnecessary development and maintenance overhead. The built-in Testing module provides native Assert and Test a desktop flow actions designed specifically for structured desktop flow validation. Reference: https://learn.microsoft.com/en-us/power-automate/desktop-flows/test-desktop-flows

A is correct: Restricted SharePoint Search is designed as a temporary measure to help maintain momentum with Copilot deployment while implementing comprehensive data security. It restricts enterprise search and Copilot experiences to an allowed list of curated SharePoint sites while administrators audit and remediate permissions. Reference: https://learn.microsoft.com/en-us/sharepoint/restricted-sharepoint-search

B is incorrect: Delaying deployment indefinitely forfeits business value while competitors adopt AI capabilities. Restricted SharePoint Search and SharePoint Advanced Management allow organizations to deploy Copilot safely while addressing oversharing concerns progressively. Reference: https://learn.microsoft.com/en-us/sharepoint/restricted-sharepoint-search

C is incorrect: Revoking all permissions disrupts active collaboration across the organization and blocks users from their own content. This approach is disproportionate and operationally infeasible at scale compared to using Restricted SharePoint Search as a targeted interim control. Reference: https://learn.microsoft.com/en-us/sharepoint/get-ready-copilot-sharepoint-advanced-management

D is incorrect: Disabling Microsoft Graph would break Copilot functionality entirely and also impact other Microsoft 365 services that depend on Graph. The recommended approach is to use Restricted SharePoint Search and SharePoint Advanced Management for targeted, granular controls. Reference: https://learn.microsoft.com/en-us/copilot/microsoft-365/microsoft-365-copilot-architecture

A is incorrect: Power Automate email summaries provide delayed, non-interactive reports that lack real-time monitoring, interactive drill-down, and KQL-based visualization. The Copilot Studio Dashboard offers a unified, shareable, and customizable monitoring experience directly within Application Insights.

B is incorrect: The built-in Copilot Studio analytics page provides metrics but does not support custom KQL queries or the extensive customization of Azure Workbooks. Granting operations staff maker access solely for monitoring violates least-privilege principles.

C is correct: The Copilot Studio Dashboard, a built-in Azure Workbook available in Application Insights, visualizes key agent metrics using Azure Monitor signals. It is editable, supports custom KQL tiles for additional attributes, and can be shared with team members having appropriate permissions.

D is incorrect: Exporting raw data to CSV and creating Excel dashboards results in a manual, disconnected monitoring process lacking real-time updates, KQL query capabilities, and interactive drill-down. Azure Workbooks offer an automated, customizable, and integrated alternative within Application Insights.

A is incorrect: Individual fairness is a concept where similar individuals should receive similar predictions, but it does not require identical scores for all. The Responsible AI dashboard does not use individual fairness; it specifically employs group fairness, evaluating how model predictions create disparities across demographic groups defined by sensitive features. Reference: https://learn.microsoft.com/en-us/azure/machine-learning/concept-fairness-ml

B is incorrect: Architectural fairness is not a concept used by the Responsible AI dashboard. Model architecture analysis relates to interpretability, not fairness assessment. The dashboard evaluates group fairness by measuring disparity metrics across subgroups defined by sensitive features to identify which populations face disproportionate impacts. Reference: https://learn.microsoft.com/en-us/azure/machine-learning/concept-fairness-ml

C is incorrect: Sequential fairness is not a recognized fairness concept in the Responsible AI dashboard. The dashboard conceptualizes fairness through group fairness, evaluating whether model predictions disproportionately affect groups defined by sensitive attributes such as gender, ethnicity, or age, not the ordering of inference outputs. Reference: https://learn.microsoft.com/en-us/azure/machine-learning/concept-fairness-ml

D is correct: The Responsible AI dashboard uses an approach known as group fairness, which asks 'Which groups of individuals are at risk for experiencing harm?' The term sensitive features indicates that the system designer should be sensitive to these features when assessing group fairness. Fairness is quantified through disparity metrics that evaluate model behavior across groups as ratios or differences. Reference: https://learn.microsoft.com/en-us/azure/machine-learning/concept-fairness-ml

A is correct: Code Interpreter has additional charges beyond token-based fees, and each simultaneous thread creates its own session. With 15 units running concurrently, 15 separate Code Interpreter sessions would be created, each active for up to 1 hour. This directly impacts cost projections. Reference: https://learn.microsoft.com/en-us/azure/ai-foundry/openai/how-to/code-interpreter

B is incorrect: The sandboxed Python environment is a managed service capability. Costs are driven by session duration and token consumption, not by the number of pre-installed libraries. There is no separate licensing for Python packages in the sandbox. Reference: https://azure.microsoft.com/en-us/pricing/details/azure-openai/

C is incorrect: Code Interpreter explicitly has additional charges beyond the token-based fees for Azure OpenAI usage. Session costs must be factored into the total cost of ownership alongside inference token costs for accurate financial planning. Reference: https://learn.microsoft.com/en-us/azure/ai-foundry/openai/how-to/code-interpreter

D is incorrect: Code Interpreter does not use flat monthly subscription pricing. It uses session-based pricing with additional charges beyond token fees. Each session created by simultaneous thread usage contributes independently to cost. Reference: https://azure.microsoft.com/en-us/pricing/details/azure-openai/

A is incorrect: A high confidence threshold increases precision but significantly reduces recall, meaning more genuinely negative feedback would be missed. This directly contradicts the requirement to flag all potentially negative feedback, even at the cost of reviewing additional non-negative entries. Reference: https://learn.microsoft.com/en-us/azure/ai-foundry/responsible-ai/language-service/transparency-note-sentiment-analysis

B is incorrect: Positive, negative, and neutral are independent sentiment dimensions with scores that add up to 1. A low positive score does not necessarily indicate negative sentiment because the text could be neutral. This approach incorrectly conflates absence of positive sentiment with presence of negative sentiment. Reference: https://learn.microsoft.com/en-us/azure/ai-services/language-service/sentiment-opinion-mining/how-to/call-api

C is correct: According to Microsoft guidance, using a lower threshold on the negative sentiment confidence score captures more potentially negative feedback. This results in fewer false negatives (missed negative feedback) but more false positives (non-negative entries flagged), which aligns with the requirement to not miss negative feedback. Reference: https://learn.microsoft.com/en-us/azure/ai-foundry/responsible-ai/language-service/transparency-note-sentiment-analysis

D is incorrect: The 'mixed' label appears when a document contains both positive and negative sentences. Relying solely on this label misses documents that are entirely negative or have subtle negative sentiment within otherwise neutral-labeled documents and is not comprehensive. Reference: https://learn.microsoft.com/en-us/azure/ai-services/language-service/sentiment-opinion-mining/how-to/call-api