Free Microsoft AI Transformation Leader (AB-731) Sample Questions

Correct answer: B — Guards against unauthorized access and data breaches.

• A. Incorrect. Preventing overfitting is a model training and generalization concern, not a primary benefit of secure AI design, which focuses on protecting data and systems from threats.
• B. Correct. A primary benefit of secure AI design is protecting systems from unauthorized access, misuse, and data breaches, ensuring confidentiality, integrity, and trust in AI solutions.
• C. Incorrect. Boosting novelty and creativity is a benefit of generative AI capabilities, not of secure AI design, which does not directly influence output creativity.
• D. Incorrect. Reducing computational resources is an efficiency or optimization goal, not a primary benefit of secure AI design, which is about mitigating risks and protecting assets.

Correct answer: B — Model evaluation

• A. Incorrect. Problem framing defines the scope, objectives, and ethical requirements, but it does not directly measure or validate fairness and transparency. The direct assessment of these properties occurs later in the lifecycle.
• B. Correct. Model evaluation is the stage where the team tests the model for performance, fairness, bias, and explainability before deployment. Techniques like SHAP values, fairness metrics, and bias detection are applied here to ensure that decisions are equitable and can be understood and justified.
• C. Incorrect. Model deployment focuses on integrating the model into production and operating it reliably. While monitoring can detect drift or issues afterward, fairness and transparency are primarily addressed during evaluation.
• D. Incorrect. Data preprocessing can reduce bias by cleaning and balancing data, but it is not the primary stage for ensuring explainability and fairness. These concerns are mainly validated during model evaluation, after the data and model have been prepared.

Correct answer: C — Few-shot with template examples

• A. Incorrect. Chain-of-thought prompting guides the model through step-by-step reasoning, which is useful for complex reasoning tasks but not for enforcing a rigid output structure or template. It does not reliably guarantee adherence to a specific company-defined format.
• B. Incorrect. Zero-shot with format instructions can specify an output structure, but without examples, the model may deviate from the required template, especially for consistent enterprise use cases. It is less reliable for strict template adherence compared to providing concrete examples.
• C. Correct. Few-shot with template examples is the most appropriate technique because it provides the model with concrete examples of the exact structure to imitate. This greatly improves the likelihood that summaries will follow the company-defined template consistently.
• D. Incorrect. Low temperature sampling reduces randomness and makes outputs more deterministic, but it does not inherently enforce a specific template or structure. While it can improve consistency, prompt design with examples is still superior for strict formatting requirements.

Correct answer: A — Add a safety filter to the generated text.

• A. Correct. Adding a safety filter is a direct and effective method to detect and block offensive language in generated text before it reaches users. This is the most straightforward safety control.
• B. Incorrect. Expanding the training corpus might improve general behavior but does not guarantee the elimination of offensive language; the model can still generate harmful content. Training data alone is not a reliable enforcement mechanism.
• C. Incorrect. Decreasing the learning rate is an optimization technique for fine-tuning, not a safety control. It affects how the model learns but does not specifically prevent offensive language generation.
• D. Incorrect. Increasing model parameters may improve capacity but does not inherently address safety or content moderation. Larger models can still produce offensive outputs without additional safeguards.

Correct answer: B — Linking generated content to verified external references.

• A. Incorrect. Grounding is not about relying only on internal training data; it specifically involves anchoring outputs to external, verified sources to improve accuracy and reduce hallucinations.
• B. Correct. Grounding in generative AI refers to linking generated content to verified external references such as databases, documents, or APIs, ensuring factual accuracy and traceability.
• C. Incorrect. Building the model architecture from fundamental components describes model design or engineering, not grounding. Grounding concerns the connection of outputs to real-world, external information.
• D. Incorrect. Enhancing prompts to increase output relevance is prompt engineering, not grounding. Grounding is about anchoring responses in external verified data, not simply refining prompts.

Correct answers: A, C — Missing sensor data for some periods; Only normal conditions, no failures

• A. Missing sensor readings create gaps in the time series, distorting feature calculations and anomaly detection. This directly reduces model reliability for predictive maintenance.
• B. The storage format (CSV) does not affect data quality or model reliability as long as data can be accurately parsed. It is a format choice, not a quality issue.
• C. A dataset with only normal operating conditions lacks failure examples, causing severe class imbalance. The model cannot learn to predict faults, undermining reliability.
• D. Sampling at 1-minute intervals is a design choice, not inherently a quality issue. It may affect granularity but does not necessarily reduce reliability.
• E. Consistent timestamp formatting is a positive data quality attribute that supports data integration, not a problem. Proper formatting does not threaten model reliability.

Correct answer: A — True

• A. The statement is true because few-shot prompting involves providing a language model with a few examples in the prompt to guide its output, which improves relevance and accuracy by helping the model infer the desired pattern, tone, or format.
• B. The statement is false because the scenario described explicitly includes providing examples in the prompt, which is the definition of few-shot prompting; therefore, the statement is true.

Correct answer: A — True

• A. The statement is true because underrepresentation in training data can lead to biased model outputs, as the model lacks sufficient examples to learn fair patterns for minority groups. Representative datasets are critical to ensure equitable and accurate predictions.
• B. The statement is false because biased predictions due to underrepresentation actually highlight the need for representative datasets, not the opposite. Therefore, the original statement is true.

Correct answer: A — True

• A. The statement is true because machine learning models can learn from new data and adapt to evolving fraud patterns dynamically, whereas static rules require manual updates and cannot keep pace with rapidly changing fraud tactics.
• B. The statement is true because machine learning models adapt to new fraud patterns and are more effective than static rules in dynamic environments; therefore, selecting 'False' is incorrect.

Correct answer: A — grounding

• A. Correct. Grounding in AI refers to the process of linking model outputs to reliable, verified external sources to ensure factual accuracy and reduce unsupported claims. It anchors the model to evidence.
• B. Incorrect. Hallucination is the opposite: a model generates plausible but factually incorrect or fabricated information not supported by real sources. It does not describe connecting outputs to verified data.
• C. Incorrect. Overfitting occurs when a model learns training data too specifically, including noise, leading to poor generalization. It is unrelated to ensuring factual accuracy through external sources.

Correct answer: A — machine learning

• A. Correct. Machine learning is designed to handle complex, evolving data patterns by learning from data rather than relying on static rules. It adapts over time, making it more suitable than fixed rule-based approaches when logic is complex or frequently changing.
• B. Incorrect. Manual scripting relies on explicitly written rules that require human intervention to update, making it rigid and ill-suited for scenarios where patterns evolve automatically.
• C. Incorrect. Static algorithms are fixed and do not learn from new data. They cannot adapt to changing data relationships, unlike machine learning which continuously improves.

Correct answer: C — Ensuring privacy for all data during processing.

• A. Incorrect. The total cost of a training run is a budget and efficiency concern, not a security or privacy safeguard. Regulations such as GDPR or HIPAA require protecting data, and a cheaper training run does nothing to meet that requirement.
• B. Incorrect. Inference latency affects performance and user experience, but it does not protect sensitive data. A fast response time cannot satisfy strict regulatory privacy obligations.
• C. Correct. When regulatory data privacy requirements are strict, the model and its surrounding pipeline must keep all data private during storage, training, inference, and processing. Meeting regulations such as GDPR or HIPAA means privacy has to be prioritized over other selection criteria.
• D. Incorrect. A model type's industry standing may increase confidence in it, but popularity or reputation does not guarantee compliance with privacy regulations. Compliance depends on how data is handled, not on how well known the model is.

Correct answer: B — Copilot in Teams for meeting recaps and action items

• A. Incorrect. Copilot in Word can assist with drafting documents, including meeting minutes, but it does not automatically capture action items or summarize live discussions during a meeting. It is more focused on post-meeting document creation rather than real-time meeting intelligence.
• B. Correct. Copilot in Teams is specifically designed to enhance meeting productivity by generating recaps, summarizing discussions, and capturing action items in real-time or post-meeting. This directly matches the need to automatically capture action items and summarize discussions.
• C. Incorrect. Copilot Chat can answer questions about meetings if provided with context, but it does not automatically capture action items or summarize discussions during a meeting. It is a general-purpose conversational tool, not specialized for meeting summarization.
• D. Incorrect. Copilot in Excel is focused on data analysis and spreadsheet management, not on capturing meeting discussions or action items. It is not the appropriate tool for improving meeting productivity in this scenario.

Correct answer: B — Index the feedback content using Graph connectors

• A. Incorrect. Copilot connectors integrate external apps or services with Copilot but do not index content for grounding. Graph connectors are needed to index data and enable Copilot to ground responses in specific sources like feedback.
• B. Correct. Graph connectors index external data sources (e.g., emails, surveys, call transcripts) into Microsoft Graph, allowing Copilot to ground its responses in this content. This is the recommended approach for making external data accessible to Copilot.
• C. Incorrect. Building a custom model in Azure AI Studio is for creating tailored AI applications, not for grounding Microsoft 365 Copilot in existing data. The scenario requires indexing data via Graph connectors.
• D. Incorrect. Azure Data Factory is an ETL and data integration tool; it does not directly enable Copilot grounding. Copilot requires indexed content in Microsoft Graph, which is provided by Graph connectors.

Correct answer: A — Copilot in PowerPoint to get design ideas

• A. Correct. Copilot in PowerPoint can automatically generate design ideas and reorganize slide content into a better layout. This is a built-in feature of Microsoft 365 Copilot for PowerPoint, leveraging AI to improve slide design without manual effort.
• B. Incorrect. Copilot Chat provides conversational assistance and can suggest layout ideas, but it does not directly apply or automatically reorganize slide content in PowerPoint. It is less actionable for direct design changes.
• C. Incorrect. The Designer feature in PowerPoint offers design suggestions, but it is a manual tool that requires user interaction. It does not leverage Copilot's automatic reorganization capabilities as described in the question.
• D. Incorrect. Converting a PowerPoint slide to Word and using Copilot there would not preserve slide layout or design. This approach is inefficient and not intended for reorganizing PowerPoint content.

Correct answer: A — Sensitivity labels with encryption and access control

• A. Correct. Sensitivity labels with encryption and access control classify and protect sensitive data like salary information. Copilot respects these labels and will not surface or generate responses containing labeled content, preventing inadvertent disclosure.
• B. Incorrect. DLP policies primarily prevent data exfiltration through sharing or uploads but do not directly control Copilot's access to or disclosure of sensitive information in responses. DLP is more about preventing data loss from sharing, not filtering Copilot outputs.
• C. Incorrect. Insider risk management is designed to detect and investigate risky user behavior, not to prevent Copilot from disclosing confidential data. It is a monitoring and investigation capability, not a content protection control for Copilot responses.
• D. Incorrect. Information barrier policies restrict communication and collaboration between defined groups but do not control or filter the content Copilot generates or accesses. They are for compliance in regulated environments, not for protecting against data disclosure via Copilot.

Correct answer: A — Click 'Summarize' in the Copilot sidebar

• A. In Outlook on the web, Microsoft 365 Copilot provides a 'Summarize' option in the Copilot sidebar, which quickly extracts key points, decisions, and action items from email threads without leaving Outlook.
• B. Forwarding the email to Copilot Chat is not the standard workflow; the user should use Copilot within Outlook itself for efficiency and direct summarization.
• C. The 'Action Items' prompt is designed to extract tasks or follow-ups, not to summarize key decisions. The question asks for understanding key decisions, so a summary is more appropriate.
• D. Opening the email in Word is unnecessary; Copilot summarization is available natively in Outlook on the web, making this approach inefficient.

Correct answers: A, B, D — The solution requires training on proprietary company data.; It must integrate with legacy on-premises databases.; The organization needs to fine-tune the model continuously.

• A. Correct. Building a custom solution with Azure AI allows training on proprietary company data, providing more control over data ingestion, retrieval, and governance compared to Microsoft 365 Copilot's standard grounding model. This is essential when data must remain private or meet specific compliance requirements.
• B. Correct. A custom Azure AI solution can be designed to connect to legacy on-premises databases through custom integrations, APIs, or hybrid networking, whereas Microsoft 365 Copilot is primarily cloud-based and may not support such integrations natively.
• C. Incorrect. Native embedding in Outlook and Teams is a strength of Microsoft 365 Copilot, not a reason to build a custom solution. If the goal is to work directly inside Microsoft 365 apps, Copilot is usually the better choice.
• D. Correct. Custom solutions built with Azure AI can be fine-tuned continuously to adapt to evolving business needs, whereas Microsoft 365 Copilot's models are managed by Microsoft and do not offer the same level of customization or iterative control.
• E. Incorrect. Automatic compliance with Microsoft 365 data policies is an advantage of Microsoft 365 Copilot within the Microsoft 365 ecosystem. While a custom Azure AI solution can be made compliant, it requires the organization to design and enforce those controls, making this a strength of Copilot, not a reason to build custom.
• F. Incorrect. A two-week timeline favors using an off-the-shelf service like Microsoft 365 Copilot, which can be deployed quickly. Building a custom solution typically takes longer due to design, integration, testing, and deployment efforts.

Correct answers: B, D, F — Assign Copilot licenses to specific users or groups.; Analyze Copilot usage metrics in reports.; Disable Copilot in specific Office desktop apps.

• A. Incorrect. Turning off web content grounding for Copilot is not a setting in the Microsoft 365 admin center; it is configured via Microsoft Purview or related service policies.
• B. Correct. Admins can assign Copilot licenses to specific users or groups in the Microsoft 365 admin center under the Licenses section, enabling access for designated users.
• C. Incorrect. Uploading a custom logo for the Copilot pane is not available in the Microsoft 365 admin center; branding options are managed in other areas of Microsoft 365.
• D. Correct. The Microsoft 365 admin center provides Copilot usage reports and analytics under Reports, allowing admins to track adoption and measure value.
• E. Incorrect. Defining how long Copilot stores chat history is not a setting in the admin center; it is governed by Microsoft 365 compliance policies or Microsoft Purview.
• F. Correct. Admins can disable Copilot in specific Office desktop apps via Settings > Org settings > Microsoft Copilot in the Microsoft 365 admin center, tailoring availability.

Correct answer: B — False

• A. The statement is false because Microsoft 365 Copilot Chat is a web-based experience that does not require a desktop installation of Microsoft 365 apps; it can be accessed directly through a web browser.
• B. The statement is false because the web experience is designed to work in a browser independent of locally installed Microsoft 365 applications, so no desktop installation is needed.

Correct answer: B — Azure Vision

• A. Incorrect. Azure AI Search is designed for indexing and searching large volumes of structured and unstructured data, not for image analysis or object detection. It cannot directly detect stock-outs or misplaced items from images.
• B. Correct. Azure Vision (part of Azure AI Services) provides computer vision capabilities, including object detection, image analysis, and custom vision models. It is ideal for automatically detecting stock-outs, misplaced items, and product recognition from shelf images.
• C. Incorrect. Azure Machine Learning is a platform for building, training, and deploying custom machine learning models. While it could solve this problem with significant custom development, it does not offer out-of-the-box computer vision features for image analysis like Azure Vision does.
• D. Incorrect. Azure Bot Service is used to create conversational AI bots and virtual assistants. It does not analyze images or detect inventory issues from shelf photos.

Correct answer: B — Custom object detection via Edge

• A. Incorrect. Cloud-based image analysis introduces network latency because images must be uploaded and results returned over the internet, making it unsuitable for the strict 100-millisecond alert requirement. Real-time conveyor-belt inspection requires local inference to avoid round-trip delays.
• B. Correct. Custom object detection via Edge leverages Azure AI Vision on edge devices (e.g., a camera or gateway near the conveyor belt). This enables on-device or near-device inference with very low latency, meeting the sub-100-millisecond requirement for real-time defect detection and alert triggering.
• C. Incorrect. Video live stream analysis in the cloud still relies on sending video frames over the network for processing, which adds latency and can be unreliable for ultra-low-latency industrial inspection. It is better suited for streaming analytics scenarios where response times can be relaxed.
• D. Incorrect. Batch image processing is designed for offline or scheduled analysis of collections of images, not immediate real-time detection. It cannot provide the instantaneous feedback needed to inspect parts and trigger alerts within 100 milliseconds.

Correct answer: B — Azure Vision classification and object detection

• A. Incorrect. Azure AI Search with an image indexer is designed for indexing and searching images based on metadata or visual features, not for medical image analysis that requires classification, localization, and explainability. It lacks the ability to highlight regions of interest or provide diagnostic explanations.
• B. Correct. Azure Vision with image classification and object detection can analyze X-ray images, classify findings, and detect regions of interest through bounding boxes or segmentation. While it may not natively generate textual explanations, it can be integrated with other services to highlight and explain findings, meeting the requirement.
• C. Incorrect. Azure Machine Learning designer is a drag-and-drop tool for building, training, and deploying custom machine learning models. It does not provide pre-built vision capabilities for medical image analysis, and building such a model would require significant custom development. It is not the most direct Foundry tool for this scenario.
• D. Incorrect. Azure Bot Service with a QnA knowledge base is designed for creating conversational AI and chatbots that answer questions from a knowledge base. It cannot analyze images or provide explainable AI insights for X-ray analysis.

Correct answer: C — Vector embeddings of images

• A. Incorrect. OCR extracts text from images (e.g., labels, receipts) but does not generate embeddings for visual similarity matching. It cannot compare visual features between a query image and catalog images.
• B. Incorrect. Image classification assigns predefined labels (e.g., 'shirt', 'dress') to an image. While useful for tagging, it does not enable finding visually similar items based on an uploaded photo, as it lacks similarity comparison capabilities.
• C. Correct. Vector embeddings of images represent visual content as numerical vectors. Combined with Azure AI Search, these vectors can be indexed and compared to find semantically or visually similar items in the catalog, making this the core capability for visual search.
• D. Incorrect. Video analysis processes video content (e.g., motion detection, scene understanding) over time. It is not relevant for static image-based visual search in an e-commerce catalog.

Correct answer: C — Faceted navigation

• A. Incorrect. Semantic ranker improves search result relevance by understanding query intent and context, but it does not provide category-based filtering. It reorders results based on semantic meaning rather than enabling users to narrow by predefined categories.
• B. Incorrect. Vector search retrieves semantically similar content using vector embeddings, which is useful for natural-language queries but does not filter results by predefined metadata categories like author or date.
• C. Correct. Faceted navigation allows users to filter and browse search results dynamically using predefined categories (facets) such as author, date, or product type. It leverages indexed metadata to enable drill-down, making it the feature designed for this purpose.
• D. Incorrect. Key phrase extraction identifies important terms or phrases from content to assist with indexing or summarization, but it does not offer interactive filtering of search results based on categories.

Correct answer: B — Image description generation

• A. Incorrect. Face detection identifies and localizes human faces in images, which is useful for tagging or privacy-related scenarios, but it does not generate descriptive alt-text for accessibility.
• B. Correct. Image description generation (part of Azure Computer Vision's 'Describe' feature) analyzes an image and produces a natural-language caption, which is directly useful for creating alt-text to improve accessibility on a website.
• C. Incorrect. Logo detection recognizes brand logos in images, which is helpful for marketing or asset management, but it does not provide a general description of the image content for alt-text.
• D. Incorrect. Landmark recognition identifies famous places or landmarks in an image, but it is a narrow capability and does not generate general-purpose alt-text for accessibility.

Correct answers: B, D — They significantly reduce development time and expertise required.; They are cost‑effective for common tasks and deploy rapidly.

• A. Incorrect. Pre-built AI models do not guarantee a fixed performance improvement over custom models. Their performance depends on the task, data quality, and domain specificity. Custom models can outperform pre-built models in specialized scenarios.
• B. Correct. Pre-built models significantly reduce development time because they are ready to use without training from scratch. They also require less specialized expertise compared to building and tuning a custom model, making AI adoption easier for teams.
• C. Incorrect. Even pre-built models require validation and testing to ensure they perform correctly for the intended use case. Testing is essential for accuracy, reliability, and compliance.
• D. Correct. Pre-built models are typically cost-effective for common tasks because they eliminate the expense of custom development and training. They can also be deployed quickly, reducing time to value.
• E. Incorrect. Pre-built models do not automatically adapt to domain-specific jargon. They handle general language well, but customization or fine-tuning is usually required for specialized terminology.

Correct answer: A — True

• A. The statement is true because Microsoft Foundry is built for enterprise use and includes role-based access control (RBAC) for managing permissions. It also operates within Microsoft's compliance framework, which includes certifications such as SOC 2 and HIPAA, ensuring data protection and regulatory alignment.
• B. The statement is false because Microsoft Foundry is a consumer-only or lightly managed service and does not include enterprise security features or compliance certifications like SOC2 and HIPAA.

Correct answer: B — suitability and trade‑offs

• A. Incorrect. 'Color theme' is irrelevant to technical and functional evaluation of an AI model. Accuracy, latency, and data privacy are model-selection criteria, not visual styling characteristics.
• B. Correct. 'Suitability and trade-offs' directly captures the need to balance competing factors such as accuracy, latency, and data privacy when selecting an AI model for a specific use case. This phrase reflects the practical constraints and priorities organizations must weigh.
• C. Incorrect. 'Download count' is a popularity metric and does not reflect the model's performance, reliability, or alignment with organizational requirements. It is not a meaningful criterion for enterprise model selection.

Correct answer: C — Optical character recognition (OCR).

• A. Incorrect. Image classification (scene labeling) assigns descriptive labels to an image as a whole, such as recognizing that it shows a document or an office. It does not read the text on the page, so it cannot extract specific fields like the claim number, date, or amount.
• B. Incorrect. Object detection with bounding boxes locates and identifies physical objects within an image, such as vehicles or people. It does not read characters, so it cannot pull claim numbers, dates, or amounts from scanned forms.
• C. Correct. Optical character recognition (OCR) extracts printed and handwritten text from images and scanned documents. In Azure Vision in Foundry, OCR reads the text on each claim form so key fields like the claim number, date, and amount can be captured and passed to downstream automation.
• D. Incorrect. Facial recognition (face matching) identifies or verifies individuals by comparing facial features. It has no role in reading text, so it does not help extract claim number, date, or amount from claim forms.

Correct answers: A, C, D — Design conversational topics using trigger phrases.; Deploy as a bot to Microsoft Teams and other channels.; Extract entities from user input for slot filling.

• A. Correct. Copilot Studio lets developers design conversational topics that are triggered by phrases the user types. This is a core low-code capability for building a custom copilot and needs no extra tooling.
• B. Incorrect. Speech recognition through Azure Speech services is not an out-of-the-box Copilot Studio capability. It needs an external Azure service and additional configuration, so it is not a built-in feature.
• C. Correct. Copilot Studio has built-in multi-channel publishing. A copilot can be deployed as a bot to Microsoft Teams, websites, mobile apps and other supported channels without additional tools.
• D. Correct. Copilot Studio extracts entities from user input for slot filling out of the box. This captures key values such as dates or names during a conversation as part of the built-in conversational design experience.
• E. Incorrect. Training a custom large language model from scratch is not a capability of Copilot Studio. It orchestrates prompts and connects to data and models, but model training requires advanced tools such as Azure AI.
• F. Incorrect. Automatically creating visualizations from data is not a native Copilot Studio feature. That would require integration with analytics and BI tools such as Power BI.

Correct answers: A, B, C — Show total sales by region and product.; Create a bar chart of revenue by month.; Add a column calculating profit margin.

• A. Correct. Copilot in Excel can interpret natural language prompts to summarize and aggregate data, such as showing total sales by region and product. This is a typical data analysis request that fits Copilot's capabilities in Excel.
• B. Correct. Copilot in Excel can create visualizations such as bar charts from a natural language request like 'Create a bar chart of revenue by month.' This is a common way to analyze and present sales trends.
• C. Correct. Copilot in Excel can add formula-driven columns, such as a profit margin calculation, from a natural language prompt. This fits its role in transforming and analyzing spreadsheet data.
• D. Incorrect. Sending an email is not a capability of Copilot in Excel, because that action requires Outlook or another email client. Copilot in Excel works with worksheet data rather than sending messages.
• E. Incorrect. Highlighting cells above a threshold is a conditional formatting task, which this question does not treat as one of the three core data analysis prompt types. It changes how cells look rather than summarizing, visualizing, or calculating sales data.
• F. Incorrect. Printing a worksheet with page breaks is a document layout task, not an analysis prompt for Copilot in Excel. It does not involve interpreting, summarizing, or calculating sales data.

Correct answer: B — To expose APIs for accessing Microsoft 365 data and relationships.

• A. Incorrect. Microsoft Graph is not the store for Copilot prompts, responses, or usage transcripts. Its purpose is to expose Microsoft 365 data such as emails, calendars, files, and chats through a unified API. Interaction history and audit records are handled by other Microsoft 365 services, not by Graph.
• B. Correct. Microsoft Graph provides a unified set of APIs for accessing Microsoft 365 data, such as emails, calendars, files, chats, and users, along with the relationships between them. Microsoft 365 Copilot uses this to ground its responses in the user's work context and to take relevant actions. This data access and relationship layer is its primary role.
• C. Incorrect. Microsoft Graph can expose directory and identity data, but it is not the primary system for administering user permissions or licenses for Copilot groups. Those tasks are performed through Microsoft Entra ID and the Microsoft 365 admin center. Graph's main role is data access and integration.
• D. Incorrect. Compliance and security rules for AI responses are enforced by Microsoft 365 governance, security, and compliance controls such as Microsoft Purview, not by Microsoft Graph itself. Graph is a data access layer and does not set or enforce policies on Copilot's output.

Correct answer: B — Azure Vision OCR with AI Language PII redaction.

• A. Incorrect. Azure AI Search synonym maps expand query terms to improve search relevance, so they do nothing to find or remove sensitive data. They also can't read text from scanned images or detect PHI, so they can't do the redaction the hospital needs.
• B. Correct. Azure Vision OCR extracts the text from scanned medical records, and Azure AI Language PII detection then identifies and redacts personal health information such as names, dates, and ID numbers. Used together, they automate PHI redaction before the records are shared for research.
• C. Incorrect. Azure Machine Learning for image synthesis is about generating or manipulating images, not extracting text or detecting sensitive data. It has no built-in OCR or PII redaction, so it doesn't address the PHI redaction workflow.
• D. Incorrect. Azure Bot Service is for building conversational agents, not for document OCR or PII redaction. Custom code could orchestrate a workflow around it, but the service itself isn't the capability that pairs with Azure Vision for this task.

Correct answer: B — Custom object detection

• A. Optical character recognition extracts printed or handwritten text from images or video frames. It does not find objects such as hard hats.
• B. Correct. Custom object detection lets you supply your own labeled images (for example, photos with and without hard hats), train a model on them, and get back the coordinates in the frame where each label is found. That is what locating hard hats in a video feed requires.
• C. Image classification applies one or more labels to an entire image, such as 'construction site'. It does not return the coordinates of individual objects, so it cannot locate each hard hat in a frame.
• D. Facial recognition identifies or verifies individuals from their facial features. It is not designed to detect safety gear such as hard hats.