AB-731 Questions and Answers

model inaccuracy

The scenario describes a model producing plausible-sounding content that is factually wrong —a common generative AI failure mode often referred to as a “hallucination.” Since “hallucination” is not offered in the dropdown, the best matching choice is model inaccuracy because the core problem is that the model’s output is incorrect even though it appears confident and coherent.

The other options do not fit the definition of the behavior: data leakage is about sensitive information being exposed (for example, proprietary prompts, secrets, or personal data). Prompt injection is an attack technique where a user tries to override system instructions or cause unsafe actions. Overreliance describes a human/organizational risk —trusting the model too much—rather than the model’s intrinsic behavior of generating incorrect facts. Overreliance can be a consequence of this behavior, but it is not what the behavior itself is called.

In practice, you mitigate this kind of inaccuracy by grounding responses in trusted sources (for example, RAG), constraining prompts with explicit requirements, using verification steps (citations, cross-checking, tool-based validation), and adding human review for high-impact use cases.

Generative AI’s defining capability is producing new content (text, images, code) in response to instructions—most commonly provided as natural language prompts. Option A best captures that general-purpose description for stakeholders: users ask questions or provide instructions, and the system generates responses or drafts content accordingly.

B is a specific application (translation) that generative AI can do, but it’s not the defining description. C describes predictive analytics/forecasting, which is a different AI category. D describes recommendation systems, typically driven by user behavior and ranking algorithms, which can be enhanced by AI but is not the core definition of generative AI.

A pretrained generative AI model is trained initially on a large, broad, and diverse dataset so it learns general language (or multimodal) patterns and capabilities. Fine-tuning then takes that pretrained base and performs additional training on a smaller, task- or domain-specific dataset to specialize behavior—improving performance for a particular use case, tone, style, or domain knowledge representation. That is exactly what option C states, making it the correct answer.

Option A is incorrect because both pretraining and fine-tuning may use labeled or unlabeled data depending on the technique; the distinction is not “labeled vs. unlabeled.” Option B is incorrect because a pretrained model is not “faster to train” due to fewer parameters; pretraining is typically the most compute-intensive phase precisely because it’s done at large scale, while fine-tuning is smaller but still trains the same model architecture. Option D is reversed: the pretrained model is the general-purpose foundation, while the fine-tuned model is the specialized variant for a specific task or dataset.

Answer Area

Azure Vision in Foundry Tools can extract and analyze key phrases from PDF files. Answer: No

Azure Vision in Foundry Tools can generate images based on natural language descriptions. Answer: No

Azure Document Intelligence in Foundry Tools can be used to automate the processing of invoices and credit notes. Answer: Yes

No — Azure Vision in Foundry Tools focuses on computer vision tasks such as image analysis and OCR (reading text from images and documents). While it can extract text from scanned PDFs via OCR, key phrase extraction is a natural language processing capability provided by Azure Language in Foundry Tools , not Azure Vision. Key phrase extraction analyzes text to identify main concepts, which is a different service family than vision.

No — Azure Vision can analyze existing images (for example, generate captions/descriptions of an image), but generating new images from a text prompt is a generative model capability (for example, DALL·E through Azure OpenAI/Azure AI Foundry model endpoints), not an Azure Vision feature. Vision describes what it “sees”; it doesn’t synthesize new images from natural language.

Yes — Azure Document Intelligence in Foundry Tools is designed for intelligent document processing , including automating extraction of structured fields from financial documents. Microsoft provides prebuilt models for invoices and supports custom extraction for similar document types, which makes it suitable for automating workflows involving invoices and credit-note style documents (field extraction, validation, routing).

The requirement is specific: users must build and use declarative agents that can access work data (tenant data / organizational context). Microsoft’s licensing guidance for Copilot extensibility ties use of declarative agents to having the appropriate Copilot entitlement that enables tenant grounding and organizational data access. In Microsoft’s cost and licensing considerations for declarative agents, Microsoft states that to use a declarative agent, users must have a Microsoft 365 Copilot add-on license (or an equivalent Copilot Chat add-on path tied to eligible licensing). Therefore, among the provided options, the best recommendation is A.

Option B (Copilot Studio user license) is primarily about authoring/building agents in Copilot Studio, but it is not, by itself, the licensing prerequisite that grants end users the right to use those agents with full Microsoft 365 Copilot capabilities and work-data grounding inside the Microsoft 365 Copilot environment. Publishing/building can be separate from the end-user entitlement to use the agent with organizational context.

Option C (Copilot Chat pay-as-you-go) can enable usage-based access to declarative agents in some configurations, but the question asks for the best license recommendation for users who need work-data access through declarative agents. The Microsoft 365 Copilot add-on is the straightforward, fully supported entitlement for that scenario.

promote the ethical and accountable use of AI.

Microsoft’s Responsible AI principles exist to guide organizations toward trustworthy, ethical, and accountable development and use of AI systems. Among the provided options, “ promote the ethical and accountable use of AI ” is the most accurate and comprehensive statement because it reflects the purpose of the principles themselves: ensuring AI is built and deployed in a way that respects people, reduces harm, and assigns clear responsibility for outcomes.

The other choices are either too absolute or too narrow. “Ensure that AI models deliver consistent and equitable results” is aspirational but not guaranteed—responsible AI practices reduce bias and improve reliability, but they cannot ensure perfect consistency or equity across all contexts. “Help organizations increase trust in fully autonomous AI systems” is not the intent; responsible AI emphasizes oversight and accountability, not pushing toward full autonomy. “Standardize model development practices across AI teams” can be a side benefit of governance, but it’s not the primary reason to adopt responsible AI principles.

In practice, adopting these principles leads to concrete actions: defining governance and accountability, performing fairness and safety evaluations, protecting privacy and security, being transparent with users, and monitoring systems after deployment. These actions build stakeholder trust and enable sustainable adoption, but the foundational reason is that the principles provide a framework for ethical and accountable AI use.

Construction blueprints in PDFs often contain a mix of text, symbols, linework, and diagrams . The requirements include understanding both visual layout (where systems are located) and textual annotations , producing summaries, and answering Q & A. That combination requires a multimodal generative AI approach—models that can reason over images and text together. Therefore, A is best.

OCR alone (B) can extract printed text, but it won’t reliably interpret diagram geometry, symbols, or spatial relationships (e.g., “electrical riser is on the east core near gridline B-4”). Text completion (C) is too generic and doesn’t address image understanding. Document summarization (D) is only one requirement (summary) and still depends on first extracting/understanding both visual and textual elements.

A multimodal solution can ingest the PDF pages as images (or rendered page images) plus extracted text, then answer questions grounded in both modalities. In practice, you may combine OCR and layout extraction with a multimodal LLM so the model can reference drawing regions, legends, callouts, and system diagrams to produce accurate explanations and field extractions.

The correct answer is data security. The principle of least privilege means users, apps, and services receive only the minimum access required to perform their work. Microsoft 365 Copilot respects existing Microsoft 365 permissions and does not grant users new access to content they are not already authorized to view. That behavior is a security control because it helps prevent unauthorized access to company information. Data classification is about labeling or categorizing information by sensitivity. Data compliance focuses on meeting policy, legal, or regulatory requirements. Data loss prevention helps prevent sensitive information from being shared inappropriately. Least privilege is primarily an access-control practice, so the best category is data security.

The correct answer is B. Testing the AI solution to identify and mitigate unfair or inconsistent outcomes directly supports Microsoft responsible AI principles, especially fairness, reliability and safety, accountability, and inclusiveness. Before deployment, organizations should evaluate whether the system behaves consistently across different user groups, avoids discriminatory outcomes, and produces safe and reliable outputs. Automatically approving or rejecting loan applications increases risk unless strong governance and human oversight are in place. Collecting personal data can increase privacy exposure if it is not necessary and controlled. Prioritizing model performance alone is also insufficient because a technically accurate model can still produce biased, unsafe, or non-transparent outcomes. Responsible AI requires evaluation, mitigation, oversight, and monitoring.

By default, Microsoft 365 Copilot is grounded in your organization’s Microsoft 365 data through Microsoft Graph , and it respects the user’s existing permissions. For “find and summarize information across internal documents,” the most direct default document repositories in Microsoft 365 are SharePoint (team sites and shared libraries) and OneDrive (a user’s work files). That is why C (Microsoft SharePoint) and D (Microsoft OneDrive) are the correct selections. Microsoft explicitly describes Copilot as accessing organizational content via Microsoft Graph, including user documents and related work content.

The other options are not “by default” sources. A (on-premises file share) is not automatically part of Microsoft Graph unless you integrate/migrate content or use connectors to make it discoverable in Microsoft 365 experiences. B (custom CRM) similarly requires an integration approach (for example, Microsoft 365 Copilot connectors / Graph connectors) to index and expose that data for Copilot to use. E (Microsoft Sway) is not a primary default content source for Copilot’s document grounding and is not typically referenced as a core internal document repository compared to SharePoint/OneDrive.

So, for immediate, out-of-the-box internal document discovery and summarization within Microsoft 365, SharePoint + OneDrive are the default answer.

Answer Area

Larger datasets can increase the cost of a generative AI solution that uses an Azure Machine Learning workspace. Answer: Yes

The cost of consuming Azure OpenAI models is primarily identified by the number of input and output tokens processed. Answer: Yes

The cost of custom generative AI solutions always remains the same regardless of the model version or capability used. Answer: No

Yes — In Azure Machine Learning, cost is driven by the resources you consume to store, process, and train. Larger datasets typically require more storage, more data transfer, and more compute time for preprocessing, training, evaluation, and experimentation. Even if you are not training foundation models, handling larger corpora can increase pipeline duration and the number/size of compute instances used, which increases overall cost.

Yes — For Azure OpenAI usage under Standard (on-demand), pricing is primarily tied to token-based consumption (input tokens + output tokens). The more context you send and the longer the generated responses, the more tokens you consume, and the higher the cost. This is why prompt optimization, response length controls, and grounding strategies matter for cost management.

No — Costs vary with model choice and capability. Different model families and versions have different price points, and larger/more capable models generally cost more per token or per unit of throughput. Additionally, architecture choices (RAG, vector search, caching), usage patterns, and throughput requirements can significantly change total cost—so it is never “always the same.”

To align an AI customer support agent with Microsoft’s Responsible AI principles, two high-impact actions are fairness/inclusiveness validation and transparency to users . B is correct because testing for inclusive and culturally sensitive responses directly supports fairness and helps reduce harm. In practice, you evaluate responses across diverse user personas, languages/dialects, accessibility scenarios, and sensitive contexts. You look for biased assumptions, stereotyping, exclusionary language, and disparate quality of service. This also implies ongoing monitoring because model behavior can drift as prompts, knowledge sources, and user inputs evolve.

E is correct because a clear disclaimer supports transparency: customers should know they are interacting with an AI system, understand the type of assistance it can provide, and know what to do if the response is incorrect or they need a human. A disclosure is also a practical risk-control that reduces overreliance and sets expectations about limitations.

The other options are not best for Responsible AI alignment: A (retain all conversations) can conflict with privacy/data minimization; retention must be justified and governed, not automatic. C (operate independently) undermines accountability and human oversight. D (multiple purposes) increases scope and risk rather than improving responsible use.

Even when training data is already consistent and uniform, the first step in building a custom Azure Machine Learning model is still to prepare the training data. “Consistent” data reduces the amount of cleaning you may need, but preparation is broader than cleaning: you still must confirm the schema, validate data types, handle missing values (if any), ensure label quality (for supervised learning), select/engineer features, and split data into training/validation/test sets. Those actions determine whether training will be stable and whether evaluation metrics will be meaningful.

If you skip preparation and go directly to training (C), the model might learn from the wrong columns, inconsistent labels, or poorly partitioned data, producing misleading results. Evaluation (B) comes after training because you need a trained model to score and measure. Hyperparameter tuning (D) is an optimization activity that presupposes you already have a working training pipeline and a baseline model to improve. Deployment (E) is last, after you have validated performance and selected the model candidate.

Azure Machine Learning commonly operationalizes these steps through pipelines, where data preparation is a foundational stage that precedes training and evaluation (and can also be iterated as you refine features and quality).

create a Microsoft Word document.

Microsoft 365 Copilot is a productivity-focused AI capability embedded across Microsoft 365 apps such as Word, Excel, PowerPoint, Outlook, and Teams. A core, common use case is drafting and generating documents in Word from prompts and existing context. Therefore, “create a Microsoft Word document” is the best completion.

The other options are not appropriate for Microsoft 365 Copilot’s scope. “Monitor network traffic and alerts in real time” is a security/operations function more aligned to security monitoring tools and, in Microsoft’s product set, closer to Microsoft Security Copilot or dedicated SIEM/SOAR platforms (for example, Microsoft Sentinel). “Modify administrative permissions for SharePoint files” is an admin/governance action that is intentionally protected by role-based access control and is not something end-user Copilot features are meant to perform as a default capability. “Create a list in Microsoft SharePoint” is closer, because Copilot can assist in generating content and can work within Microsoft 365 collaboration contexts; however, the most universally accurate and directly supported capability among the options is creating/drafting a Word document, which is a primary Copilot scenario in Word.

In short, Microsoft 365 Copilot is designed to help users create, summarize, rewrite, and transform content within Microsoft 365 apps, and document creation in Word is a canonical example of that value.

Answer Area

A text-to-image generator can be used to translate content into other languages. Answer: No

A predictive analytics model can generate new marketing content for a company’s online ads. Answer: No

A generative AI chatbot can engage customers in personalized conversations and recommend products. Answer: Yes

No — A text-to-image generator’s primary function is to create images from text prompts , not translate text between languages. Translation is a natural language processing task typically handled by language models or dedicated translation services. A text-to-image model could illustrate translated content (e.g., generate an image based on a translated prompt), but it is not the tool used to perform the translation itself.

No — Predictive analytics models are designed to predict outcomes (forecasts, probabilities, classifications) from historical patterns, such as predicting click-through rate, churn, or next-quarter demand. They are not designed to create new ad copy or marketing creatives. Generating new marketing content is a generative AI capability (text generation), not predictive analytics.

Yes — A generative AI chatbot is well-suited to interactive, natural-language conversations . With access to product catalogs and business rules, it can ask clarifying questions, tailor responses to customer needs, and recommend products (for example, suggesting tents based on group size, season, and budget). This combines conversational generation with retrieval/recommendation logic behind the scenes, enabling personalized customer engagement at scale.

The problem is harmful output language (inappropriate or exclusionary/ableist content). The requirement says you must prevent those responses while minimizing costs . The most cost-effective and direct control is to add a content-moderation filter (B) to screen and block (or rewrite/escalate) responses that violate your safety or inclusion standards. Moderation can be applied at the output stage (and often also at input) without retraining the model, which keeps costs and delivery time low. It also provides an immediate safety layer even if the underlying model occasionally produces biased or exclusionary phrasing.

Option A is not reliable: a newer model version might reduce issues but does not guarantee elimination of ableist language, and you still need policy enforcement. Option C (retraining on only inclusive content) can help, but it is typically expensive (data curation, re-training, re-evaluation, regression testing, re-deployment) and not the “minimize costs” path—also it can reduce coverage/utility if overly restrictive. Option D is clearly wrong because it would amplify the harmful behavior.

In practice, the lowest-cost, high-impact approach is to implement moderation thresholds and handling actions (block, warn, regenerate with constraints, human review) and then, if needed, follow up later with deeper mitigations like prompt constraints, targeted fine-tuning, red-teaming, and continuous evaluation.

Answer Area

A generative AI solution is well-suited to predict next-quarter sales trends. Answer: No

A generative AI solution can summarize lengthy policy documents. Answer: Yes

A generative AI solution can create product descriptions from product specifications. Answer: Yes

No — Predicting next-quarter sales trends is primarily a forecasting/predictive analytics problem. Microsoft differentiates predictive AI (forecasting outcomes from historical patterns) from generative AI (creating content like text, images, or code). While you can use LLMs to assist analysts (explain trends, draft narratives), the core forecasting model is typically traditional ML/time-series methods rather than generative AI as the main engine.

Yes — Summarization is a classic, high-value generative AI capability. Given a long policy, an LLM can compress it into executive summaries, key obligations, risks, and action items, often with formatting constraints (bullets, sections, “do/don’t” lists). Microsoft highlights summarization and analysis as common generative AI use cases in business contexts.

Yes — Generative AI is well-suited to transform structured inputs (features/specs) into natural-language outputs (product descriptions). This is straightforward “content generation,” where you control tone, length, and required fields (benefits, differentiators, disclaimers). Microsoft also points to generating product descriptions and similar marketing/customer-facing text as a practical generative AI scenario.

improves the accuracy and reliability of the predictions and outputs of AI.

High-quality grounding data improves a generative AI solution by anchoring responses to trusted, relevant, and up-to-date information , which increases the likelihood that outputs are accurate, consistent, and aligned with the organization’s expectations. This is why the best completion is “ improves the accuracy and reliability of the predictions and outputs of AI .” When the model is given authoritative context (for example, approved policy text, product specifications, knowledge base articles, or controlled enterprise content), it has less need to “guess” based on general patterns in its training data. That reduces hallucinations and improves response relevance to the user’s question and the business domain.

It does not “ensure that all responses are factually accurate” because grounding reduces errors but cannot eliminate them completely—retrieval can return incomplete or irrelevant passages, user prompts can be ambiguous, and the model can still misinterpret context. It also does not inherently “increase performance of an AI model” in the sense of speed/throughput or model capability; grounding is an architecture and data strategy that improves output quality, not compute efficiency. Finally, grounding is not about “increasing storage required to host an AI model.” While you may store documents in an index or repository, the core benefit is improved response quality through better context, not larger model hosting requirements.

Answer Area

A manufacturer can use Azure Vision in Foundry Tools to identify product defects on an assembly line. Answer: Yes

A logistics company can use Azure Vision in Foundry Tools to recognize package shipping labels. Answer: Yes

The HR department at your company can only use Azure Vision in Foundry Tools to extract written content from Microsoft Word files. Answer: No

Azure Vision in Foundry Tools provides computer vision capabilities to analyze images, including identifying visual features and reading text with OCR. Because it is designed to “analyze images” and support vision scenarios, it can be applied to manufacturing quality inspection use cases where the goal is to detect anomalies/defects from images captured on a production line. This aligns with statement 1 being Yes .

Statement 2 is also Yes because recognizing shipping labels is fundamentally text extraction from images (often plus some layout/field parsing). Azure Vision supports optical character recognition (OCR) to read printed text from images, and Microsoft documentation explicitly notes OCR can extract text from images such as product labels and similar real-world text surfaces—making shipping labels a direct fit.

Statement 3 is No because it is incorrectly restrictive. Azure Vision is not limited to extracting written content from Word documents, nor is OCR restricted to Word files. Vision capabilities apply broadly to images (and, depending on the capability, various document/image inputs) for tasks like image analysis and text recognition. HR could use it for many scenarios such as extracting text from scanned images, photos, or other visual inputs—not “only” Word files.

generative

The sentence describes an AI system that creates new, original content (for example, text, images, audio, code) by learning patterns from large datasets. That is the defining characteristic of generative AI , so “generative” is the correct completion.

To clarify the contrasts: predictive AI primarily uses historical data to forecast outcomes or estimate probabilities (for example, predicting next-quarter sales, predicting equipment failure, or classifying whether a transaction is fraudulent). It focuses on what is likely to happen or which category something belongs to , rather than producing novel content. Prescriptive AI goes a step further by recommending actions or decisions (for example, “increase inventory by X,” “route tickets to team Y,” “schedule maintenance next week”) often using optimization and business constraints.

Generative AI, by comparison, is focused on content synthesis . It identifies statistical patterns and relationships in training data and then generates new sequences that resemble the learned distribution while being responsive to prompts. In business use cases, that translates to drafting emails and documents, summarizing reports, generating meeting notes, creating product descriptions, producing marketing copy, and building chat-based assistants. While all three categories can “recognize patterns,” only generative AI is explicitly characterized by the ability to produce new and original content , which is what the question is testing.

Microsoft 365 Copilot

Microsoft Copilot Studio

Microsoft 365 Copilot is the correct match for creating financial summaries and visual reports from Excel workbooks because it is designed to work inside Microsoft 365 productivity apps, including Excel, and helps users analyze, summarize, and create business content. Microsoft Copilot Studio is the correct match for building an autonomous agent by using a low-code approach because Copilot Studio is Microsoft’s tool for creating, customizing, and managing agents that connect to business data and automate interactions. GitHub Copilot is focused on developer coding productivity, and Copilot in Power Automate is focused on workflow automation rather than building a full low-code conversational or autonomous agent. Therefore, Excel-based business analysis maps to Microsoft 365 Copilot, while low-code agent creation maps to Microsoft Copilot Studio.