Provisioned (PTUs)

For continuous, large-scale workloads, the key needs are predictable throughput, consistent latency, and cost/performance stability . Azure OpenAI Provisioned Throughput Units (PTUs) are designed for this scenario because you reserve model capacity to meet sustained demand. This reduces the risk of variability that can occur with purely on-demand usage during peak periods and provides a more predictable operating model when the workload is always “on” and high volume.

Standard (On-Demand) is best when usage is variable or you’re starting small (PoCs, pilots, spiky workloads), because it is flexible pay-as-you-go but can be less predictable at sustained scale. Batch API is optimized for asynchronous, non-interactive processing where you can tolerate delayed results (for example, large offline summarization jobs), not for always-on, real-time interactions at scale.

Therefore, for continuous high-volume production workloads, Provisioned (PTUs) is the best pricing model choice.

The task is high-volume content generation with consistent structure and immediate publishing: product descriptions that reliably include chosen product attributes (brand, specs, materials, dimensions, use cases) and can be produced quickly by non-technical staff. The best fit is a fine-tuned LLM (D) because fine-tuning can standardize tone, format, and completeness against your catalog schema, reducing variability and minimizing manual editing before posting. With a fine-tuned model, you can strongly enforce style guidelines (length, voice, prohibited claims), and you can template prompts so staff only supply product fields and get publish-ready copy.

Option A is not best: Azure Machine Learning is excellent for predictive models but is unnecessary for straightforward text generation. B (Researcher) is optimized for multistep research across work data + web, not deterministic product copy generation. C (interactive agent) can help collect requirements, but it’s more complexity than needed; the core need is consistent text generation from structured product data, which fine-tuning addresses directly while keeping user interaction simple (fill fields → generate description).

The requirement describes a predictive analytics / machine learning scenario: using historical data to learn patterns and then predict future outcomes . The Microsoft service that directly supports the end-to-end machine learning lifecycle—data preparation, model training, evaluation, deployment, and MLOps—is Azure Machine Learning , which is why C is the best choice. Azure Machine Learning is explicitly designed to help data scientists and engineers train and deploy models and manage the ML project lifecycle, making it the right fit for building a predictive model from your dataset.

The other options focus on different problem classes: Azure Document Intelligence is for extracting structured data from documents (OCR, key-value pairs, tables), not for general predictive modeling. Azure Content Understanding is for deriving structured insights from multimodal content (documents, images, audio, video) into a user-defined schema; it’s not the primary service for training predictive models from a tabular historical dataset. Microsoft Foundry is a broader platform for building AI apps/agents and orchestrating models/tools, but the specific need here is classical ML training and prediction—handled most directly by Azure Machine Learning.

This scenario is a clear violation of the fairness principle. Fairness in Microsoft’s Responsible AI framework is about ensuring AI systems do not create unjustified bias or discriminatory outcomes—especially when decisions affect people’s access to opportunities such as credit, employment, housing, or education. A rule or learned behavior that rejects all applicants over a certain age creates a systematic, categorical disadvantage for a protected demographic group and indicates a discriminatory decision boundary rather than an individualized assessment of creditworthiness.

Even if the model designers believed age correlates with risk, using a hard cutoff that rejects every applicant older than 60 is not an equitable approach. It suggests the model is either using age directly as a dominant feature or reflects biased training data/labels that encoded discriminatory outcomes. Fairness requires you to evaluate model outcomes across groups (for example, age brackets), measure disparate impact, and apply mitigations such as feature review (removing or constraining sensitive attributes), rebalancing training data, adjusting thresholds, or using fairness-aware training/evaluation methods. It also requires governance and review of high-stakes automated decisions.

The other principles are not the best match: transparency concerns explainability and user understanding, accountability concerns human oversight and ownership, and reliability and safety concerns consistent and safe operation. The core issue here is discriminatory treatment across an age group— fairness .

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.

Microsoft Foundry is positioned as a unified platform experience for building, optimizing, and governing AI applications and agents. Microsoft explicitly emphasizes “fleetwide security and governance” and the ability to build and manage AI in a unified environment, which directly supports statement 1 being Yes : it is designed to help organizations build and operate generative AI solutions with centralized governance controls (for example, environment setup, data isolation, access control, and operational management).

For statement 2, Foundry supports scaling as demand increases . Microsoft documentation for Foundry-related model usage notes that as usage grows, Foundry can automatically increase quotas by moving users to higher tiers (and allows requesting additional quota). This is a concrete scalability mechanism tied to increased workload demand, so the statement is Yes .

For statement 3, Foundry is not limited to text-only generative AI. Microsoft provides “Azure Vision in Foundry Tools,” which delivers computer vision capabilities such as analyzing images, reading text (OCR), and other image-processing features. That means Foundry can be used for image recognition/computer vision workloads, so the statement is Yes .

Answer Area

Microsoft Copilot in Outlook can automatically send follow-up replies to emails. Answer: No

Microsoft Copilot in Outlook can identify and surface tasks mentioned in your emails. Answer: Yes

Microsoft Copilot in Outlook can summarize long email threads and highlight action items and deadlines. Answer: Yes

1) No — Copilot in Outlook can draft email responses and help you take actions, but sending messages is not positioned as an automatic “follow-up reply” behavior. Microsoft describes capabilities such as “Draft with Copilot” (generating email drafts from prompts) and Copilot chat actions that update the mailbox, which implies user-directed execution rather than autonomous follow-up sending.

2) Yes — Copilot supports contextual help over your inbox and can help you “catch up” and understand what’s due, which includes surfacing action-oriented items from recent communications. This aligns with using Copilot to pull together summaries of email messages and identify what requires attention.

3) Yes — Microsoft explicitly documents email-thread summarization in Outlook: Copilot scans a conversation and produces a summary of key points. Microsoft also describes Copilot’s summarization capability as extracting key points from threads, which is commonly used to surface what matters most (including actions to take) when the thread is long.

Answer Area

Retrieval Augmented Generation (RAG) architectures require that a large language model (LLM) be retrained on proprietary data. Answer: No

Retrieval-Augmented Generation (RAG) grounds a language model to produce more factual and context-relevant responses. Answer: Yes

Retrieval Augmented Generation (RAG) retrieves information from external knowledge sources at runtime instead of relying solely on the knowledge of a generative AI model. Answer: Yes

1) No — RAG does not require retraining or fine-tuning the base LLM on proprietary data. The defining idea of RAG is to keep the model as-is and instead supply it with relevant context retrieved from trusted sources at inference time. Fine-tuning can be optional for style or specialized behavior, but it is not a requirement for RAG.

2) Yes — RAG is a grounding approach. By retrieving authoritative passages (policies, manuals, product specs, internal knowledge bases) and injecting them into the prompt context, the model’s answer is constrained by evidence that is relevant to the user’s question. This improves factuality and domain relevance and helps reduce hallucinations.

3) Yes — RAG explicitly depends on runtime retrieval from external knowledge sources, such as indexed documents, databases, or enterprise repositories. The retrieval layer finds the best matching content for the query, and the generation layer uses that retrieved content to craft the response. This is why RAG is valuable when information changes frequently: you update the source documents/index rather than retraining the LLM.

Overall, RAG is best understood as an architecture pattern that combines search/retrieval + generation , improving accuracy and freshness without the cost and risk of retraining the underlying model each time the knowledge base changes.

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).

Answer Area

Using incomplete or poor-quality data during generative AI model training can increase costs. Answer: Yes

AI models rely on training data to learn patterns and identify relationships to produce outputs. Answer: Yes

Generative AI models trained on non-representative datasets can produce inaccurate or unbalanced results. Answer: Yes

Yes — Poor-quality or incomplete training data increases cost because it drives more iterations: additional data cleaning, relabeling, re-training, and re-evaluation to reach acceptable performance. It can also increase operational costs after deployment if the model produces low-quality outputs that require human rework, escalations, or incident handling. In practice, data quality debt becomes model cost debt.

Yes — Training data is the primary mechanism by which AI models learn statistical patterns and relationships. For generative models, the training corpus shapes language fluency, factual associations, style tendencies, and the kinds of content the model can produce. Without sufficient and appropriate training signals, outputs degrade.

Yes — If the training dataset is not representative of the real-world population or business context, the model can systematically underperform for certain groups, topics, or edge cases. This can manifest as biased language, missing perspectives, and uneven accuracy, producing “unbalanced” results. That is why Responsible AI practice emphasizes representative data, evaluation across slices, and continuous monitoring.