AI-900 Microsoft Azure AI Fundamentals Questions and Answers

Text extraction.

According to the Microsoft Azure AI Fundamentals (AI-900) study guide and Microsoft Learn documentation for Azure AI Vision (formerly Computer Vision), text extraction—also known as Optical Character Recognition (OCR)—is the computer vision capability that detects and extracts printed or handwritten text from images and video frames.

In this scenario, a traffic monitoring system collects vehicle registration numbers (license plates) from CCTV footage. These registration numbers are alphanumeric text that must be read and converted into digital form for processing, storage, or analysis. The Azure AI Vision service’s OCR (text extraction) feature performs this function. It analyzes each frame from the video feed, detects text regions (the license plates), and converts the visual text into machine-readable text data.

This process is widely used in Automatic Number Plate Recognition (ANPR) systems that support law enforcement, toll booths, and parking management solutions. The OCR model can handle variations in font, lighting, and angle to accurately extract license plate numbers.

The other options describe different vision capabilities:

Image classification assigns an image to a general category (e.g., “car,” “truck,” or “bike”), not text extraction.

Object detection identifies and locates objects in images using bounding boxes (e.g., detecting the car itself), but not the text written on the car.

Spatial analysis tracks people or objects in a defined physical space (e.g., counting individuals entering a building), not reading text.

Therefore, for a traffic monitoring system that identifies vehicle registration numbers from CCTV footage, the most accurate Azure AI Vision capability is Text extraction (OCR).

According to the Microsoft Azure AI Fundamentals (AI-900) Official Study Guide and the Microsoft Learn module “Explore computer vision in Azure,” the Computer Vision service can analyze images to detect objects, landmarks, celebrities, and brands.

The brand detection capability in the Computer Vision Image Analysis API uses pre-trained models to identify well-known brand logos within images. When an image is analyzed, the service returns brand names, confidence scores, and bounding box coordinates where the logos appear.

Let’s examine the other options:

A. Face: Detects and analyzes human faces, not brand logos.

B. Custom Vision: Used for training custom models to recognize unique objects (e.g., company-specific products), not pre-built brand detection.

D. Form Recognizer: Extracts text and data from structured or semi-structured documents like invoices and receipts.

Thus, since the question specifies identifying well-known brands using a pre-built AI model, the correct Azure service is Computer Vision.

Extracting relationships and insights from large volumes of unstructured data (such as documents, text files, or images) aligns with the Knowledge Mining workload in Microsoft Azure AI. According to the Microsoft AI Fundamentals (AI-900) study guide and Microsoft Learn module “Describe features of common AI workloads,” knowledge mining involves using AI to search, extract, and structure information from vast amounts of unstructured or semi-structured content.

In a typical knowledge mining solution, tools like Azure AI Search and Azure AI Document Intelligence work together to index data, apply cognitive skills (such as OCR, key phrase extraction, and entity recognition), and then enable users to discover relationships and patterns through intelligent search. The process transforms raw content into searchable knowledge.

The key characteristics of knowledge mining include:

Using AI to extract entities and relationships between data points.

Applying cognitive skills to text, images, and documents.

Creating searchable knowledge stores from unstructured data.

Hence, B. Knowledge Mining is correct.

The other options—computer vision, NLP, and anomaly detection—deal with image recognition, language understanding, and data irregularities, respectively, not large-scale information extraction.

Object detection is similar to tagging, but the API returns the bounding box coordinates (in pixels) for each object found. For example, if an image contains a dog, cat and person, the Detect operation will list those objects together with their coordinates in the image. You can use this functionality to process the relationships between the objects in an image. It also lets you determine whether there are multiple instances of the same tag in an image.

The Detect API applies tags based on the objects or living things identified in the image. There is currently no formal relationship between the tagging taxonomy and the object detection taxonomy. At a conceptual level, the Detect API only finds objects and living things, while the Tag API can also include contextual terms like "indoor", which can't be localized with bounding boxes.

According to the Microsoft Azure AI Fundamentals (AI-900) official study materials and the Microsoft Learn module “Describe Azure Machine Learning and Automated ML,” Azure Machine Learning designer (formerly known as Azure Machine Learning Studio) is a drag-and-drop, low-code/no-code environment that allows users to create, train, and evaluate machine learning models visually — without the need for extensive programming knowledge.

The designer provides a visual interface, known as the canvas, where users can:

Import and prepare data using modules for data transformation and cleaning.

Split data into training and testing datasets.

Select and configure algorithms (classification, regression, or clustering).

Train and evaluate the model.

Deploy the model as a web service directly from the designer.

The official Microsoft Learn content emphasizes that “Azure Machine Learning designer enables users to build, test, and deploy models by adding and connecting prebuilt modules on a visual interface.” This allows business analysts, data professionals, and beginners to experiment with machine learning workflows without writing code.

By comparison:

Automatically performing common data preparation tasks refers to Automated ML, not the designer.

Automatically selecting an algorithm is also part of Automated ML, which optimizes models algorithmically.

Using a code-first notebook experience applies to Azure Machine Learning notebooks, intended for data scientists familiar with Python and SDKs.

Therefore, as per the AI-900 study guide and Microsoft Learn documentation, the verified and correct answer is:

✅ Adding and connecting modules on a visual canvas, which accurately describes how Azure Machine Learning designer operates.

A transformer model is the foundational architecture behind many modern natural language processing systems such as GPT and BERT. It processes text data through multiple key stages. According to the Microsoft Azure AI Fundamentals (AI-900) curriculum and Microsoft Learn materials, the major stages of a transformer-based large language model are tokenization, embedding calculation, and next token prediction.

Tokenization (C) – The first step converts raw text into smaller units called tokens (words, subwords, or characters). This process allows the model to handle text in a structured numerical form rather than as raw language.

Embedding Calculation (B) – After tokenization, the tokens are mapped into high-dimensional numeric vectors, known as embeddings. These embeddings capture semantic relationships between words and phrases so that the model can understand context and meaning.

Next Token Prediction (D) – This stage is the heart of transformer operation, where the model predicts the next likely token in a sequence based on prior tokens. Repeated next-token predictions enable text generation, summarization, or translation.

Options A (object detection) and E (anonymization) are incorrect because they relate to vision and privacy workflows, not language modeling.

The correct answers are A and D because both scenarios involve extracting structured information from documents, which is exactly what Azure Form Recognizer is designed to do.

According to the Microsoft Azure AI Fundamentals (AI-900) official study guide and Microsoft Learn module “Explore computer vision”, Form Recognizer is an Azure Cognitive Service that uses advanced Optical Character Recognition (OCR) and machine learning to extract key-value pairs, tables, and text from structured and semi-structured documents such as receipts, invoices, business cards, and forms. It allows organizations to automate data entry and digitize document processing.

A. Extract the invoice number from an invoice → CorrectForm Recognizer can identify fields such as invoice number, total amount, date, vendor name, and billing address directly from invoices. It uses prebuilt models for invoices and receipts that automatically detect and extract relevant information without requiring extensive manual labeling. As stated in Microsoft Learn, “Form Recognizer extracts information from documents like receipts and invoices and returns structured data including key-value pairs.”

D. Identify the retailer from a receipt → CorrectThe prebuilt receipt model in Form Recognizer can read printed or scanned receipts and extract data points such as retailer name, transaction date, total amount, and tax information. This makes it ideal for expense reporting, auditing, or financial reconciliation.

The following options are incorrect:

B. Translate a form from French to English → This task involves language translation, which is performed by Azure Translator, not Form Recognizer.

C. Find an image of a product in a catalog → This requires object detection or image classification, which are part of Computer Vision, not Form Recognizer.

Therefore, based on Microsoft’s AI-900 learning objectives and documentation, the two correct scenarios are:

✅ A. Extract the invoice number from an invoice

✅ D. Identify the retailer from a receipt

According to the Microsoft Azure AI Fundamentals (AI-900) official study guide and Microsoft Learn module “Identify features of classification machine learning”, classification is a supervised machine learning technique used when the output variable represents discrete categories or classes. In this case, the brain scan images are already labeled into predefined haemorrhage types, such as “subarachnoid,” “epidural,” or “intraventricular.” The model’s goal is to learn patterns from labeled examples and then predict the correct class for new, unseen images.

The use of categorized brain scan images clearly indicates a supervised learning setup because both the input (image data) and output (haemorrhage type) are known during training. This aligns with Microsoft’s definition: classification problems “predict which category or class an item belongs to,” often using algorithms such as logistic regression, decision trees, neural networks, or convolutional neural networks (CNNs) for image-based data.

In contrast:

A. Clustering is an unsupervised learning approach that groups data into clusters based on similarity when no predefined labels exist.

B. Regression predicts continuous numeric values (e.g., predicting age or temperature), not categories.

Because this project aims to automatically classify medical images into known diagnostic categories, it is a textbook example of classification.

https://docs.microsoft.com/en-us/azure/machine-learning/algorithm-module-reference/split-data

The correct answer is B. Randomly split the data into rows for training and rows for evaluation.

According to the Microsoft Azure AI Fundamentals (AI-900) official study guide and the Microsoft Learn module “Describe fundamental principles of machine learning on Azure”, the process of developing a machine learning model involves dividing the available dataset into two or more parts—commonly training data and evaluation (or testing) data. The goal is to ensure that the model can learn patterns from one subset of the data (training set) and then be objectively tested on unseen data (evaluation set) to measure how well it generalizes to new situations.

The training dataset contains both features (the measurable inputs) and labels (the target outputs). The model learns from the patterns and relationships between these features and labels. The evaluation dataset also contains features and labels, but it is kept separate during the training phase. Once the model has been trained, it is tested on this unseen evaluation data to calculate metrics like accuracy, precision, recall, or F1 score.

Microsoft emphasizes that the data split should be random and based on rows, not columns. Each row represents a complete observation (for example, one customer record, one transaction, or one image). Randomly splitting ensures that both subsets represent the same distribution of data, avoiding bias. Splitting by columns would separate features themselves, which would make the model training invalid.

The AI-900 materials often illustrate this using Azure Machine Learning’s data preparation workflow, where data is randomly divided (commonly 70% for training and 30% for testing). This ensures the model learns from diverse examples and is fairly evaluated.

Therefore, the verified and correct approach, as per Microsoft’s official guidance, is B. Randomly split the data into rows for training and rows for evaluation.

Conversational AI workloads involve human-like dialogue with AI systems.

A: A smart assistant (e.g., smart speaker) uses voice-based conversational AI.

B: A knowledge-based chatbot interacts with users via natural language.Options C and D describe automation/IoT workloads, not conversational AI.

✅ Final Answer (Q110): A and B

These matches are based on the Microsoft Azure AI Fundamentals (AI-900) Official Study Guide and the Microsoft Learn module “Explore Azure Cognitive Services.”

Microsoft Azure provides Cognitive Services that enable developers to integrate artificial intelligence capabilities—such as vision, speech, language understanding, and decision-making—into applications without requiring in-depth AI expertise.

Convert a user’s speech to text → Speech ServiceThe Azure Speech Service supports speech-to-text (STT) conversion, which transcribes spoken language into written text. This feature is commonly used in voice assistants, transcription systems, and voice-enabled apps. The service uses advanced speech recognition models to handle different accents, languages, and background noises.

Identify a user’s intent → Language ServiceThe Azure AI Language Service (which includes capabilities from LUIS – Language Understanding) is used to interpret what a user means or wants to achieve based on their words. It identifies intents (the goal or action behind the input) and entities (key pieces of information) from natural language text. This is a key component in conversational AI applications, allowing chatbots and virtual assistants to respond intelligently.

Provide a spoken response to the user → Speech ServiceThe Speech Service also supports text-to-speech (TTS) functionality, which converts textual responses into natural-sounding speech. This enables applications to communicate audibly with users, completing the conversational loop.

Translator Text is not used here because it’s primarily designed for language translation between different languages, not for speech recognition or intent understanding.

The dataset described contains numeric columns (ColumnA through ColumnE). The model’s task is to predict the value of ColumnE based on the other numeric columns (A–D). This is a classic regression problem.

According to the Microsoft Azure AI Fundamentals (AI-900) study guide and Microsoft Learn module “Identify common types of machine learning,” a regression model is used when the target variable (the value to predict) is continuous and numeric, such as price, temperature, or—in this case—a numerical value in ColumnE.

Regression models analyze relationships between independent variables (inputs: Columns A–D) and a dependent variable (output: ColumnE) to predict a continuous outcome. Common regression algorithms include linear regression, decision tree regression, and neural network regression.

Option analysis:

A. Regression: ✅ Correct. Used for predicting numerical, continuous values.

B. Analysis: ❌ Incorrect. “Analysis” is a general term, not a machine learning model type.

C. Clustering: ❌ Incorrect. Clustering is unsupervised learning, grouping similar data points, not predicting values.

Therefore, the type of machine learning model used to predict ColumnE (a numeric value) from other numeric columns is Regression, which fits perfectly within Azure’s supervised learning models.

According to the Microsoft Azure AI Fundamentals (AI-900) Official Study Guide and the Microsoft Learn module “Explore computer vision in Microsoft Azure,” there are multiple types of computer vision tasks, each designed for different goals such as recognizing, categorizing, or locating objects within an image.

In this question, the application performs two distinct tasks: locating birds within an image and identifying their species. Each of these corresponds to a different type of computer vision capability.

Locate the birds → Object detection

Object detection is used when an AI system needs to identify and locate multiple objects within a single image.

It not only recognizes what the object is but also provides bounding boxes that indicate the exact position of each object.

In this scenario, locating the birds (drawing rectangles around each bird) is achieved through object detection models, such as those available in the Azure Custom Vision Object Detection domain.

Identify the species of the birds → Image classification

Image classification focuses on identifying what is in the image rather than where it is.

It assigns a single label (or multiple labels in multilabel classification) to an entire image based on its contents.

In this case, determining the species of a bird (e.g., robin, eagle, parrot) is achieved through image classification, where the model compares visual features against learned patterns from training data.

Incorrect options:

Automated captioning generates descriptive sentences about an image, not object locations or classifications.

Optical character recognition (OCR) extracts text from images, irrelevant in this case.

Full Detailed Explanation (250–300 words):

The correct answer is A. AI systems should treat people fairly.

This statement aligns with one of Microsoft’s six Responsible AI principles, which are:

Fairness – AI systems should treat all people fairly and avoid bias.

Reliability and Safety

Privacy and Security

Inclusiveness

Transparency

Accountability

The principle of Fairness ensures that AI models do not discriminate based on factors such as race, gender, age, or socioeconomic background. For example, a loan approval or hiring model must provide equal opportunity to all qualified applicants regardless of demographic differences.

B (Not revealing design details) contradicts Transparency, which promotes openness about AI functionality.

C (Black-box models) goes against Microsoft’s push for Explainable AI.

D (Protect developers’ interests) is not part of Microsoft’s Responsible AI framework.

Therefore, the verified correct answer is A. AI systems should treat people fairly.

Azure Custom Vision is a cognitive service that lets you build, deploy, and improve your own image classifiers. An image classifier is an AI service that applies labels (which represent classes) to images, according to their visual characteristics. Unlike the Computer Vision service, Custom Vision allows you to specify the labels to apply.

Note: The Custom Vision service uses a machine learning algorithm to apply labels to images. You, the developer, must submit groups of images that feature and lack the characteristics in question. You label the images yourself at the time of submission. Then the algorithm trains to this data and calculates its own accuracy by testing itself on those same images. Once the algorithm is trained, you can test, retrain, and eventually use it to classify new images according to the needs of your app. You can also export the model itself for offline use.

According to the Microsoft Azure AI Fundamentals (AI-900) official study guide and Microsoft Learn Cognitive Services documentation, developing a voice-controlled personal assistant app involves integrating multiple Azure AI services that specialize in different aspects of language and speech processing. The three services in focus—Azure AI Speech, Azure AI Language Service, and Azure AI Translator Text—perform unique but complementary roles in conversational AI systems.

Convert a user’s speech to text → Azure AI SpeechThe Azure AI Speech service provides speech-to-text (STT) capabilities. It enables applications to recognize spoken language and convert it into written text in real time. This is often the first step in voice-enabled applications, transforming audio input into a machine-readable format that can be analyzed further.

Identify a user’s intent → Azure AI Language serviceOnce speech has been transcribed, the Azure AI Language service (which includes capabilities like Conversational Language Understanding and Text Analytics) interprets the meaning of the text. It detects the user’s intent (what the user wants to accomplish) and extracts entities (key data points) from the input. This service helps the assistant understand commands like “Book a flight” or “Set a reminder.”

Provide a spoken response to the user → Azure AI SpeechAfter determining an appropriate response, the system uses the text-to-speech (TTS) feature of Azure AI Speech to convert the assistant’s text-based reply back into natural-sounding spoken language, allowing the user to hear the response.

Together, these services form the backbone of a conversational AI system: Speech-to-Text → Language Understanding → Text-to-Speech, aligning precisely with the AI-900 curriculum’s explanation of how Azure Cognitive Services enable intelligent voice-based interactions.

As described in the AI-900 study guide under “Identify features of machine learning,” regression is a supervised learning technique used to predict continuous numerical values. In this scenario, the goal is to predict energy output (a continuous variable) based on density (a numeric input).

Regression models find relationships between variables by fitting a line or curve that best represents the trend of the data. In Azure Machine Learning, regression algorithms such as linear regression, decision tree regression, and boosted decision trees are commonly used for such predictions.

Classification (A) predicts discrete labels (e.g., “High” or “Low”), not numeric values.

Clustering (B) groups similar data points but does not perform prediction.

Knowledge mining (C) extracts insights from unstructured data using tools like Azure AI Search and Cognitive Skills.

Hence, based on AI-900 fundamentals, predicting energy based on density requires a regression workload.

facial analysis.

According to the Microsoft Azure AI Fundamentals (AI-900) official study guide and the Microsoft Learn module “Describe features of computer vision workloads on Azure,” facial analysis is a computer vision capability that detects faces and extracts attributes such as facial expressions, emotions, pose, occlusion, and image quality factors like exposure and noise. It does not identify or verify individual identities; rather, it interprets facial features and image characteristics to analyze conditions in an image.

In this question, the AI solution helps photographers take better portrait photos by providing feedback on exposure, noise, and occlusion — tasks directly linked to facial analysis. The model analyzes the detected face to determine if the image is well-lit, clear, and unobstructed, thereby improving photo quality. These capabilities are part of the Azure Face service in Azure Cognitive Services, which includes both facial detection and facial analysis functionalities.

Here’s how the other options differ:

Facial detection only identifies that a face exists in an image and provides its location using bounding boxes, without further interpretation.

Facial recognition goes a step further — it attempts to identify or verify a person’s identity by comparing the detected face with stored images. This is not what the scenario describes.

Thus, when an AI solution evaluates image quality aspects like exposure, noise, and occlusion, it’s performing facial analysis, which focuses on understanding image and facial characteristics rather than identification.

In summary, based on Microsoft’s AI-900 study material, this scenario demonstrates facial analysis, a subcategory of computer vision tasks within Azure Cognitive Services.

✅ Yes – Extract key phrases

❌ No – Generate press releases

✅ Yes – Detect sentiment

The Azure AI Language service is a powerful set of natural language processing (NLP) tools within Azure Cognitive Services, designed to analyze, understand, and interpret human language in text form. According to the Microsoft Azure AI Fundamentals (AI-900) study guide and Microsoft Learn documentation, this service includes several capabilities such as key phrase extraction, sentiment analysis, language detection, named entity recognition (NER), and question answering.

Extract key phrases from documents → YesThe Key Phrase Extraction feature identifies the most relevant words or short phrases within a document, helping summarize important topics. This is useful for indexing, summarizing, or organizing content. For instance, from “Azure AI Language helps analyze customer feedback,” it may extract “Azure AI Language” and “customer feedback” as key phrases.

Generate press releases based on user prompts → NoThis functionality falls under generative AI, specifically within Azure OpenAI Service, which uses models such as GPT-4 for text creation. The Azure AI Language service focuses on analyzing and understanding existing text, not generating new content like press releases or articles.

Build a social media feed analyzer to detect sentiment → YesThe Sentiment Analysis capability determines the emotional tone (positive, neutral, negative, or mixed) of text data, making it ideal for analyzing social media posts, reviews, or feedback. Businesses often use this to gauge customer satisfaction or brand reputation.

In summary, the Azure AI Language service analyzes text to extract insights and detect sentiment but does not generate new textual content.

According to the Microsoft Azure AI Fundamentals (AI-900) official study guide and Microsoft Learn modules on machine learning concepts, ensuring that the accuracy of a predictive model can be proven requires data partitioning—specifically splitting the available data into training and testing datasets. This is a foundational concept in supervised machine learning.

When you split the data, typically about 70–80% of the dataset is used for training the model, while the remaining 20–30% is used for testing (or validation). The reason behind this approach is to ensure that the model’s performance metrics—such as accuracy, precision, recall, and F1-score—are evaluated on data the model has never seen before. This prevents overfitting and allows you to demonstrate that the model generalizes well to new, unseen data.

In the AI-900 Microsoft Learn content under “Describe the machine learning process”, it is explained that after cleaning and transforming the data, the next essential step is data splitting to “evaluate model performance objectively.” By keeping training and testing data separate, you can prove the reliability and accuracy of the model’s predictions, which is particularly crucial in sensitive domains like clinical or healthcare analytics, where decision transparency and validation are vital.

Option A (Train the model by using the clinical data) is incorrect because you should not train and evaluate on the same data—it would lead to biased results.

Option C (Train the model using automated ML) is incorrect because automated ML is a method for training and tuning, but it doesn’t inherently prove accuracy.

Option D (Validate the model by using the clinical data) is also incorrect if you use the same dataset for validation and training—it would not prove true accuracy.

Therefore, per Microsoft’s official AI-900 study content, the verified correct answer is B. Split the clinical data into two datasets.

Box 1: No

Box 2: Yes

Box 3: Yes

Anomaly detection encompasses many important tasks in machine learning:

Identifying transactions that are potentially fraudulent.

Learning patterns that indicate that a network intrusion has occurred.

Finding abnormal clusters of patients.

Checking values entered into a system.

This question evaluates your understanding of how different Azure AI workloads correspond to specific tasks in image, text, and content generation scenarios, as explained in the Microsoft Azure AI Fundamentals (AI-900) study guide and Microsoft Learn modules covering common AI workloads and Azure services.

Generate a caption from a given image → Computer VisionThis is a computer vision task because it involves analyzing the visual elements of an image and producing descriptive text (a caption). Azure AI Vision provides image analysis and captioning capabilities through its Describe Image API, which uses deep learning models to recognize objects, scenes, and actions in an image and automatically generate natural-language descriptions (e.g., “A cat sitting on a sofa”).

Generate an image from a given caption → Generative AIThis task belongs to Generative AI, which focuses on creating new content such as text, code, or images based on prompts. Tools like Azure OpenAI Service with DALL-E can interpret text descriptions and generate realistic images that match the given caption. Generative AI is capable of creative synthesis, not just analysis, making it the appropriate category.

Generate a 200-word summary from a 2,000-word article → Text AnalyticsText analytics (a subset of natural language processing) allows summarization, sentiment analysis, and entity recognition from large text corpora. Azure AI Language includes text summarization capabilities that condense long documents into concise summaries while preserving meaning and key information.

According to the Microsoft Azure AI Fundamentals (AI-900) official study guide and the Microsoft Learn module “Identify features of computer vision workloads”, Optical Character Recognition (OCR) is a core capability within the computer vision domain that enables systems to detect and extract text from images or documents. OCR technology can identify printed or handwritten characters in photographs, scanned documents, or camera feeds, and convert them into machine-readable text.

In this scenario, the task is to read the numbers on runners’ shirts in race photos. These numbers are textual or numeric characters embedded within images. OCR is specifically designed for this purpose — to locate and recognize characters within visual data and convert them into usable text. Once extracted, those numbers can be cross-referenced with a database to identify each runner.

Let’s analyze why the other options are incorrect:

A. Facial recognition focuses on identifying individuals based on unique facial features, not reading text or numbers.

C. Semantic segmentation classifies each pixel of an image into categories (for example, separating road, sky, and people), but it doesn’t read text.

D. Object detection identifies and locates objects within an image (such as detecting people or vehicles) but does not extract readable text or numbers.

Therefore, since the task involves reading textual or numeric content from an image, the appropriate type of computer vision to use is Optical Character Recognition (OCR).

According to the Microsoft Azure AI Fundamentals (AI-900) official study guide and Microsoft Learn documentation for Azure AI Custom Vision, this service is a specialized part of the Azure AI Vision family that enables developers to train custom image classification and object detection models. It allows organizations to build tailored computer vision models that recognize images or specific objects relevant to their business needs.

Detect objects in an image → YesThe Azure AI Custom Vision service supports both image classification (assigning an image to one or more categories) and object detection (identifying and locating objects within an image using bounding boxes). This means it can indeed detect and differentiate multiple objects in a single image, making this statement true.

Requires your own data to train the model → YesThe Custom Vision service is designed to be customizable. Unlike prebuilt Azure AI Vision models that work out of the box, Custom Vision requires you to upload and label your own dataset for training. The model then learns from your examples to perform specialized image recognition tasks relevant to your domain. Thus, this statement is also true.

Analyze video files → NoWhile Custom Vision can analyze images, it does not directly process or analyze video files. Video analysis is handled by a different service—Azure Video Indexer—which can extract insights such as spoken words, scenes, and faces from videos.

In summary:

✅ Yes – Detect objects in images

✅ Yes – Requires your own data

❌ No – Does not analyze video files.

Box1: Sentiment analysis

Sentiment Analysis is the process of determining whether a piece of writing is positive, negative or neutral.

Box 2: Broad entity extraction

Broad entity extraction: Identify important concepts in text, including key

Key phrase extraction/ Broad entity extraction: Identify important concepts in text, including key phrases and named entities such as people, places, and organizations.

Box 3: Entity Recognition

Named Entity Recognition: Identify and categorize entities in your text as people, places, organizations, date/time, quantities, percentages, currencies, and more. Well-known entities are also recognized and linked to more information on the web.

According to the Microsoft Azure AI Fundamentals (AI-900) study material and Azure OpenAI Service documentation, large language models (LLMs) such as GPT are capable of performing multiple natural language processing (NLP) tasks depending on the intent of the prompt. These tasks generally fall into categories like classification, generation, summarization, and translation, each with a distinct purpose and output type.

Classifying – This process involves analyzing text and assigning it to a predefined category or label based on its content. The scenario “Detect the genre of a work of fiction” clearly fits this category. The model must evaluate the text and determine whether it belongs to genres like mystery, romance, or science fiction. This is a classic text classification problem, as the output is a discrete category derived from textual features.

Summarizing – This process means condensing lengthy text into a shorter version that preserves the key information. In the scenario “Create a list of bullet points based on text input,” the model extracts essential information and reformats it as concise bullet points, which is an abstraction form of summarization. Summarization models help users quickly understand the main ideas from long documents, meeting efficiency and readability goals.

Generating – This refers to the LLM’s ability to produce new, creative content based on input instructions. The task “Create advertising slogans from a product description” represents generation because it requires the model to construct original text that didn’t previously exist. Generation tasks showcase the creativity and contextual fluency of models like GPT in marketing and content creation.

Thus, these mappings align directly with the Azure OpenAI LLM capabilities taught in AI-900, linking each NLP process with its most suitable real-world task.

According to the Microsoft Azure AI Fundamentals (AI-900) curriculum and Microsoft Learn’s modules on Computer Vision, object detection is the AI technique used to identify and locate multiple objects within an image. Unlike simple image classification, which only labels an entire image with a single category (for example, “This is a product”), object detection not only identifies the type of object but also pinpoints its exact position by providing bounding boxes and coordinates within the image.

In the scenario described — identifying the location of products on a conveyor belt — the system must be able to detect multiple items simultaneously and determine their spatial positions. Object detection algorithms (such as YOLO, Faster R-CNN, or SSD) are specifically designed for this purpose. This allows automation systems, like robotic arms or quality inspection systems, to track product locations in real time for sorting, packaging, or defect detection.

Let’s evaluate the other options:

Image classification only determines what is in the image, not where it is located. It cannot handle multiple objects or their positions.

Image processing involves operations like resizing, filtering, or adjusting contrast, not understanding object placement.

Optical character recognition (OCR) extracts text from images and documents, unrelated to locating physical items.

Thus, per Microsoft Learn’s AI-900 guidance, object detection is the correct computer vision capability when a task requires both identification and spatial localization of items in an image or video stream.

✅ Final Answer: Object detection

When using generative AI models such as those in Azure OpenAI, it’s critical to monitor and control the model’s outputs to prevent the generation of harmful, hateful, or unsafe content. According to Microsoft’s Responsible AI guidelines and Azure OpenAI safety systems, this is achieved through content filtering mechanisms.

Content filtering automatically analyzes generated text and blocks or flags responses containing hate speech, harassment, sexual content, or violent language. Azure OpenAI applies these filters by default through its built-in content moderation system, which categorizes responses into safety levels (safe, review, block).

Option review:

A. Prompt engineering: Helps guide responses but does not detect hate content automatically.

B. Abuse monitoring: Refers to manual or post-processing review, not real-time automated filtering.

C. Content filtering: ✅ Correct — actively detects and prevents hateful or unsafe responses.

D. Fine-tuning: Adjusts model behavior using additional training data but doesn’t ensure safety monitoring.

According to the Microsoft Azure AI Fundamentals (AI-900) official study materials and the Microsoft Learn module: “Describe features of common AI workloads”, conversational AI solutions like chatbots can be created using various methods—not only through custom code. Azure provides both no-code/low-code and developer-focused approaches. For instance, users can design chatbots using Power Virtual Agents, which requires no programming knowledge, or they can use Azure Bot Service with the Bot Framework SDK for fully customized scenarios. Hence, the statement “Chatbots can only be built by using custom code” is False (No) because Azure supports multiple levels of technical involvement for building bots.

The second statement is True (Yes) because the Azure Bot Service is designed specifically to host, manage, and connect conversational bots to users across different channels. Microsoft Learn explicitly explains that the service provides integrated hosting, connection management, and telemetry for bots built using the Bot Framework or Power Virtual Agents. It acts as the foundation for deploying, scaling, and managing chatbot workloads in Azure.

The third statement is also True (Yes) because Azure Bot Service supports integration with Microsoft Teams, among many other channels such as Skype, Facebook Messenger, Slack, and web chat. Microsoft documentation states that Azure-hosted bots can communicate directly with Teams users through the Teams channel, enabling intelligent virtual assistants within the Teams environment.

A: R-squared (R2), or Coefficient of determination represents the predictive power of the model as a value between -inf and 1.00. 1.00 means there is a perfect fit, and the fit can be arbitrarily poor so the scores can be negative.

C: RMS-loss or Root Mean Squared Error (RMSE) (also called Root Mean Square Deviation, RMSD), measures the difference between values predicted by a model and the values observed from the environment that is being modeled.

Clustering.

According to the Microsoft Azure AI Fundamentals (AI-900) official study guide and Microsoft Learn module “Identify features of common machine learning types”, clustering is an unsupervised machine learning technique used to group data points into distinct segments or clusters based on shared characteristics. Unlike supervised learning (classification or regression), clustering works with unlabeled data, discovering natural groupings without predefined outcomes.

In this question, Recency, Frequency, and Monetary (RFM) values are common marketing metrics used to evaluate customer behavior:

Recency – how recently a customer made a purchase.

Frequency – how often they make purchases.

Monetary – how much money they spend.

Using RFM analysis, a company can segment its customers into groups such as “loyal,” “occasional,” or “at-risk” buyers. This segmentation process does not rely on predefined labels but rather discovers patterns within the data — which is the defining characteristic of clustering.

In the AI-900 context, clustering is described as a method that “groups items with similar features so that items in the same group are more similar to each other than to those in other groups.” Common algorithms used include K-Means, Hierarchical Clustering, and DBSCAN, all available within Azure Machine Learning Designer and other Azure ML environments.

To clarify the incorrect options:

Classification is supervised learning used to predict discrete categories (e.g., yes/no, spam/not spam).

Regression predicts continuous numeric values (e.g., house prices).

Regularization is a model optimization technique, not a type of machine learning.

Therefore, when businesses use RFM values to identify customer segments without labeled outcomes, this is an application of unsupervised learning through clustering.

In Azure OpenAI, grounding refers to the process of connecting the model to external data sources (for example, a database, search index, or API) so that it can retrieve accurate and up-to-date information before generating a response. This is particularly important for scenarios requiring current facts or events, since OpenAI models like GPT-3.5 and GPT-4 are trained on data available only up to a certain cutoff date.

By adding grounding data, the model’s responses are “anchored” to factual sources retrieved at runtime, improving reliability and factual accuracy. Grounding is commonly implemented in Azure OpenAI + Azure Cognitive Search solutions (Retrieval-Augmented Generation or RAG).

Option review:

A. Modify the system message: Changes model tone or behavior but doesn’t supply real-time data.

B. Add grounding data: ✅ Correct — allows access to recent and domain-specific information.

C. Add few-shot learning: Provides examples in the prompt to improve context understanding but not factual accuracy.

D. Add training data: Refers to fine-tuning; this requires retraining and doesn’t update the model’s awareness of current events.

Hence, the best method to ensure accurate and current responses from an Azure OpenAI model is to add grounding data, enabling the model to reference real, updated sources dynamically.

According to the Microsoft Azure AI Fundamentals (AI-900) official study materials and Microsoft Learn module “Identify features of common AI workloads,” the Azure Form Recognizer service is part of Azure Cognitive Services for Document Intelligence. It enables organizations to extract, analyze, and identify information from structured and unstructured documents, including sensitive or confidential data such as names, addresses, financial figures, and identification numbers.

Form Recognizer uses optical character recognition (OCR) combined with machine learning to automatically extract key-value pairs, tables, and text fields from documents like invoices, receipts, contracts, and forms. It can be customized to identify and classify documents that contain specific sensitive data, allowing businesses to automate compliance and data governance tasks.

By contrast:

A. Custom Vision is used for image classification and object detection — it analyzes visual data, not document content.

B. Conversational Language Understanding (formerly LUIS) identifies intent and entities in text conversations, not document structure or sensitive data.

Form Recognizer is explicitly mentioned in the AI-900 course as the tool for document analysis and extraction. It can even integrate with Azure Cognitive Search or Azure Purview for further data management and compliance workflows.

Therefore, the verified and correct answer, aligned with Microsoft’s official training content, is C. Form Recognizer, as it is the Azure Cognitive Service capable of identifying and processing documents containing sensitive information.

For evaluating a classification model, the appropriate metric from the options provided is the True Positive Rate (TPR), also known as Sensitivity or Recall. According to the Microsoft Azure AI Fundamentals (AI-900) official study guide and Microsoft Learn module “Evaluate model performance”, classification models are evaluated using metrics that measure how accurately the model predicts categorical outcomes such as “yes/no,” “spam/not spam,” or “approved/denied.”

The True Positive Rate measures the proportion of correctly identified positive cases out of all actual positive cases. Mathematically, it is expressed as:

True Positive Rate (Recall)=True PositivesTrue Positives + False Negatives\text{True Positive Rate (Recall)} = \frac{\text{True Positives}}{\text{True Positives + False Negatives}}True Positive Rate (Recall)=True Positives + False NegativesTrue Positives​

This metric is important when missing positive predictions carries a high cost, such as in medical diagnosis or fraud detection. Microsoft Learn highlights classification evaluation metrics such as accuracy, precision, recall, F1 score, and AUC (Area Under the Curve) as suitable for classification models.

The other options—Mean Absolute Error (MAE), Root Mean Squared Error (RMSE), and Coefficient of Determination (R²)—are regression metrics used to evaluate models that predict numeric values rather than categories. For example, they apply to predicting house prices or temperatures, not yes/no decisions.

Therefore, the correct classification evaluation metric among the choices is A. True Positive Rate.

The correct answers are B. Azure REST API and C. Azure SDK for Python.

The Azure OpenAI Service can be accessed using multiple development interfaces. According to Microsoft Learn documentation, developers can call the service via the Azure REST API, which provides direct HTTPS-based access to the model endpoints for tasks like completions, chat, embeddings, and image generation. This interface is platform-independent and supports integration with any system capable of making HTTP requests.

Additionally, Azure SDKs offer higher-level libraries for convenient integration into applications. The Azure SDK for Python and Azure SDK for JavaScript are both supported for Azure OpenAI interaction, allowing developers to authenticate with Azure credentials, send prompts, and receive model responses programmatically.

However, among the listed options, the REST API (B) and SDK for Python (C) are most explicitly referenced in the AI-900 learning modules and Microsoft documentation as standard tools to call Azure OpenAI services.

Option A (Azure CLI) is incorrect because the CLI is used primarily for provisioning and managing Azure resources, not for directly calling OpenAI model endpoints.

Therefore, based on the Azure AI-900 and OpenAI integration guidance, the correct answers are B. Azure REST API and C. Azure SDK for Python.

According to the Microsoft Azure AI Fundamentals (AI-900) Official Study Guide and the Microsoft Learn module “Explore conversational AI in Microsoft Azure,” the QnA Maker (now integrated into the Azure AI Language Service as Custom Question Answering) is used to create, train, and publish a knowledge base of question-and-answer pairs that can power a chatbot.

There are three primary methods to create Q&A content:

Generate questions and answers from an existing webpage (Option A):QnA Maker can automatically extract question–answer pairs from structured or semi-structured data sources like FAQs, product manuals, or support webpages.

Manually enter questions and answers (Option C):Users can create Q&A pairs directly in the QnA Maker portal or Azure Language Studio, enabling custom answers to be crafted manually.

Import chit-chat content from a predefined data source (Option E):QnA Maker provides predefined “chit-chat” datasets that let a bot handle casual conversation (e.g., greetings or small talk) naturally.

The other options are incorrect:

B. Use automated machine learning – AutoML is for predictive modeling, not knowledge extraction.

D. Connect the bot to Cortana – This is a channel integration, not a method of content creation.

In this scenario, analyzing vehicle images and measuring the distance between them requires first detecting each vehicle’s position in the image. Object detection models can locate and identify multiple objects (such as cars, trucks, or motorcycles) by assigning bounding boxes. Once detected, their coordinates can be used to calculate distances or spacing.

Image classification only assigns a single label per image, not per object. Facial recognition is human-focused, and OCR deals with text extraction. Thus, object detection is the correct model type for this task.

Within Azure AI Document Intelligence (formerly Form Recognizer), the Contract prebuilt model is designed to extract key information from legal and business contracts, including parties, jurisdictions, dates, and terms. According to Microsoft Learn, this prebuilt model identifies structured entities such as contracting parties, effective dates, governing jurisdictions, and termination clauses.

Layout (B) extracts text, tables, and structure but does not identify semantic information such as parties or jurisdictions.

General document (C) extracts key-value pairs and entities but lacks domain-specific contract analysis.

Read (D) performs OCR (optical character recognition) to extract raw text but not contextual metadata.

Thus, when the requirement is to extract parties and jurisdictions from a legal document, the Contract model is the correct Azure AI Document Intelligence choice.

In the Microsoft Azure AI Fundamentals (AI-900) curriculum, computer vision workloads are grouped into distinct types, each serving a specific purpose. The three major workloads illustrated here are image classification, object detection, and optical character recognition (OCR). Understanding their use cases is essential for correctly mapping them to real-world scenarios.

Generate captions for images → Image classificationThe image classification workload is used to identify the main subject or context of an image and assign descriptive labels. In Microsoft Learn’s “Describe features of computer vision workloads,” image classification models are trained to recognize content (e.g., a cat, a beach, or a city). Caption generation expands on classification results by describing the image’s contents in human-readable language—based on what the model identifies as key visual features.

Extract movie title names from movie poster images → Optical character recognition (OCR)OCR is a vision workload that detects and extracts text from images. Azure AI Vision’s Read API or Document Intelligence OCR models can identify printed or handwritten text within posters, signs, or documents. In this case, the movie title text from a poster is best extracted using OCR.

Locate vehicles in images → Object detectionThe object detection workload identifies multiple objects within an image and provides their locations using bounding boxes. It’s ideal for tasks like counting cars in a parking lot or tracking objects in traffic images.

Statement

Yes / No

Providing an explanation of the outcome of a credit loan application is an example of the Microsoft transparency principle for responsible AI.

Yes

A triage bot that prioritizes insurance claims based on injuries is an example of the Microsoft reliability and safety principle for responsible AI.

Yes

An AI solution that is offered at different prices for different sales territories is an example of the Microsoft inclusiveness principle for responsible AI.

No

This question is based on the Responsible AI principles defined by Microsoft, which are part of the AI-900 Microsoft Azure AI Fundamentals curriculum. Microsoft’s Responsible AI framework consists of six key principles: Fairness, Reliability and Safety, Privacy and Security, Inclusiveness, Transparency, and Accountability. Each principle ensures that AI systems are developed and used in a way that benefits people and society responsibly.

Transparency Principle – YesProviding an explanation for a loan decision aligns with the Transparency principle. Microsoft defines transparency as helping users and stakeholders understand how AI systems make decisions. For example, when a credit scoring AI model approves or denies a loan, explaining the factors that influenced that outcome (such as credit history or income level) ensures that customers understand the reasoning process. This builds trust and supports responsible deployment.

Reliability and Safety Principle – YesA triage bot that prioritizes insurance claims based on injury severity relates directly to Reliability and Safety. This principle ensures AI systems operate consistently, perform accurately, and produce dependable outcomes. In the case of the triage bot, it must reliably assess the input data (injury descriptions) and rank claims appropriately to avoid harm or misjudgment, aligning with Microsoft’s emphasis on designing AI systems that are safe and robust.

Inclusiveness Principle – NoAn AI solution priced differently across sales territories is not related to Inclusiveness. Inclusiveness focuses on ensuring accessibility and eliminating bias or exclusion for all users—especially those with disabilities or underrepresented groups. Pricing strategy is a business decision, not an inclusiveness issue. Therefore, this statement is No.

In summary, based on the AI-900 Responsible AI principles, the correct selections are:

According to the Microsoft Azure AI Fundamentals (AI-900) official study guide and Microsoft Learn module “Describe features of machine learning on Azure”, a classification model outputs a probability score representing how likely each input belongs to a particular class. To decide whether a prediction is “positive” or “negative,” the model applies a threshold (often defaulted to 0.5). Adjusting this threshold directly affects the balance between false positives and false negatives.

A false positive occurs when the model incorrectly predicts a positive outcome (for example, predicting that a patient has a disease when they do not).

A false negative occurs when the model fails to predict a true positive (for example, predicting that a patient does not have a disease when they actually do).

To reduce false positives, you must make the model less likely to classify borderline cases as positive. This is done by increasing the decision threshold, thereby favoring false negatives (because the model will only classify a case as positive when the prediction confidence is very high). In other words, by moving the threshold upward, you tighten the model’s standard for what qualifies as a “positive” prediction, reducing incorrect positives.

Let’s review why other options are incorrect:

A. Include test data in training data: This contaminates your dataset and causes overfitting, which leads to unreliable performance metrics.

B. Increase the number of training iterations: This may improve learning but doesn’t specifically target false positives.

C. Modify the threshold in favor of false positives: That would increase, not reduce, false positives.

Therefore, the correct step to reduce false positives is to adjust the threshold in favor of false negatives, making the model more conservative when labeling a case as positive — hence, Answer: D.

This question tests the understanding of features and labels in machine learning, a core concept covered in the Microsoft Azure AI Fundamentals (AI-900) syllabus under “Describe fundamental principles of machine learning on Azure.”

In supervised machine learning, data is divided into features (inputs) and labels (outputs).

Features are the independent variables — measurable properties or characteristics used by the model to make predictions.

Labels are the dependent variables — the target outcome the model is trained to predict.

From the provided dataset, the goal of the Azure Machine Learning model is to predict product quality (Pass or Fail). Therefore:

Mass (kg) is a feature – Yes“Mass (kg)” represents an input variable used by the model to learn patterns that influence product quality. It helps the algorithm understand how variations in mass might correlate with passing or failing the quality test. Thus, it is correctly classified as a feature.

Quality Test is a label – YesThe “Quality Test” column indicates the outcome of the manufacturing process, marked as either Pass or Fail. This is the target the model tries to predict during training. In Azure ML terminology, this column is the label, as it represents the dependent variable.

Temperature (C) is a label – No“Temperature (C)” is an input that helps the model determine quality outcomes, not the outcome itself. It influences the quality result but is not the value being predicted. Therefore, temperature is another feature, not a label.

In conclusion, per Microsoft Learn and AI-900 study materials, features are measurable inputs (like mass and temperature), while the label is the target output (like the quality test result).

To prevent hateful or harmful responses from being returned by a generative AI solution, the correct approach is to implement content filtering. According to the Microsoft Learn documentation for Azure OpenAI Service and the Responsible AI principles, content filtering is a built-in safety mechanism that automatically screens both user prompts (inputs) and model outputs (responses) for inappropriate, harmful, or policy-violating material.

Content filters are designed to detect and block content such as:

Hate speech or harassment

Sexual or explicit material

Self-harm or violent content

Personally identifiable information (PII) misuse

In Azure OpenAI, the content filtering system is part of Microsoft’s Responsible AI standard and cannot be disabled. It ensures that generative AI models such as GPT-3.5 or GPT-4 operate safely and ethically, reducing the risk of producing offensive or discriminatory text. The filter evaluates model responses in real time and can modify, block, or flag inappropriate outputs before they reach the user.

Let’s review the other options:

A. Abuse monitoring tracks misuse after deployment but does not actively prevent hateful responses.

C. Fine-tuning customizes a model’s style or domain knowledge but does not guarantee filtering of offensive content.

D. Prompt engineering helps steer model behavior but cannot fully prevent harmful outputs.

Therefore, to proactively prevent hateful, unsafe, or offensive responses in a generative AI system built on Azure OpenAI, the correct and Microsoft-verified approach is B. Content filtering.

 You can communicate with a bot by using email → No

 You can communicate with a bot by using Microsoft Teams → Yes

 You can communicate with a bot by using a webchat interface → Yes

These answers are based on the Microsoft Azure AI Fundamentals (AI-900) Official Study Guide and the Microsoft Learn module “Explore conversational AI in Microsoft Azure.”

The Azure Bot Service allows developers to build, test, deploy, and manage intelligent chatbots that can interact with users through various channels. Channels are communication platforms or interfaces that connect users to bots. Once a bot is built and published through the Azure Bot Service, it can be connected to multiple channels such as Microsoft Teams, webchat, Skype, Facebook Messenger, Direct Line, Slack, and others.

Let’s evaluate each statement:

You can communicate with a bot by using email → NoAzure Bot Service does not support direct interaction via email as a channel. Bots are designed for real-time or conversational interactions through messaging or voice-based platforms, not asynchronous email communication.

You can communicate with a bot by using Microsoft Teams → YesMicrosoft Teams is one of the primary channels supported by Azure Bot Service. Bots can be integrated directly into Teams to handle chat-based conversations, provide information, automate workflows, or assist users interactively within Teams.

You can communicate with a bot by using a webchat interface → YesThe Web Chat channel is another core feature of Azure Bot Service. It allows embedding the bot into a website or web application using the Web Chat control or the Direct Line API, enabling users to chat directly from a browser interface.

In summary, Azure Bot Service supports real-time conversational interfaces like Teams and webchat, but not email.

This question tests your understanding of machine learning workflow tasks as described in the Microsoft Azure AI Fundamentals (AI-900) study guide and the Microsoft Learn module “Explore the machine learning process.” The AI-900 curriculum divides the machine learning lifecycle into key phases: data preparation, feature engineering and selection, model training, model evaluation, and model deployment. Each phase has specific tasks designed to prepare, build, and assess predictive models before deployment.

Examining the values of a confusion matrix → Model evaluationIn Azure Machine Learning, evaluating a model involves checking its performance using metrics such as accuracy, precision, recall, and F1-score. The confusion matrix is one of the most common tools for this purpose. According to Microsoft Learn, “model evaluation is the process of assessing a trained model’s performance against test data to ensure reliability before deployment.” Analyzing the confusion matrix helps determine whether predictions align with actual outcomes, making this task part of model evaluation.

Splitting a date into month, day, and year fields → Feature engineeringFeature engineering refers to transforming raw data into features that better represent the underlying patterns to improve model performance. The study guide describes it as “the process of creating new input features from existing data.” Splitting a date field into separate numeric fields (month, day, year) is a classic example of feature engineering because it enables the model to learn from temporal patterns that might otherwise remain hidden.

Picking temperature and pressure to train a weather model → Feature selectionFeature selection involves identifying the most relevant variables that have predictive power for the model. As defined in Microsoft Learn, “feature selection is the process of choosing the most useful subset of input features for training.” In this scenario, selecting temperature and pressure variables as inputs for a weather prediction model fits perfectly within the feature selection stage.

Therefore, the correct matches are:

✅ Examining confusion matrix → Model evaluation

✅ Splitting date field → Feature engineering

✅ Picking temperature & pressure → Feature selection

According to the Microsoft Azure AI Fundamentals (AI-900) syllabus and Microsoft Learn module “Describe features of natural language processing (NLP) workloads on Azure,” translation is a core NLP workload that converts text from one language into another while maintaining meaning and context.

In this scenario, the website stores reviews in English and must present them in the user’s native language based on geographical location. This directly requires a translation workload, which uses Azure Cognitive Services — specifically, the Translator service — to automatically translate content dynamically for each user.

Other options explained:

B. Language modeling involves predicting the next word in a sentence or understanding linguistic patterns; it’s used in model training, not translation.

C. Key phrase extraction identifies main ideas in text, not language conversion.

D. Speech recognition converts spoken words into written text but does not perform translation or handle geographic adaptation.

Microsoft’s Translator service supports real-time text translation, multi-language detection, and context preservation, making it ideal for global websites. The AI-900 study guide emphasizes translation as one of the most common NLP workloads, enabling applications to break language barriers and enhance accessibility for diverse audiences.

Therefore, based on official Microsoft Learn material, the correct answer is:

✅ A. translation.

Statements

Yes

No

A bot that responds to queries by internal users is an example of a conversational AI workload.

✅ Yes

An application that displays images relating to an entered search term is an example of a conversational AI workload.

✅ No

A web form used to submit a request to reset a password is an example of a conversational AI workload.

✅ No

According to the Microsoft Azure AI Fundamentals (AI-900) official study materials, conversational AI workloads are those that enable interaction between humans and AI systems through natural language conversation, either by text or speech. These workloads are typically implemented using Azure Bot Service, Azure Cognitive Services for Language, and Azure OpenAI Service. The key characteristic of a conversational AI workload is the presence of dialogue—the AI interprets user intent and provides a meaningful, contextual response in a conversation-like manner.

“A bot that responds to queries by internal users is an example of a conversational AI workload.” → YESThis fits the definition perfectly. A chatbot that helps employees (internal users) by answering questions about policies, IT issues, or HR procedures is a typical example of conversational AI. It uses natural language understanding to interpret questions and provide automated responses. Microsoft Learn explicitly identifies chatbots as conversational AI solutions designed for both internal and external interactions.

“An application that displays images relating to an entered search term is an example of a conversational AI workload.” → NOThis is not conversational AI because there is no dialogue or language understanding involved. It is an example of information retrieval or computer vision if it uses image recognition, but not conversation.

“A web form used to submit a request to reset a password is an example of a conversational AI workload.” → NOA password reset form is a simple UI-driven process that doesn’t require AI or conversational logic. It performs a fixed function based on user input but does not understand or respond to natural language.

Therefore, based on the AI-900 study guide, only the first statement is an example of a conversational AI workload, while the second and third statements are not.

This question assesses knowledge of the Azure Cognitive Services Speech and Text Analytics capabilities, as described in the Microsoft Azure AI Fundamentals (AI-900) official study guide and Microsoft Learn modules “Explore natural language processing” and “Explore speech capabilities.” These services are part of Azure Cognitive Services, which provide prebuilt AI capabilities for speech, language, and text understanding.

You can use the Speech service to transcribe a call to text → YesThe Speech-to-Text feature in the Azure Speech service automatically converts spoken words into written text. Microsoft Learn explains: “The Speech-to-Text capability enables applications to transcribe spoken audio to text in real time or from recorded files.” This makes it ideal for call transcription, voice assistants, and meeting captioning.

You can use the Text Analytics service to extract key entities from a call transcript → YesOnce a call has been transcribed into text, the Text Analytics service (part of Azure Cognitive Services for Language) can process that text to extract key entities, key phrases, and sentiment. For example, it can identify names, organizations, locations, and product mentions. Microsoft Learn notes: “Text Analytics can extract key phrases and named entities from text to derive insights and structure from unstructured data.”

You can use the Speech service to translate the audio of a call to a different language → YesThe Azure Speech service also includes Speech Translation, which can translate spoken language in real time. It converts audio input from one language into translated text or speech output in another language. Microsoft Learn describes this as: “Speech Translation combines speech recognition and translation to translate spoken audio to another language.”

According to the Microsoft Azure AI Fundamentals (AI-900) and Azure Machine Learning documentation, regression is the appropriate machine learning task when the goal is to predict continuous numeric values—such as prices, sales amounts, or other measurable quantities.

In this scenario, the dataset contains motor vehicle sales data, and the objective is to predict vehicle sale values. A vehicle’s sale value (price) is a continuous numeric variable, meaning it can take on a wide range of possible numbers (for example, $15,000, $28,500, $42,300). Regression models are designed to analyze relationships between input features (like make, model, mileage, age, fuel type) and a continuous output variable (price).

Automated Machine Learning (AutoML) in Azure simplifies this process by automatically testing multiple regression algorithms (e.g., Linear Regression, Random Forest, Gradient Boosted Trees) and hyperparameters to find the best-performing model.

Let’s evaluate why the other options are incorrect:

Classification – Used for predicting discrete categories (e.g., car type: sedan, SUV, or truck). Sale value prediction is not categorical.

Time series forecasting – Used when predicting future values based on time-dependent data (e.g., sales over months or years). The question focuses on predicting price based on vehicle characteristics, not over time.

Natural Language Processing (NLP) – Deals with text-based data, not numeric vehicle data.

Computer Vision – Applies to image-based tasks (e.g., detecting car types from photos).

Therefore, per Microsoft Learn’s Automated ML task selection guidance, when predicting a numeric output like vehicle sale value, the correct machine learning task type is Regression.

✅ Final Answer: Regression

In Azure OpenAI Service, fine-tuning refers to the process of uploading your own labeled dataset to customize or adapt a pretrained model (such as GPT-3.5 or Curie) for a specific use case. According to the Microsoft Learn documentation and AI-900 official study guide, fine-tuning allows organizations to improve a model’s performance on domain-specific tasks or to align responses with brand tone and context.

Fine-tuning differs from simple prompting because it requires providing structured training data (usually in JSONL format) that contains pairs of input prompts and ideal completions. The model uses this data to adjust its internal weights, thereby “learning” your organization’s language patterns, terminology, or industry context.

Option review:

A. Completion: Refers to the text generated by a model in response to a prompt. It’s the output, not the customization process.

B. Grounding: Integrates external, up-to-date data sources (like search results or databases) during inference but doesn’t alter the model’s parameters.

C. Fine-tuning: ✅ Correct — this is the process of uploading and training with your own data.

D. Prompt engineering: Involves designing effective prompts but does not change the underlying model.

Thus, fine-tuning is the term used for customizing an Azure OpenAI model using your own uploaded data.

Box 1: 11

TP = True Positive.

The class labels in the training set can take on only two possible values, which we usually refer to as positive or negative. The positive and negative instances that a classifier predicts correctly are called true positives (TP) and true negatives (TN), respectively. Similarly, the incorrectly classified instances are called false positives (FP) and false negatives (FN).

Box 2: 1,033

FN = False Negative

To generate cartoons or images from text descriptions, the correct Azure OpenAI model is DALL-E. As described in Microsoft’s OpenAI integration documentation, DALL-E is a generative image model that converts natural language prompts into images, illustrations, and artwork.

Codex is for code generation, GPT-3.5 and GPT-4 are for text and reasoning tasks, not image creation. Therefore, B. DALL-E is correct.

According to the Microsoft Azure AI Fundamentals (AI-900) official study guide and Microsoft Learn module “Describe features of conversational AI workloads on Azure,” the Azure Bot Service is the dedicated Azure service for building, connecting, deploying, and managing conversational AI experiences across multiple channels — such as Microsoft Teams, Cortana, and Amazon Alexa.

The Azure Bot Service integrates with the Bot Framework SDK to design intelligent chatbots that can communicate with users in natural language. It also connects seamlessly with other Azure Cognitive Services, such as Language Service (LUIS) for intent understanding and Speech Service for voice input/output.

The question specifies that the conversational AI must be accessible through multiple platforms, including Microsoft Teams, Cortana, and Alexa. Azure Bot Service supports this multi-channel communication model out of the box, allowing developers to configure a single bot that interacts through many endpoints simultaneously.

Other options:

B. Azure Cognitive Search: Used for information retrieval and knowledge mining, not conversational AI.

C. Language Service: Provides natural language understanding, key phrase extraction, sentiment analysis, etc., but doesn’t handle multi-channel communication.

D. Speech: Provides speech-to-text and text-to-speech conversion but is not a chatbot platform.

Therefore, the best solution for building and deploying a multi-channel conversational AI system is Azure Bot Service, as clearly defined in Microsoft’s AI-900 learning content.

The correct mappings are based on the Microsoft Azure AI Fundamentals (AI-900) curriculum topic: “Describe features of computer vision workloads.” Microsoft divides computer vision tasks into key workload types — image classification, object detection, facial recognition, and optical character recognition (OCR) — each designed for specific visual analysis objectives.

Identify celebrities in images → Facial recognitionFacial recognition goes beyond simple face detection; it can identify or verify specific individuals by comparing facial features with known profiles. According to Microsoft Learn, the Face service in Azure Cognitive Services can detect, recognize, and identify people in photos or videos. Recognizing celebrities or known individuals is a prime example of facial recognition.

Extract movie title names from movie poster images → Optical Character Recognition (OCR)OCR is used to detect and extract text content from images, such as printed or handwritten words. Azure’s Computer Vision API uses OCR technology to read text in various languages from photos, scanned documents, or posters. Therefore, extracting movie titles or actor names from a poster image is a perfect use case for OCR.

Locate vehicles in images → Object detectionObject detection identifies and locates specific objects within an image, returning bounding boxes that indicate their positions. In Azure, the Custom Vision service or Computer Vision object detection models are used to detect multiple objects like vehicles, pedestrians, or animals in a single image.

Summary:

Facial recognition → Identifies specific people (celebrities)

OCR → Extracts text (movie titles)

Object detection → Finds and locates physical items (vehicles)

Thus, the verified and official answer is:

The Azure OpenAI DALL-E model is a generative image model designed to create original images from textual descriptions (prompts). According to the Microsoft Learn documentation and the AI-900 study guide, DALL-E’s primary function is text-to-image generation—it converts creative or descriptive text input into visually relevant imagery.

“Generate captions for uploaded images” → NoDALL-E cannot create image captions. Captioning an image (describing what’s in an uploaded image) is a vision analysis task, not an image generation task. That functionality belongs to Azure AI Vision, which can analyze and describe images, detect objects, and generate captions automatically.

“Reliably generate technically accurate diagrams” → NoWhile DALL-E can create visually appealing artwork or conceptual sketches, it is not designed for producing precise or technically correct diagrams, such as engineering schematics or architectural blueprints. The model’s generative process emphasizes creativity and visual diversity rather than factual or geometric accuracy. Thus, it cannot be relied upon for professional technical outputs.

“Generate decorative images to enhance learning materials” → YesThis is one of DALL-E’s strongest use cases. It can generate decorative, conceptual, or illustrative images to enhance presentations, educational materials, and marketing content. It enables educators and designers to quickly produce unique visuals aligned with specific themes or topics, enhancing engagement and creativity.

According to the Microsoft Azure AI Fundamentals (AI-900) Official Study Guide and the Microsoft Learn module “Explore fundamental principles of machine learning,” a regression model is used when the goal is to predict a continuous numerical value based on historical data.

In this question, the task is to predict the sale price of auctioned items, which is a numeric output that can take on a wide range of values (for example, $50.25, $199.99, etc.). This makes it a regression problem because the output is continuous rather than categorical.

Regression models analyze the relationship between input features (such as item type, condition, age, bidding history, or demand) and a numerical target variable (the sale price). Common regression algorithms include linear regression, decision tree regression, and neural network regression. In Azure Machine Learning, these models are trained using labeled datasets containing known outcomes to learn patterns and make future predictions.

Let’s review the incorrect options:

Classification: Used to predict discrete categories or labels, such as “sold” vs. “unsold” or “low,” “medium,” “high.” It cannot output continuous numeric predictions.

Clustering: An unsupervised technique used to group similar data points based on shared characteristics, not to predict specific numeric outcomes.

Therefore, because predicting a sale price involves forecasting a continuous numerical value, the correct model type is Regression.

This aligns with Microsoft’s AI-900 teaching that regression is used for tasks such as:

Predicting house prices

Forecasting sales revenue

Estimating car values or auction prices

In Microsoft Azure AI Language Service, both Named Entity Recognition (NER) and Key Phrase Extraction are core features for text analytics. They serve distinct purposes in analyzing and structuring unstructured text data.

Named Entity Recognition (NER):NER is used to identify and categorize specific entities within text, such as people, organizations, locations, dates, times, and quantities. According to Microsoft Learn’s “Analyze text with Azure AI Language” module, NER scans text to extract these entities along with their types. Therefore, the statement “Named entity recognition can be used to retrieve dates and times in a text string” is True (Yes).

Key Phrase Extraction:This feature identifies the most important phrases or main topics in a block of text. It is useful for summarization or highlighting central ideas without classifying them into specific categories. Therefore, the statement “Key phrase extraction can be used to retrieve important phrases in a text string” is also True (Yes).

City Name Retrieval:While key phrase extraction highlights major phrases, it does not extract specific entities like cities or dates. Extracting such details requires Named Entity Recognition, which is designed to find named entities such as city names, people, or organizations. Hence, the statement “Key phrase extraction can be used to retrieve all the city names in a text string” is False (No).

According to Microsoft’s Responsible AI Principles, Transparency means that AI systems should clearly communicate how they operate, including data sources, limitations, and decision-making processes. Stating the source of data used to train a model helps users understand where the model’s knowledge comes from, enabling informed trust and accountability.

Transparency ensures that organizations disclose relevant details about data collection and model design, especially for compliance, fairness, and reproducibility.

Other options are incorrect:

A. Fairness: Focuses on avoiding bias and ensuring equitable outcomes.

C. Reliability and safety: Ensures AI performs consistently and safely.

D. Privacy and security: Protects user data and maintains confidentiality.

Thus, the principle illustrated by disclosing training data sources is Transparency.

For an app that helps employees write emails and reports based on user prompts, you need a text generation model capable of understanding natural language instructions and producing coherent, contextually appropriate output. Azure OpenAI GPT models—available through Azure AI Foundry (formerly Azure OpenAI Studio)—are specifically designed for such generative tasks.

By integrating GPT-3.5 or GPT-4, the app can analyze prompts like “Write a professional email to a client about project updates” and automatically generate polished text in seconds.

The other options do not fit:

A. Azure AI Speech: Converts spoken language to text or text to speech; not suitable for generating written content.

C. Azure AI Vision: Processes and analyzes images or video content.

D. Azure Machine Learning Studio: Used for training, testing, and deploying custom ML models, not directly for content generation.

Therefore, to create a writing-assistance app for emails and reports, the correct solution is B. Azure OpenAI in Foundry Models using GPT-based language generation.

The correct matches are Custom Vision, Form Recognizer, and Face — each corresponding to a distinct capability under Azure Cognitive Services as described in the Microsoft Azure AI Fundamentals (AI-900) official study guide and Microsoft Learn modules on Computer Vision workloads.

Custom Vision → Identify objects in an imageThe Custom Vision service is part of the Azure Cognitive Services suite that enables developers to train custom image classification and object detection models. Unlike the prebuilt Computer Vision API, Custom Vision allows users to upload their own labeled images and teach the model to recognize specific objects relevant to their business context. The AI-900 syllabus explains that Custom Vision is ideal for tasks such as identifying products on a shelf, categorizing images, or detecting defects in manufacturing.

Form Recognizer → Automatically import data from an invoice to a databaseForm Recognizer is a document processing AI service that extracts structured data from forms, receipts, and invoices. It uses optical character recognition (OCR) combined with layout and key-value pair detection to automatically capture information such as invoice numbers, amounts, and vendor names. The AI-900 study materials highlight this service under the Document Intelligence category, emphasizing its ability to streamline data entry and business automation workflows by importing extracted data directly into databases or applications.

Face → Identify people in an imageThe Face service provides advanced facial detection and recognition capabilities. It can locate faces in images, compare similarities between faces, identify known individuals, and even detect facial attributes such as age or emotion. The AI-900 course classifies this under Computer Vision services for person identification and security-related use cases such as access control or identity verification.

Thus, each mapping aligns precisely with the AI-900 official learning outcomes on Cognitive Services capabilities:

Custom Vision → Object recognition

Form Recognizer → Data extraction from forms

Face → People identification

✅ Final verified configuration:

Custom Vision → Identify objects in an image

Form Recognizer → Automatically import data from an invoice to a database

Face → Identify people in an image

Description

Correct Service

Enables the use of natural language to query a knowledge base.

QnA Maker

Enables the real-time transcription of speech-to-text.

Speech

This question tests understanding of Azure Cognitive Services and their use cases as outlined in the Microsoft Azure AI Fundamentals (AI-900) study guide.

“Enables the use of natural language to query a knowledge base.” → QnA MakerAccording to Microsoft Learn’s AI-900 module “Identify features of Natural Language Processing (NLP) workloads and services,” QnA Maker is a cloud-based service that allows developers to build a question-and-answer layer over structured or unstructured content. It enables users to ask questions in natural language, and the service retrieves the most relevant answer from a knowledge base (such as FAQs, manuals, or documents).QnA Maker uses Natural Language Processing (NLP) techniques to interpret user intent and return an appropriate response. It is often integrated into chatbots built with Azure Bot Service to make them capable of conversational question-answering. In the newer Azure Cognitive Services lineup, QnA Maker capabilities are merged into Azure Cognitive Service for Language (Question Answering).

“Enables the real-time transcription of speech-to-text.” → SpeechThe Azure Speech service (part of Azure Cognitive Services) provides the ability to convert spoken language into written text in real time. This feature, called Speech-to-Text, uses deep neural network models to recognize and transcribe human speech with high accuracy.Microsoft’s AI-900 documentation specifies that Speech service capabilities also include text-to-speech, speech translation, and speaker recognition. Real-time transcription is widely used in applications such as voice assistants, captioning systems, call analytics, and accessibility tools.

Other listed services such as Azure Storage and Language Understanding (LUIS) serve different purposes:

Azure Storage handles data storage, not AI workloads.

LUIS identifies user intent from natural language but does not query knowledge bases directly.

The correct matching aligns directly with the Microsoft Azure AI Fundamentals (AI-900) official study guide and Microsoft Learn modules under “Identify features of Azure Machine Learning”. Azure Machine Learning provides a suite of tools that serve different functions within the model development lifecycle — from creating workspaces, to training models, to automating experimentation.

The Azure portal → Create a Machine Learning workspace.The Azure portal is a web-based graphical interface for managing all Azure resources. According to Microsoft Learn, you use the portal to create and configure the Azure Machine Learning workspace, which acts as the central environment where datasets, experiments, models, and compute resources are organized. Creating a workspace through the portal involves specifying a subscription, resource group, and region — tasks that are part of the setup stage rather than model development.

Machine Learning designer → Use a drag-and-drop interface used to train and deploy models.The Machine Learning designer (formerly “Azure ML Studio (classic)”) provides a visual, no-code/low-code interface for building, training, and deploying machine learning pipelines. The designer uses a drag-and-drop workflow where users connect modules representing data transformations, model training, and evaluation. This tool is ideal for beginners and those who want to quickly experiment with machine learning concepts without writing code.

Automated machine learning (Automated ML) → Use a wizard to select configurations for a machine learning run.Automated ML simplifies model creation by automatically selecting algorithms, hyperparameters, and data preprocessing options. Users interact through a guided wizard (within the Azure Machine Learning studio) that walks them through configuration steps such as selecting datasets, target columns, and performance metrics. The system then iteratively trains and evaluates multiple models to recommend the best-performing one.

Together, these tools streamline the machine learning workflow:

Azure portal for setup and resource management,

Machine Learning designer for visual model creation, and

Automated ML for guided, automated model selection and tuning.

According to the Microsoft Learn documentation for Azure AI Document Intelligence (formerly Form Recognizer) and the AI-900 official study materials, the prebuilt receipt model in Azure AI Document Intelligence supports analyzing image and PDF files up to a maximum size of 10 MB per document.

Azure AI Document Intelligence is a cloud-based service that applies advanced optical character recognition (OCR) and machine learning to extract structured information from documents such as receipts, invoices, identity cards, and business forms. The prebuilt receipt model is specifically designed to extract key data fields from retail receipts—such as merchant name, transaction date, items purchased, subtotal, tax, and total—without requiring users to build or train a custom model.

As per Microsoft’s service limits, the input file for the prebuilt models (including receipts, invoices, business cards, and identity documents) must:

Not exceed 10 MB in file size.

Not exceed 17 x 17 inches (43 x 43 cm) in physical dimensions.

Be in a supported image or document format such as JPG, PNG, BMP, TIFF, or PDF.

Let’s examine why other options are incorrect:

A. 5 MB → Too small; the service allows up to 10 MB.

C. 50 MB and D. 100 MB → Exceed the official maximum file size supported by Azure AI Document Intelligence.

Therefore, when using the prebuilt receipt model, you must ensure that the input file is 10 MB or smaller to be successfully processed by the service.

This question tests understanding of AI workload types, a fundamental topic in the Microsoft Azure AI Fundamentals (AI-900) curriculum. Each workload type—Computer Vision, Natural Language Processing, Machine Learning (Regression), and Anomaly Detection—serves a specific function within the AI landscape, as explained in Microsoft Learn’s module “Describe features of common AI workloads.”

Computer Vision enables computers to “see” and interpret visual information such as images or videos. Identifying handwritten letters requires analyzing image patterns, shapes, and strokes, which is a classic image recognition task. Azure’s Computer Vision API and Custom Vision services are specifically designed for such tasks.

Natural Language Processing (NLP) involves interpreting human language, both written and spoken. Determining the sentiment of a social media post (positive, negative, or neutral) is a typical text analytics use case within NLP, often implemented using Azure’s Text Analytics for Sentiment Analysis.

Anomaly Detection focuses on identifying data points that deviate from normal patterns. Detecting fraudulent credit card payments requires finding transactions that are unusual compared to historical spending behavior. Azure’s Anomaly Detector API applies machine learning to identify such irregularities.

Machine Learning (Regression) is used for predicting continuous numerical outcomes based on historical data. Estimating next month’s toy sales is a regression problem—an example of supervised learning where the model predicts future sales values from past sales data.

Thus, based on Microsoft’s official AI-900 learning objectives, the correct mapping of workloads to scenarios is:

Computer Vision → Identify handwritten letters

NLP → Predict sentiment

Anomaly Detection → Fraud detection

Machine Learning (Regression) → Predict toy sales

Yes, No, Yes.

According to the Microsoft Azure AI Fundamentals (AI-900) official study guide and the Microsoft Learn module “Identify capabilities of Azure Cognitive Services for Language”, the Azure Translator service is a cloud-based machine translation service used to translate text or entire documents between languages in real time. It uses REST APIs or client libraries to translate text input, detect languages, and support multiple target languages in a single request.

“The following service call will accept English text as an input and output Italian and French text: /translate?from=en&to=it,fr – Yes.This URL format is correct because the Translator service API allows multiple target languages to be specified in a single to parameter separated by commas. In this case, from=en defines the source language (English), and to=it,fr requests translations into Italian (it) and French (fr). The API would return results in both target languages simultaneously. This syntax is officially documented in Microsoft Learn as the valid format for multi-language translation.

“The following service call will accept English text as an input and output Italian and French text: /translate?from=en&to=fr&to=it – No.This format is incorrect, as the Translator API does not support repeating the to parameter multiple times. Only one to parameter is valid, and multiple target languages must be provided as a comma-separated list within the same to parameter.

“The Translator service can be used to translate documents from English to French.” – Yes.This statement is true. The Translator service supports both text translation and document translation. The document translation capability allows the translation of whole files such as Word, PowerPoint, or PDF documents while preserving formatting and structure. This feature is included in the official Translator API under “Document Translation.”

In summary, the AI-900 study content clarifies that:

✅ /translate?from=en&to=it,fr → Valid syntax

❌ /translate?from=en&to=fr&to=it → Invalid syntax

✅ Translator can translate full documents between languages

According to the Microsoft Azure AI Fundamentals (AI-900) Official Study Guide and the Microsoft Learn module “Explore conversational AI in Microsoft Azure,” conversational bots are AI applications that can understand and respond to natural language inputs through text or speech. Building such a bot typically involves two key Azure services:

Azure Bot Service (Option B):This service provides the framework and infrastructure needed to create, test, and deploy intelligent chatbots that interact with users across multiple channels (webchat, Teams, email, etc.). It handles conversation flow, integration, and user message management.

Azure Language Service (Option A):This service powers the natural language understanding (NLU) capability of the bot. It enables the bot to interpret user input, extract intent, and query a knowledge base using Question Answering (formerly QnA Maker). This allows the bot to respond intelligently to user questions by finding the most relevant answers.

The other options are incorrect:

C. Form Recognizer is used for extracting structured data from documents like invoices or forms.

D. Anomaly Detector is used for identifying unusual patterns in time-series data.

Hence, to build a bot that understands and answers user questions in natural language, the solution must combine Azure Bot Service for conversation management and Azure Language Service for knowledge-based question answering and natural language understanding.

According to the Microsoft Azure AI Fundamentals (AI-900) official study materials and the Microsoft Learn module “Identify features of common machine learning types,” the evaluation phase occurs after training and testing a machine learning model. Evaluation metrics are used to measure how well the model performs when applied to data it has not seen before (the validation data).

The machine learning workflow includes the following key steps:

Data Preparation – Importing, cleaning, and transforming data.

Splitting the Data – Dividing it into training and validation (or test) sets.

Model Training – Using the training data to teach the model patterns or relationships.

Model Evaluation – Assessing the trained model using the validation data and evaluation metrics such as accuracy, precision, recall, F1 score, and root mean square error (RMSE).

As stated in the AI-900 content, evaluation metrics are crucial after testing, as they help determine if the model is accurate enough or if it requires retraining with different parameters or algorithms.

A. After you clean the data → incorrect, as metrics cannot be reviewed before training.

B. Before you train a model → incorrect, since the model has not yet learned patterns.

C. Before you choose the type of model → incorrect, as metrics depend on the model’s output.

Therefore, the verified answer is D. After you test a model on the validation data, which is when you review evaluation metrics to determine model performance and readiness for deployment.

When training a machine learning model, it is standard practice to randomly split the dataset into training and testing subsets. The purpose of this is to evaluate how well the model generalizes to unseen data. According to the AI-900 study guide and Microsoft Learn module “Split data for training and evaluation”, this ensures that the model is trained on one portion of the data (training set) and evaluated on another (test or validation set).

The correct answer is C. to test the model by using data that was not used to train the model.

Random splitting prevents data leakage and overfitting, which occur when a model memorizes patterns from the training data instead of learning generalizable relationships. By testing on unseen data, developers can assess true performance, ensuring that predictions will be accurate on future, real-world data.

Options A and B are incorrect because:

A. Train the model twice does not improve accuracy; model accuracy depends on data quality, feature engineering, and algorithm choice.

B. Train multiple models simultaneously refers to model comparison, not the purpose of splitting data.

Thus, the correct reasoning is that random splitting provides a reliable estimate of the model’s predictive power on new data.

Safety system.

According to the Microsoft Learn documentation and the AI-900: Microsoft Azure AI Fundamentals official study guide, the safety system layer in generative AI architecture plays a crucial role in monitoring, filtering, and mitigating harmful or unsafe model outputs. This layer works alongside the model and user experience layers to ensure that generative AI systems—such as those powered by Azure OpenAI—produce responses that are safe, aligned, and responsible.

The safety system layer uses various techniques including content filtering, prompt moderation, and policy enforcement to prevent outputs that could be harmful, biased, misleading, or inappropriate. It evaluates both user inputs (prompts) and model-generated outputs to identify and block unsafe or unethical content. The system might use predefined rules, classifiers, or human feedback signals to decide whether to allow, modify, or stop a response.

In contrast, the other layers serve different purposes:

The model layer contains the core large language or generative model (e.g., GPT or DALL-E) that processes inputs and produces outputs.

The metaprompt and grounding layer ensures the model’s responses are contextually relevant and factually supported, often linking to organizational data sources or system prompts.

The user experience layer defines how users interact with the AI system, including the interface and conversational flow, but does not manage safety enforcement.

Therefore, the layer that uses system inputs and context to mitigate harmful outputs from a generative AI model is the Safety system layer.

This aligns with Microsoft’s responsible AI principles—Fairness, Reliability and Safety, Privacy and Security, Inclusiveness, Transparency, and Accountability—ensuring generative AI operates ethically and safely.

According to the Microsoft Azure AI Fundamentals (AI-900) curriculum, computer vision encompasses several key capabilities: image classification, object detection, facial detection, and optical character recognition (OCR). When the task requires counting the number of distinct objects (in this case, animals) in an image, object detection is the correct type of vision model.

Object detection not only classifies what is present in an image but also identifies where each object appears by drawing bounding boxes around them. Each detected object is individually labeled, enabling the system to count or track them accurately. In contrast, image classification would only tell you the overall category (e.g., “This is an image of animals”) without counting how many animals are present.

Facial detection focuses solely on identifying human faces, while OCR extracts text from images — neither applies here.

Therefore, the AI-900 official learning modules confirm that object detection is the appropriate solution for identifying and counting multiple entities within an image.

The service that can complete a paragraph based on a sentence is Azure OpenAI. According to Microsoft Learn’s AI-900 study guide, Azure OpenAI provides access to advanced language models like GPT-3.5 and GPT-4, which can generate and continue text, summarize, or create content based on prompts. The task described—text completion—is precisely what GPT models are designed for.

Azure AI Language performs language understanding and analysis (sentiment, key phrases, translation), Azure Machine Learning trains custom models, and Azure AI Vision handles images. Hence, Azure OpenAI is the correct choice.

This question is derived from the Microsoft Azure AI Fundamentals (AI-900) learning module, particularly under “Describe features of conversational AI workloads on Azure.” It tests understanding of chatbot capabilities and design principles within the context of Azure Bot Service and Conversational AI.

Chatbots can support voice input – YesAccording to the AI-900 official materials, conversational AI systems such as chatbots can interact with users through text or voice. Using speech recognition services like Azure Cognitive Services Speech-to-Text, bots can interpret spoken input, and with Text-to-Speech, they can respond verbally. This enables voice-based chatbots used in virtual assistants, call centers, and customer support. Hence, voice input is fully supported by conversational AI solutions in Azure.

A separate chatbot is required for each communication channel – NoThe Azure Bot Service is designed to provide multi-channel communication from a single bot instance. A single chatbot can communicate across several channels such as Microsoft Teams, Web Chat, Slack, Facebook Messenger, and email without needing separate bots for each platform. This centralized design allows developers to create, deploy, and manage one bot while configuring multiple channel connections through the Azure portal. Therefore, the statement is false.

Chatbots manage conversation flows by using a combination of natural language and constrained option responses – YesIn Microsoft’s AI-900 training, chatbots are described as using Natural Language Processing (NLP) to understand free-form user input while also guiding interactions with predefined options such as buttons or quick replies. This hybrid approach ensures both flexibility and control, improving user experience and accuracy. Bots can interpret natural language via services like Language Understanding (LUIS) and also present structured options to guide conversations efficiently.

According to the Microsoft Azure AI Fundamentals (AI-900) official study guide and Microsoft Learn module “Extract insights from text with the Text Analytics service”, key phrase extraction is a feature of the Text Analytics service that identifies the most important words or phrases in a given document. It helps summarize the main ideas by isolating significant concepts or terms that describe what the text is about.

In this scenario, the goal is to determine which documents share similar topics or themes. By extracting key phrases from each document (for example, “policy renewal,” “coverage limits,” “claim process”), you can compare and categorize documents based on overlapping keywords. This is exactly how key phrase extraction is used—to summarize and group text content by topic relevance.

The other options do not fit this use case:

A. Translation uses the Translator service, not key phrase extraction.

B. Generating video captions involves speech recognition and computer vision.

C. Identifying sentiment relates to sentiment analysis, not key phrase extraction.

According to the Microsoft Azure AI Fundamentals (AI-900) official study guide and the Microsoft Learn module “Describe features of conversational AI workloads,” the correct answer is Power Virtual Agents. This service is part of the Microsoft Power Platform and is specifically designed to enable users to create intelligent chatbots without writing any code.

Power Virtual Agents (PVA) provides a no-code/low-code environment where business users and developers can collaboratively design conversational experiences. It integrates built-in natural language processing (NLP) models to understand user intent and respond intelligently to text or speech inputs. The platform’s interface allows chatbot creators to design dialogues visually, connect to back-end data via Power Automate, and publish bots on websites, Teams, or other communication channels.

This approach is highlighted in Microsoft Learn as an ideal solution for organizations that want to deploy conversational bots quickly without requiring specialized AI or programming expertise. PVA automatically leverages Microsoft’s language understanding models, allowing it to interpret user input and map it to predefined topics or actions.

Let’s analyze the other options:

Azure Health Bot: A specialized solution for the healthcare industry that provides prebuilt medical compliance and healthcare content. It is not a general-purpose, no-code chatbot builder.

Microsoft Bot Framework: A developer-focused SDK for building highly customized bots through code, offering maximum flexibility but not no-code functionality.

Therefore, the most appropriate choice that “can be used to build no-code apps that use built-in natural language processing models” is Power Virtual Agents — the official Microsoft no-code chatbot solution for conversational AI workloads.

The correct answer is Document Intelligence.

According to the Microsoft Azure AI Fundamentals (AI-900) study materials and Microsoft Learn documentation, the Azure AI Document Intelligence service (formerly known as Form Recognizer) is specifically designed to extract structured data from documents, including scanned invoices, receipts, forms, and business cards.

This service combines optical character recognition (OCR) with machine learning to analyze both the layout and semantic meaning of document content. When processing scanned invoices, Document Intelligence identifies and extracts fields such as invoice numbers, dates, totals, taxes, vendor names, and line-item details. The extracted information can then be automatically imported into business systems like accounting software or databases, eliminating manual data entry and improving operational efficiency.

Here’s why the other options are incorrect:

Generative AI: Focuses on creating new content such as text, images, or code (for example, using GPT-4 or DALL·E). It is not used for structured data extraction.

Natural Language Processing (NLP): Deals with understanding and generating human language from text-based input, not document scanning or layout interpretation.

The Document Intelligence workload excels at handling semi-structured documents where the location and format of data vary between samples. Microsoft’s prebuilt models—like Invoice, Receipt, Identity Document, and Contract—simplify extraction without requiring custom training.

In summary, if the task involves extracting data from scanned invoices, the appropriate Azure AI service is Azure AI Document Intelligence, which uses AI-powered document understanding to convert unstructured document images into structured, usable data.

QnA Maker supports automatic extraction of question-and-answer pairs from structured files such as PDF, Microsoft Word, or Excel documents, as well as from public webpages. This makes PDF the correct file format for populating a knowledge base.

Other options are invalid:

B. PPTX – Not supported.

C. XML – Not a recognized input for QnA Maker.

D. ZIP – Used for packaging, not Q&A content.

This question maps each Azure AI service to its correct capability based on the Microsoft Azure AI Fundamentals (AI-900) syllabus and Microsoft Learn documentation on Azure Cognitive Services.

Classify and label images → Azure AI VisionAzure AI Vision (formerly Computer Vision) provides capabilities to analyze visual content, detect objects, classify images, and extract information from pictures. It includes object detection, image classification, and tagging, which are core vision tasks. This service enables businesses to build solutions that understand visual input, such as identifying products, reading signs, or detecting faces in images.

Generate conversational responses → Azure OpenAI ServiceAzure OpenAI Service integrates powerful large language models such as GPT-3.5 and GPT-4, capable of generating human-like text responses, summarizations, translations, and dialogues. These models are designed for natural language generation (NLG) and conversational AI, making them ideal for chatbots, virtual agents, and intelligent assistants that produce dynamic, context-aware replies.

Convert speech to text in real time → Azure AI SpeechAzure AI Speech provides speech-to-text capabilities (speech recognition) that convert spoken language into written text instantly. It is commonly used in transcription services, voice command systems, and live captioning applications. Additionally, the Speech service supports text-to-speech (speech synthesis) and speech translation, making it versatile for voice-based AI applications.

By understanding each service’s specialization—Vision for visual data, OpenAI for generative text, and Speech for audio processing—you can correctly match the capabilities.

In Azure Machine Learning designer, a low-code drag-and-drop interface, users can visually build machine learning workflows. According to the AI-900 study guide and Microsoft Learn module “Create and publish models with Azure Machine Learning designer”, two key components that can be dragged onto the designer canvas are datasets and modules.

Datasets (A): These are collections of data that serve as the input for training or evaluating models. They can be registered in the workspace and then dragged onto the canvas for use in transformations or model training.

Modules (D): These are prebuilt processing and modeling components that perform operations such as data cleaning, feature engineering, model training, and evaluation. Examples include “Split Data,” “Train Model,” and “Evaluate Model.”

Compute (B) and Pipeline (C) are not drag-and-drop items within the designer. Compute targets are infrastructure resources used to run the pipeline, while a pipeline represents the overall workflow, not a component that can be added like a dataset or module.

Hence, the correct answers are A. Dataset and D. Module.

According to the Microsoft Azure AI Fundamentals (AI-900) official study guide and Microsoft Learn documentation on Azure AI Vision services, different Azure AI Vision capabilities are suited for different use cases such as object detection, brand recognition, facial recognition, and spatial analysis.

App1: Uses a set of images and photos to extract brand names → Image AnalysisThe Azure AI Vision – Image Analysis service (formerly part of Computer Vision) can detect and extract brands, objects, text, and other visual features from images. It uses advanced image classification and object detection models to recognize logos and identify brand names (for example, “Microsoft” or “Coca-Cola”) in photos. The Image Analysis API can also return descriptive tags, scene descriptions, and confidence scores. Therefore, since App1 analyzes static images to extract brand names, it specifically relies on the Image Analysis feature of Azure AI Vision.

App2: Enables touchless access control for buildings → FaceThe Azure AI Face service is designed for facial detection, verification, and identification. It can recognize and match faces in real time, making it ideal for access control, identity verification, and attendance tracking systems. A “touchless access control” system uses a camera to detect a person’s face and verify identity against a stored profile, allowing or denying entry without physical interaction.

The other options are not suitable:

Optical Character Recognition (OCR) extracts text, not brand logos.

Spatial Analysis is for detecting movement or presence in video feeds.

Video Analysis is for analyzing dynamic video content rather than still images.

“When evaluating the performance of a model, the confusion matrix displays the predicted and actual positives and negatives by using a grid of 0 and 1 values.”

According to the Microsoft Azure AI Fundamentals (AI-900) official study materials and Microsoft Learn module “Identify features of common machine learning types”, a confusion matrix is a tool used to evaluate the performance of classification models. It visually summarizes how many predictions were correctly or incorrectly classified by comparing the predicted labels to the actual (true) labels.

A confusion matrix is a table, typically 2×2 for binary classification, with the following components:

True Positives (TP): The model correctly predicted the positive class.

True Negatives (TN): The model correctly predicted the negative class.

False Positives (FP): The model incorrectly predicted the positive class.

False Negatives (FN): The model incorrectly predicted the negative class.

The confusion matrix allows data scientists and analysts to derive important performance metrics such as accuracy, precision, recall, and F1-score, which together provide a more complete understanding of how well a model performs beyond a single number.

In Microsoft Learn’s AI-900 curriculum, the confusion matrix is highlighted as a key visualization tool that “compares actual values to predicted values to evaluate classification performance.” The grid format (using 0s and 1s for predicted classes) helps identify where misclassifications occur.

By contrast:

AUC metric (Area Under Curve) and ROC curve evaluate model discrimination ability.

Threshold defines decision cutoffs but doesn’t display classifications.

Therefore, based on the official Microsoft AI-900 study guide and Microsoft Learn resources, the correct answer is Confusion Matrix, as it provides a grid view comparing actual versus predicted values in classification models.

According to the Microsoft Azure AI Fundamentals (AI-900) Official Study Guide and the Microsoft Learn module “Explore fundamental principles of machine learning,” anomaly detection is a specialized machine learning technique used to identify data points, patterns, or events that deviate significantly from normal behavior.

Anomaly detection is widely used for monitoring time-series data and detecting unexpected or rare occurrences that may indicate problems, opportunities, or fraud. For example:

Detecting fraudulent transactions in banking systems.

Identifying equipment malfunctions in industrial IoT applications.

Monitoring network intrusions in cybersecurity.

Detecting unexpected spikes or drops in web traffic or sales.

In Azure, this workload is supported by the Azure AI Anomaly Detector service, which uses statistical and machine learning algorithms to learn from historical data and establish a baseline of normal behavior. When the system detects data points that fall outside expected patterns, it flags them as anomalies.

Let’s evaluate the incorrect options:

A. A machine learning technique that understands written and spoken language → This describes Natural Language Processing (NLP), not anomaly detection.

B. A machine learning technique that classifies objects based on user-supplied images → This refers to image classification, typically using computer vision.

D. A machine learning technique that classifies images based on their contents → Also describes computer vision, not anomaly detection.

Therefore, the correct answer is C, since anomaly detection specifically refers to analyzing data over time and identifying unusual or abnormal patterns that differ from the expected trend.

According to the Microsoft Azure AI Fundamentals (AI-900) official study materials and Microsoft Learn module “Describe features of Computer Vision workloads on Azure”, the Azure AI Custom Vision service allows users to build, train, and deploy custom image classification or object detection models. It is specifically designed for scenarios where you need a model tailored to your unique image dataset — in this case, personal digital photographs.

Custom Vision lets you upload and label your own images (for example, “family,” “friends,” “vacations,” or “pets”) and then train a model that learns to recognize those categories. The system automatically extracts relevant features from the training images and creates a model capable of classifying new images into the predefined labels. You can iteratively refine your model by adding more images or re-training to improve accuracy.

The other options do not fit this requirement:

A. Azure AI Language deals with text-based tasks such as language understanding, sentiment analysis, and key phrase extraction — not image processing.

B. Azure AI Computer Vision provides prebuilt image analysis models (e.g., object detection, tag generation, scene description), but it cannot learn custom categories unique to your dataset. It’s great for general image recognition but not for specialized labeling tasks.

C. Azure AI Document Intelligence (Form Recognizer) is used to extract information from structured or semi-structured documents such as forms, invoices, or receipts — not photographs.

Therefore, when you need to label or categorize personal photos with custom-defined tags, the right service is Azure AI Custom Vision, because it allows you to build a model trained specifically on your own collection of images.

According to the Microsoft Azure AI Fundamentals (AI-900) Official Study Guide and the Microsoft Learn module “Explore fundamental principles of machine learning,” regression is a type of supervised machine learning used to predict continuous numeric values.

In this question, the goal is to predict how many vehicles will travel across a bridge on a given day. The predicted output (the number of vehicles) is a continuous value—meaning it can take on any numerical value depending on various factors like time, weather, or day of the week. This makes it a regression problem, as the model learns from historical numeric data to estimate a continuous outcome.

How Regression Works:

Regression models find patterns between input features (such as temperature, weekday/weekend, traffic trends) and a numerical output (number of vehicles). Common regression algorithms include linear regression, decision trees for regression, and neural network regression. In Azure Machine Learning, regression tasks are used for business scenarios such as:

Predicting sales revenue for a future month.

Estimating house prices based on property characteristics.

Forecasting energy consumption or traffic flow, as in this case.

Why not the other options?

Classification: Used for predicting discrete categories (e.g., “spam” vs. “not spam”). It does not handle continuous numeric values.

Clustering: An unsupervised learning technique used to group data points based on similarity without predefined labels (e.g., segmenting customers into groups).

Therefore, the task of predicting the number of vehicles—a numeric, continuous value—is a regression problem.

The Translator service, part of Microsoft Azure Cognitive Services, is designed specifically for text translation between multiple languages. It is a cloud-based neural machine translation service that supports more than 100 languages. According to Microsoft Learn’s module “Translate text with the Translator service”, this service provides two main capabilities: text translation and automatic language detection.

“You can use the Translator service to translate text between languages.” → YesThis statement is true. The primary purpose of the Translator service is to translate text accurately and efficiently between supported languages, such as English to Spanish or French to Japanese. It maintains contextual meaning using neural machine translation models.

“You can use the Translator service to detect the language of a given text.” → YesThis statement is also true. The Translator service includes automatic language detection, which determines the source language before translation. For instance, if a user submits text in an unknown language, the service can identify it automatically before performing translation.

“You can use the Translator service to transcribe audible speech into text.” → NoThis statement is false. Transcribing speech (audio) into text is a function of the Azure Speech service, specifically the Speech-to-Text API, not the Translator service.

Therefore, the Translator service is used for text translation and language detection, while speech transcription belongs to the Speech service.

According to the Microsoft Azure AI Fundamentals (AI-900) official study materials, object detection is a type of computer vision workload that not only identifies objects within an image but also determines their location by drawing bounding boxes around them. This functionality is clearly described in the Microsoft Learn module “Identify features of computer vision workloads.”

In this scenario, the AI system analyzes an image to find a vehicle and then returns a bounding box showing where that vehicle is located within the image frame. That ability — to detect, classify, and localize multiple objects — perfectly defines object detection.

Microsoft’s study content contrasts object detection with other computer vision workloads as follows:

Image classification: Determines what object or scene is present in an image as a whole but does not locate it (e.g., “this is a car”).

Object detection: Identifies what objects are present and where they are, usually returning coordinates for bounding boxes (e.g., “car detected at position X, Y”).

Optical Character Recognition (OCR): Extracts text content from images or scanned documents.

Facial detection: Specifically locates human faces within an image or video feed, often as part of face recognition systems.

In Azure, object detection capabilities are available through services such as Azure Computer Vision, Custom Vision, and Azure Cognitive Services for Vision, which can be trained to detect vehicles, products, or other objects in various image datasets.

Therefore, based on the AI-900 study guide and Microsoft Learn materials, the verified and correct answer is Object detection, as it accurately describes the process of returning a bounding box indicating an object’s position in an image.

According to Microsoft’s Responsible AI principles, one of the six core principles is fairness, which ensures that AI systems treat all individuals equitably and that their outcomes are not influenced by biases present in the training data or algorithms. The official Microsoft Learn module “Identify the guiding principles for responsible AI” clearly defines fairness as the requirement that AI systems should not amplify or perpetuate existing societal biases.

In this scenario, the statement emphasizes that AI systems should NOT reflect biases from the datasets used to train them, which directly aligns with the fairness principle. Bias in AI models can arise when the data used for training is unbalanced or not representative of the real-world population. For instance, if a facial recognition model is trained mostly on images of one demographic group, it may perform poorly on others—an example of unfair bias. Microsoft advocates building and testing AI systems with diverse, high-quality datasets to ensure fair performance across all groups.

The other principles listed—accountability, inclusiveness, and transparency—are also important but do not directly address bias mitigation:

Accountability ensures that people remain responsible for AI systems and their decisions.

Inclusiveness promotes accessibility and usability for all people, including those with disabilities.

Transparency focuses on explaining how AI systems make decisions.

However, Fairness explicitly deals with avoiding discrimination and bias in AI outcomes and training data.

Thus, in Microsoft’s Responsible AI framework, ensuring that systems do not reflect biases from datasets is part of the Fairness principle, which promotes equitable and unbiased treatment for all individuals in AI-driven decisions.

In the context of Microsoft Azure AI Fundamentals (AI-900) and general machine learning principles, regression refers to a type of supervised learning used to predict continuous numerical values based on historical data. The goal of regression is to model the relationship between input variables (features) and a continuous output variable (target).

In this scenario, the task is to predict how many vehicles will travel across a bridge on a given day. The number of vehicles is a numerical value that can vary continuously depending on factors such as time of day, weather, weekday/weekend, or traffic trends. Because the output is numeric and not categorical, this problem type clearly fits into regression analysis.

Microsoft’s official learning content for AI-900, under “Identify features of regression and classification machine learning models,” specifies that regression models are used to predict values such as sales forecasts, demand estimation, temperature prediction, or traffic volume—all of which share the same underlying objective: predicting a quantity.

To clarify other options:

Classification is used when predicting categories or discrete classes, such as determining whether an email is spam or not spam, or if an image contains a cat or a dog.

Clustering is an unsupervised learning technique used to group similar data points without predefined labels (for example, grouping customers by purchasing behavior).

Since predicting the number of vehicles results in a continuous numerical output, it aligns precisely with the regression workload type described in the Microsoft AI-900 study materials.

Statement

Yes / No

Providing an explanation of the outcome of a credit loan application is an example of the Microsoft transparency principle for responsible AI.

Yes

A triage bot that prioritizes insurance claims based on injuries is an example of the Microsoft reliability and safety principle for responsible AI.

Yes

An AI solution that is offered at different prices for different sales territories is an example of the Microsoft inclusiveness principle for responsible AI.

No

This question is based on the Responsible AI principles defined by Microsoft, a major topic in the AI-900: Microsoft Azure AI Fundamentals certification. The goal of Responsible AI is to ensure that artificial intelligence is developed and used ethically, safely, and transparently to benefit people and society. Microsoft’s framework defines six core principles: Fairness, Reliability and Safety, Privacy and Security, Inclusiveness, Transparency, and Accountability.

Transparency Principle – YesProviding an explanation for a loan application decision clearly reflects transparency. According to Microsoft’s Responsible AI guidelines, transparency involves ensuring that users and stakeholders understand how AI systems make decisions. When a financial AI model explains why a loan was approved or denied, it promotes user trust and confidence in automated decision-making. Transparency helps individuals understand influencing factors (like income or credit score), thereby fostering ethical AI deployment.

Reliability and Safety Principle – YesA triage bot that prioritizes insurance claims based on injury severity demonstrates reliability and safety. This principle ensures that AI systems consistently operate as intended, handle data accurately, and do not cause unintended harm. For a triage bot, safety means it must correctly interpret medical or claim information and consistently provide appropriate prioritization. Microsoft emphasizes that reliable AI systems must be tested rigorously, function correctly in various scenarios, and maintain user safety at all times.

Inclusiveness Principle – NoAn AI solution priced differently for various sales territories is unrelated to inclusiveness. Inclusiveness focuses on designing AI systems that are accessible and fair to all users, including those with disabilities or from different demographic backgrounds. Price variation across territories is a business strategy, not an ethical AI inclusion concern. Hence, this statement does not align with any Responsible AI principle.

According to the AI-900 official study guide and Microsoft Learn module “Describe features of conversational AI workloads on Azure”, conversational AI refers to artificial intelligence systems that interact with users through natural language via text or speech. These systems include chatbots, virtual assistants, and interactive voice response (IVR) systems that simulate human conversation.

B. Chatbot that provides users with the ability to find answers on a website by themselvesThis is a classic example of conversational AI. Chatbots use natural language understanding (LUIS) and Azure Bot Service to interpret user input, identify intent, and provide relevant responses automatically. They help users self-serve information without human support, such as retrieving account details or answering FAQs.

C. Telephone voice menus to reduce the load on human resourcesAutomated telephone systems or IVRs use conversational AI to interpret spoken commands and route calls intelligently. This is often implemented using Azure Cognitive Services Speech (for speech-to-text and text-to-speech) combined with Azure Bot Service for managing dialogue flow.

Generative AI focuses on creating new content rather than just analyzing existing data. As per Microsoft’s AI-900 curriculum and Azure OpenAI documentation, typical use cases include generating text, images, code, or other creative outputs based on input prompts.

A. Generating draft responses for customer service agents — ✅ Correct. GPT-based models can automatically generate draft replies to customer queries, enabling agents to refine responses and increase efficiency.

B. Creating original artwork from textual descriptions — ✅ Correct. DALL-E, available through Azure OpenAI, can produce unique images based on natural language prompts.

Options C and D are incorrect because they involve predictive or classification models, not generative ones:

C. Predicting sales revenue → Regression (machine learning).

D. Classifying email messages → Classification (machine learning).

Correct answers: A and B.

According to the Microsoft Azure AI Fundamentals (AI-900) official study guide and Microsoft Learn module “Identify features of classification machine learning”, classification is a type of supervised machine learning used when the goal is to predict a categorical outcome. That means the output variable represents discrete labels such as Yes/No, True/False, or Category A/B/C.

In this example, the model is predicting whether a person uses a bicycle (Yes or No) — a binary categorical outcome. The input (distance from home to work) is numeric, but the prediction is a class or category, which makes it a classification problem.

To compare:

A and D (predicting how many cups of coffee or race minutes) involve numeric predictions, which are regression tasks.

B (grouping images by similar colors) involves clustering, an unsupervised learning method used to find natural groupings in data.

Thus, the use case that fits classification is predicting whether someone uses a bicycle, since the answer is categorical.

The Transformer model architecture is a foundational deep learning model introduced in the 2017 research paper “Attention Is All You Need.” It serves as the core architecture for modern large language models such as GPT, BERT, and T5, all of which are used in Azure OpenAI Service.

“A transformer model architecture uses self-attention.” – YesThe self-attention mechanism is the defining feature of transformer models. It allows the model to evaluate the relationships between words (tokens) in a sequence and assign weights based on contextual relevance. This means that each word in an input sentence can "attend" to every other word, capturing dependencies regardless of their position in the text. This mechanism replaced older recurrent (RNN) and convolutional (CNN) architectures for sequence processing because it provides parallelization and better context understanding.

“A transformer model architecture includes an encoder block and a decoder block.” – YesThe original Transformer architecture includes both an encoder and a decoder. The encoder processes the input sequence into contextual representations, and the decoder generates the output sequence based on both the encoder’s output and previously generated tokens. Models like BERT use only the encoder stack, while GPT models use only the decoder stack, but the full Transformer design conceptually includes both.

“A transformer model architecture includes an encryption block or a decryption block.” – NoTransformers are not related to cryptography. They perform encoding and decoding of language data for representation learning—not encryption or decryption for data security. The terms “encoder” and “decoder” here refer to neural network components, not cryptographic processes.

To improve the quality and relevance of responses generated by a generative AI solution using GPT-3.5, the following three actions are emphasized in the Microsoft Learn: Azure OpenAI Service best practices and AI-900/AI-102 training materials:

(A) Add grounding data to prompts: Grounding ensures that the model’s output is based on factual, domain-specific information. By adding context or external knowledge sources, responses become more accurate and aligned with the organization’s data rather than relying on the model’s general training corpus.

(B) Provide additional examples to prompts: Also known as few-shot prompting, this method demonstrates desired response patterns by including examples in the prompt. This significantly improves output quality, consistency, and adherence to desired formats.

(E) Modify system messages: In Azure OpenAI chat completions, system messages define the model’s behavior, style, and tone. Adjusting system messages allows fine-tuning of the model’s response quality, ensuring it follows context or persona guidelines.

The remaining options are incorrect:

(C) Modify tokenization is a low-level text-processing technique not used to improve model response quality directly.

(D) Add training data to prompts is not possible at runtime since GPT-3.5 models are pre-trained; only prompt engineering can influence output behavior.

Therefore, based on Azure OpenAI and AI-900 guidance, the three best ways to enhance generative