AI-900 — Azure AI Fundamentals

Full question text, answer options, and explanations for this practice set — a spoiler-free alternative is the interactive quiz above for scored, shuffled practice.

1. 1. A retail company wants to build a system that automatically reads and interprets images of product labels to update inventory records. Which type of AI workload does this represent?

   • A. Natural language processing
   • B. Computer vision(correct)
   • C. Anomaly detection
   • D. Recommendation system

   Explanation: Interpreting images (product labels) to extract information is a computer vision workload. Natural language processing (A) handles text and speech. Anomaly detection (C) identifies unusual patterns in time-series or tabular data. Recommendation systems (D) suggest items based on user behaviour — none of these involve image interpretation.

2. 2. An AI system used in loan approvals has been found to approve loans at a lower rate for applicants from certain postal codes, even when their financial profiles are similar to approved applicants. Which Responsible AI principle is most directly at risk?

   • A. Reliability and safety
   • B. Fairness(correct)
   • C. Privacy and security
   • D. Transparency

   Explanation: Fairness in AI requires that AI systems treat all people equitably and do not produce systematically different outcomes for groups based on characteristics such as geography, which may act as a proxy for race or socioeconomic status. Reliability/safety (A) concerns consistent and predictable behaviour. Privacy/security (C) concerns data protection. Transparency (D) concerns explainability and disclosure.

3. 3. A hospital AI system occasionally provides incorrect drug dosage recommendations due to rare edge cases in training data. A pharmacist must review every AI recommendation before it is acted on. Which Responsible AI principle does this human-review step primarily support?

   • A. Inclusiveness
   • B. Accountability(correct)
   • C. Reliability and safety
   • D. Transparency

   Explanation: Accountability means that humans remain responsible for AI-assisted decisions, especially in high-stakes domains. Requiring a pharmacist to review each recommendation establishes a human accountability layer. Inclusiveness (A) concerns designing AI for all users including those with disabilities. Reliability/safety (C) is about consistent, safe performance — the mandatory review is a governance control. Transparency (D) is about explaining how the AI works, not who is responsible.

4. 4. Which of the following is a characteristic of a document processing AI workload?

   • A. Generating new images from text descriptions
   • B. Extracting structured data such as invoice totals and dates from scanned forms(correct)
   • C. Recommending movies based on a user's viewing history
   • D. Detecting network intrusion patterns in firewall logs

   Explanation: Document processing workloads extract structured information (fields, tables, key-value pairs) from unstructured documents such as invoices, receipts, and forms. Generating images (A) is a generative AI workload. Movie recommendations (C) are a personalisation/recommendation workload. Network intrusion detection (D) is an anomaly detection or security workload.

5. 5. A data scientist wants to predict the sale price of a house based on features such as square footage, number of bedrooms, and neighbourhood. Which machine learning technique is most appropriate?

   • A. Classification
   • B. Clustering
   • C. Regression(correct)
   • D. Reinforcement learning

   Explanation: Regression predicts a continuous numeric output (sale price). Classification (A) predicts a discrete category label. Clustering (B) groups unlabelled data points by similarity with no target variable. Reinforcement learning (D) trains an agent through reward and penalty signals, not on labelled datasets.

6. 6. In a supervised machine learning workflow, a data scientist splits a labelled dataset into a training set and a validation set. What is the primary purpose of the validation set?

   • A. To provide additional training examples when the training set is too small
   • B. To evaluate the model's performance on data it has not seen during training and guide hyperparameter tuning(correct)
   • C. To store the final trained model artefact for deployment
   • D. To generate synthetic training data using data augmentation

   Explanation: The validation set is held back during training and used to measure generalisation performance after each epoch or tuning iteration, enabling the data scientist to detect overfitting and select optimal hyperparameters. It is not used for training (A), model storage (C), or data augmentation (D).

7. 7. A business analyst with no machine learning coding experience needs to build a customer churn prediction model using historical CRM data. The analyst wants to automatically explore many algorithm types and select the best-performing model with minimal manual effort. Which Azure Machine Learning capability should they use?

   • A. Manual hyperparameter tuning scripts written in Python
   • B. Automated machine learning (AutoML) to automatically try multiple algorithms and return the best model(correct)
   • C. Azure AI Vision to classify customer images
   • D. Azure AI Language to run sentiment analysis on CRM notes

   Explanation: Automated machine learning (AutoML) in Azure Machine Learning automatically iterates through algorithm and feature engineering combinations, evaluating each run and returning the best-performing model — ideal for analysts without deep ML coding expertise. Manual Python tuning (A) requires coding expertise. Azure AI Vision (C) processes images, not tabular CRM data. Azure AI Language (D) analyses text sentiment, not churn prediction from structured data.

8. 8. Which architecture forms the foundation for modern large language models (LLMs) such as GPT-4?

   • A. Convolutional Neural Network (CNN)
   • B. Recurrent Neural Network (RNN) with LSTM cells
   • C. Transformer architecture with self-attention mechanisms(correct)
   • D. Support Vector Machine (SVM) with radial basis function kernel

   Explanation: The Transformer architecture, introduced in the 2017 'Attention Is All You Need' paper, uses self-attention to process all tokens in parallel and model long-range dependencies — it is the foundation for GPT-4 and other LLMs. CNNs (A) are designed for spatial feature extraction in images. RNNs with LSTM (B) process sequences but are limited by sequential computation and vanishing gradients at long range. SVMs (D) are shallow, non-neural classifiers.

9. 9. A developer uses Azure AI Vision to scan images of passports and extract the name, date of birth, and document number from the printed text. Which computer vision capability is being used?

   • A. Image classification
   • B. Object detection
   • C. Optical character recognition (OCR)(correct)
   • D. Facial analysis

   Explanation: OCR reads and extracts printed or handwritten text from images. Image classification (A) assigns a label to the whole image. Object detection (B) finds and locates objects with bounding boxes. Facial analysis (D) identifies and analyses human faces, not printed document text.

10. 10. An online retail company wants to identify whether product images uploaded by sellers belong to the correct category (e.g., 'Electronics', 'Clothing', 'Furniture'). Each image should receive exactly one category label. Which computer vision solution type is most appropriate?

    • A. Object detection
    • B. Image classification(correct)
    • C. OCR
    • D. Facial detection

    Explanation: Image classification assigns a single label to an entire image from a set of predefined classes — exactly the requirement here. Object detection (A) locates and labels multiple objects within one image. OCR (C) extracts text characters. Facial detection (D) is specific to human faces.

11. 11. A smart building company installs cameras to monitor whether employees are wearing hard hats in a construction zone. The system must detect the presence of hard hats and their locations in real time. Which Azure AI service and capability should they use?

    • A. Azure AI Language with entity recognition
    • B. Azure AI Vision object detection to locate hard hats in each video frame(correct)
    • C. Azure AI Face detection to identify individual employees
    • D. Azure AI Vision image classification to label the entire scene as 'safe' or 'unsafe'

    Explanation: Object detection via Azure AI Vision identifies and localises specific objects (hard hats) with bounding boxes in each frame, enabling accurate counting and position tracking. Azure AI Language (A) processes text, not images. Face detection (C) identifies people, not equipment on their heads. Image classification (D) would label the whole scene but cannot pinpoint the location of individual hard hats.

12. 12. Which Azure AI service provides face detection capabilities that can identify facial attributes such as estimated age, emotion, and whether a person is wearing glasses?

    • A. Azure AI Vision (Image Analysis)
    • B. Azure AI Face detection service(correct)
    • C. Azure AI Document Intelligence
    • D. Azure AI Video Indexer

    Explanation: Azure AI Face is the dedicated service for detecting human faces and analysing facial attributes (approximate age, emotion, accessories like glasses) and, with appropriate access policies, face verification and identification. Azure AI Vision's image analysis (A) can detect faces at a basic level but the dedicated Face service provides richer attribute analysis. Document Intelligence (C) processes documents. Video Indexer (D) analyses video content for broader insights, using the Face service internally but is not itself the face attribute service.

13. 13. A news aggregation service automatically tags articles by topics such as 'Finance', 'Technology', and 'Sports'. Users can also search articles in any language, and the system must identify which language each article is written in. Which Azure AI Language feature should be used for language identification?

    • A. Sentiment analysis
    • B. Language detection(correct)
    • C. Key phrase extraction
    • D. Entity recognition

    Explanation: Language detection in Azure AI Language identifies the language of a piece of text and returns a confidence score. Sentiment analysis (A) scores emotional tone. Key phrase extraction (C) returns important terms. Entity recognition (D) identifies named entities such as people and places — none of these determine the document's language.

14. 14. A customer service team wants to automatically score thousands of product reviews each day to understand whether customers are satisfied. They need a numerical score from positive to negative without reading each review individually. Which Azure AI Language feature should they use?

    • A. Key phrase extraction to list the most important words in each review
    • B. Sentiment analysis to return a sentiment score and label (positive, neutral, negative) for each review(correct)
    • C. Translation to convert all reviews into English before analysis
    • D. Named entity recognition to identify product names mentioned in reviews

    Explanation: Sentiment analysis in Azure AI Language scores text on a positive-to-negative scale and assigns a label, allowing teams to track customer satisfaction at scale. Key phrase extraction (A) returns important terms but not emotional tone. Translation (C) changes the language but does not assess sentiment. NER (D) identifies entities but not the emotional valence of the review.

15. 15. A multinational company wants to build a voice assistant that listens to spoken commands and responds in spoken language. Which TWO Azure AI Speech capabilities must be combined for this use case?

    • A. Speech-to-text to convert the user's spoken input into text, and text-to-speech to convert the assistant's response back to audio(correct)
    • B. Text-to-speech to convert the user's spoken commands into text
    • C. Optical character recognition to read printed commands
    • D. Named entity recognition to classify the language of the spoken input

    Explanation: A voice assistant requires speech-to-text (STT) to convert incoming audio to text for processing, and text-to-speech (TTS) to convert the generated text response back to audio. Text-to-speech alone (B) goes in the wrong direction for input. OCR (C) reads printed text, not audio. NER (D) extracts entities from text and does not identify language or process audio.

16. 16. An enterprise application needs to extract the most important concepts from long meeting transcripts and identify company names and project codes mentioned. Which TWO Azure AI Language features should be used together? Choose 2.

    • A. Key phrase extraction to identify the most important topics and terms(correct)
    • B. Named entity recognition (NER) to identify company names and project codes as entities(correct)
    • C. Language detection to determine which language the meeting was held in
    • D. Sentiment analysis to score the meeting as positive or negative
    • E. Translation to convert the transcripts to a different language

    Explanation: Key phrase extraction (A) surfaces the most salient concepts and topics from long text. NER (B) identifies and classifies specific entities like company names and project codes — together they address both requirements. Language detection (C) identifies the language but is not needed if the transcript language is already known. Sentiment analysis (D) scores tone, not content concepts. Translation (E) is a different task entirely.

17. 17. Which of the following best describes a generative AI model?

    • A. A model that classifies input data into predefined categories
    • B. A model that generates new content such as text, images, or code based on a prompt(correct)
    • C. A model that clusters unlabelled data into groups based on similarity
    • D. A model that predicts a continuous numeric value from input features

    Explanation: Generative AI models learn the distribution of training data and generate new, original content — text, images, audio, or code — in response to prompts. Classification models (A) assign labels. Clustering models (C) group data without labels. Regression models (D) predict numeric values. None of these create new content.

18. 18. A software development team wants to use a generative AI service on Azure to automatically generate unit test code for functions they write. Which Azure service is most appropriate?

    • A. Azure AI Vision
    • B. Azure OpenAI Service(correct)
    • C. Azure AI Document Intelligence
    • D. Azure AI Speech

    Explanation: Azure OpenAI Service provides access to large language models including code-generation models such as GPT-4 that can generate unit tests from function signatures and descriptions. Azure AI Vision (A) processes images. Document Intelligence (C) extracts data from documents. Azure AI Speech (D) handles voice and audio.

19. 19. What is the primary function of Azure AI Foundry in the context of building generative AI solutions?

    • A. It is a database service for storing AI training datasets
    • B. It is a unified platform for building, evaluating, and deploying AI models and applications, including generative AI, with collaboration and governance features(correct)
    • C. It is a code editor extension for writing Python machine learning scripts
    • D. It is a service for creating and training custom speech recognition models only

    Explanation: Azure AI Foundry is Microsoft's unified AI development platform that provides a hub and project structure for collaborative AI development, model catalog access, prompt flow, evaluation tools, and deployment — spanning generative AI and beyond. It is not a database (A), a code editor (C), or a speech-only service (D).

20. 20. A startup is building a chatbot that must answer questions about its proprietary product documentation. The documentation changes frequently. The team wants the model's answers to always reflect the latest documents without retraining the model. Which generative AI pattern is most suitable?

    • A. Fine-tuning the language model on the documentation each time it is updated
    • B. Retrieval-Augmented Generation (RAG), where relevant documentation chunks are retrieved at query time and included in the prompt(correct)
    • C. Increasing the model's maximum token limit to fit all documentation into every prompt
    • D. Using Azure AI Vision to convert documentation pages into images for the model to read

    Explanation: RAG pairs a retrieval system (e.g., Azure AI Search) with a generative model. When a user asks a question, relevant document chunks are fetched and inserted into the prompt, so the model's answer reflects current documentation without retraining. Fine-tuning (A) is expensive and cannot keep pace with frequent documentation changes. Fitting all documentation into the prompt (C) is infeasible for large document sets. Converting to images (D) is not how language models consume documents.

21. 21. Which Azure AI Foundry feature allows developers to discover, compare, and deploy a wide variety of foundation models from Microsoft and third-party providers (such as Meta and Mistral AI) in one place?

    • A. Azure AI Foundry prompt flow
    • B. Azure AI Foundry model catalog(correct)
    • C. Azure Machine Learning automated ML
    • D. Azure AI Search semantic ranking

    Explanation: The Azure AI Foundry model catalog is a central repository of foundation models from Microsoft (e.g., Phi), OpenAI, Meta (Llama), Mistral, and others, allowing users to browse benchmarks, compare models, and deploy them with a few clicks. Prompt flow (A) is for orchestrating AI workflows. AutoML (C) trains custom models on your data. Semantic ranking (D) improves search relevance.

22. 22. A company is evaluating the use of generative AI to produce marketing copy. Leadership asks what Responsible AI considerations are specific to generative AI that differ from traditional classification models. Which answer best captures a generative-AI-specific concern?

    • A. Generative AI models can produce factually incorrect or fabricated information (hallucinations) that appears plausible, which can mislead users(correct)
    • B. Generative AI models require labelled training datasets, making data collection the primary concern
    • C. Generative AI models cannot process natural language, so they require structured input
    • D. Generative AI models are always deterministic and produce the same output for the same input

    Explanation: Hallucination — generating plausible-sounding but factually wrong content — is a Responsible AI consideration unique to generative models, requiring mitigation strategies such as grounding, fact-checking, and human review. Labelled datasets (B) are a concern for supervised learning, not specifically generative AI. Generative AI does process natural language (C) — that statement is false. Generative models are stochastic, not deterministic (D).

23. 23. Which TWO statements correctly describe the Azure OpenAI Service? Choose 2.

    • A. It provides access to OpenAI models (such as GPT-4 and DALL-E) hosted on Microsoft Azure infrastructure with enterprise security and compliance(correct)
    • B. It allows developers to call models via REST API or SDKs and integrate them into custom applications(correct)
    • C. It is exclusively a data storage service for training datasets used by OpenAI models
    • D. It can only be used by US-based customers due to data residency restrictions
    • E. It requires users to train a model from scratch before any text generation is possible

    Explanation: Azure OpenAI Service hosts OpenAI's models (GPT-4, DALL-E, Whisper, etc.) on Microsoft's Azure infrastructure under Microsoft's security, compliance, and data privacy controls (A). Developers integrate these models via REST API, Python SDK, or .NET SDK into their own applications (B). It is not a data storage service (C). It is available in multiple Azure regions globally, not only to US customers (D). Developers use pre-trained foundation models — no from-scratch training is required (E).

24. 24. A content creation agency is using an Azure OpenAI model to draft social media posts. A reviewer notices the model occasionally generates posts that contain subtle demographic stereotypes. According to Microsoft's Responsible AI guidelines, what is the most appropriate first step to mitigate this?

    • A. Accept the outputs as-is because AI models cannot be changed after deployment
    • B. Enable Azure AI Content Safety filters and configure system prompts with explicit instructions to avoid stereotypes, then evaluate outputs regularly(correct)
    • C. Delete the Azure OpenAI resource and switch to a manual content workflow
    • D. Increase the model's max_tokens parameter to give it more space to self-correct

    Explanation: Microsoft's Responsible AI approach for generative AI includes using content safety filters to detect harmful content and engineering the system prompt to specify behavioural constraints (such as avoiding stereotypes), combined with ongoing human evaluation of outputs. Models can be constrained after deployment through these controls (refuting A). Deleting the resource (C) abandons the AI initiative unnecessarily. Increasing max_tokens (D) affects response length, not bias in content.

25. 25. A developer wants to create an application where users can describe an image they want in plain English and the application generates that image automatically. Which Azure OpenAI model capability should they use?

    • A. GPT-4 text completion to write a description of the desired image
    • B. DALL-E image generation to create an image from a text prompt(correct)
    • C. Whisper speech-to-text to convert verbal descriptions to text
    • D. Azure AI Vision image classification to label the image category

    Explanation: DALL-E is the Azure OpenAI model that generates novel images from natural-language text prompts — exactly the described use case. GPT-4 text completion (A) generates text, not images. Whisper (C) transcribes audio to text, which could be a useful upstream step but is not the image-generation component. Azure AI Vision image classification (D) labels an existing image; it does not create one.