Free AI-900 Practice Questions with Detailed Explanations

Test your Azure AI Fundamentals readiness with 25 free practice questions covering machine learning, computer vision, NLP, and generative AI.

15 min read
Updated April 2026
AI-900 Fundamentals

The AI-900 (Microsoft Azure AI Fundamentals) exam validates your understanding of AI and machine learning concepts and how they are implemented using Azure AI services. It's ideal for business and technical roles who work with or want to understand AI solutions — no programming experience required.

These 25 questions cover all five AI-900 domains, emphasizing the latest exam content including generative AI and Azure OpenAI Service. Compare with our AZ-900 vs AI-900 comparison if you're deciding between the two. Full details on our AI-900 certification page.

What You'll Get:

  • 25 scenario-based questions across all 5 exam domains
  • Generative AI questions including Azure OpenAI Service
  • Detailed explanations for every answer option
  • Scoring guide to assess your readiness

What These Questions Cover

5
AI Workloads
Concepts & responsible AI
5
Machine Learning
ML types & Azure ML
5
Computer Vision
Azure AI Vision
5
NLP
Language & speech services
5
Generative AI
Azure OpenAI, Copilot

📝 Practice Test Instructions

  • • Each question has ONE correct answer
  • • Questions are service-identification focused — know which Azure AI service does what
  • • Note your answers before scrolling to the answer key
  • • Aim to complete all 25 questions in 25 minutes
🤖

AI Workloads & Machine Learning

Questions 1–10

1

AI Workload Types

A hospital wants to build a system that reads chest X-ray images and automatically flags potential abnormalities for radiologist review. The system should learn from thousands of labeled historical X-rays.

Which type of AI workload best describes this system?

A)Computer vision
B)Natural language processing
C)Knowledge mining
D)Conversational AI
2

Responsible AI — Fairness

A company trains a loan approval AI model using 10 years of historical approval data. After deployment, they discover the model approves loans at significantly lower rates for certain demographic groups, even when financial profiles are identical.

Which responsible AI principle has been violated?

A)Reliability and safety
B)Fairness
C)Transparency
D)Privacy and security
3

Responsible AI — Transparency

An AI system makes decisions about employee performance reviews, but employees cannot understand why they received a specific rating or how the AI arrived at its conclusion.

Which responsible AI principle is most relevant to address this concern?

A)Fairness
B)Inclusiveness
C)Explainability (Transparency)
D)Reliability and safety
4

Machine Learning Types

You want to build an ML model that predicts a customer's expected lifetime value (a numeric dollar amount) based on their purchase history, demographics, and engagement data. You have thousands of historical records with known lifetime values.

Which type of machine learning task is this?

A)Classification
B)Reinforcement learning
C)Clustering
D)Regression
5

Unsupervised Learning

A retailer has millions of customer transactions but no predefined categories. They want the ML system to automatically identify natural groupings of customers with similar purchase behaviors so marketing can target each group differently.

Which ML technique should be used?

A)Clustering (unsupervised learning)
B)Supervised classification
C)Regression analysis
D)Anomaly detection
6

Azure Machine Learning

A data science team needs a platform to build, train, deploy, and monitor machine learning models at scale. They need experiment tracking, a model registry, automated ML capabilities, and a managed compute environment.

Which Azure service provides this end-to-end ML lifecycle platform?

A)Azure Databricks
B)Azure Machine Learning
C)Azure Synapse Analytics
D)Azure Cognitive Search
7

Automated ML

A business analyst without coding experience wants to build a classification model to predict customer churn. They have a labeled CSV dataset. They need the system to automatically try multiple algorithms and select the best one.

Which Azure Machine Learning feature should they use?

A)Azure ML Designer
B)Azure ML Notebooks
C)Automated ML (AutoML)
D)Azure ML Endpoints
8

Azure ML Designer

A data scientist wants to build a machine learning pipeline using a visual drag-and-drop interface without writing code. They want to connect data preparation, feature engineering, model training, and evaluation steps graphically.

Which Azure ML feature provides this visual pipeline builder?

A)Automated ML
B)Azure ML CLI
C)Azure ML SDK
D)Azure ML Designer
9

Model Evaluation

You trained a binary classification model to detect fraudulent transactions. The dataset has 99% legitimate and 1% fraudulent transactions. The model predicts "legitimate" for every transaction and achieves 99% accuracy.

Why is accuracy alone misleading for this model, and which metric better measures its effectiveness?

A)Accuracy is misleading because the dataset is imbalanced. Recall (sensitivity) is better — it measures what percentage of actual fraudulent transactions the model correctly identifies
B)Accuracy is fine — 99% is excellent performance for any model
C)The model needs more training data to improve accuracy
D)Precision is the only valid metric for classification models
10

Feature Engineering

You are preparing data for a regression model that predicts house prices. Your dataset has age of house (in years), number of rooms, and zip code (categorical text). Before training, you need to convert zip codes into a format usable by the model.

What is the recommended technique for encoding categorical variables like zip codes?

A)Normalization
B)One-hot encoding
C)Imputation
D)Feature scaling
👁️

Computer Vision & NLP

Questions 11–20

11

Azure AI Vision — OCR

A logistics company needs to automatically extract text from scanned shipping documents, invoices, and forms — including handwritten notes in some cases — and store the extracted data in a database.

Which Azure AI service capability should they use?

A)Azure AI Vision — Image Analysis
B)Azure AI Document Intelligence (Form Recognizer)
C)Azure AI Vision — Optical Character Recognition (OCR)
D)Azure AI Language — Key Phrase Extraction
12

Azure AI Vision — Image Analysis

An e-commerce platform wants to automatically generate descriptive captions and identify objects in product images uploaded by sellers. The goal is to improve search indexing and accessibility.

Which Azure AI Vision feature handles image captioning and object detection?

A)Azure AI Vision — OCR
B)Azure AI Face service
C)Azure AI Custom Vision — Classification
D)Azure AI Vision — Image Analysis (caption and detect objects)
13

Azure AI Custom Vision

A manufacturing company wants to detect defects in products on an assembly line using images. The defects are unique to their specific product and cannot be identified by general-purpose image classification. They have 500 labeled images of defective and non-defective products.

Which Azure AI service is best suited for training a custom product defect detector?

A)Azure AI Custom Vision
B)Azure AI Vision — Image Analysis
C)Azure Machine Learning Designer
D)Azure AI Document Intelligence
14

Azure AI Face Service

A security system needs to verify that the person at a secure entry point matches a stored employee photo. The system should compare facial features between a live camera feed and an employee photo database.

Which Azure AI capability handles face verification?

A)Azure AI Vision — Image Analysis
B)Azure AI Face — Verify (face comparison)
C)Azure AI Custom Vision — Classification
D)Azure AI Vision — Spatial Analysis
15

Azure AI Language — Sentiment Analysis

A brand monitoring tool needs to analyze thousands of customer reviews and social media comments to determine whether each post expresses a positive, negative, or neutral sentiment about a product.

Which Azure AI Language feature addresses this need?

A)Named Entity Recognition
B)Key Phrase Extraction
C)Sentiment Analysis and Opinion Mining
D)Language Detection
16

Azure AI Language — Entity Recognition

A legal document processing system needs to automatically identify and extract mentions of people, organizations, locations, dates, and legal case numbers from unstructured text.

Which Azure AI Language feature is best for this task?

A)Sentiment Analysis
B)Translation
C)Text Summarization
D)Named Entity Recognition (NER)
17

Azure AI Translator

A global customer support platform needs to translate customer messages from over 100 languages into English in real time so agents can respond, then translate the agent's English response back to the customer's language.

Which Azure AI service provides real-time text translation across 100+ languages?

A)Azure AI Translator
B)Azure AI Language — Language Detection
C)Azure AI Language — Custom Text Classification
D)Azure OpenAI Service
18

Azure AI Speech

A call center wants to automatically transcribe customer phone calls into text in real time for quality monitoring, compliance logging, and sentiment analysis.

Which Azure AI Speech capability should they use?

A)Text to Speech
B)Speech to Text (Real-time transcription)
C)Speech Translation
D)Speaker Recognition
19

Conversational AI — Azure Bot Service

A bank wants to build a customer service chatbot that can handle FAQs, check account balances via API, and escalate to a human agent when needed. The chatbot should work on both their website and Microsoft Teams.

Which Azure service orchestrates the bot's conversation logic and connects to multiple channels?

A)Azure AI Language — CLU (Conversational Language Understanding)
B)Azure AI Language — Question Answering
C)Azure Bot Service
D)Azure OpenAI Service
20

Azure AI Language — Question Answering

A company has a large FAQ document with hundreds of questions and answers. They want to create a bot that automatically answers user questions by finding the best matching answer from the FAQ document.

Which Azure AI Language feature is purpose-built for this use case?

A)Conversational Language Understanding (CLU)
B)Custom Text Classification
C)Azure OpenAI Service with RAG
D)Question Answering (formerly QnA Maker)

Generative AI

Questions 21–25

21

Large Language Models

A developer is building an application that can generate human-quality text, answer questions in natural language, summarize documents, and write code. They want to use a pre-trained model rather than building from scratch.

Which Azure service provides access to large language models like GPT-4 for these generative tasks?

A)Azure OpenAI Service
B)Azure Machine Learning Automated ML
C)Azure AI Language — Text Summarization
D)Azure AI Search with semantic ranking
22

Prompt Engineering

A developer is using Azure OpenAI Service to build a customer service assistant. They need the model to always respond formally, never discuss competitor products, and format all answers as bulleted lists. These instructions should apply to every conversation.

Where should these persistent behavioral instructions be placed?

A)In each user message before the question
B)In the system message (system prompt)
C)As fine-tuning data for the model
D)In the Azure OpenAI deployment configuration
23

Retrieval Augmented Generation

A law firm wants to use Azure OpenAI to answer questions about their specific internal legal documents, case files, and policies. The model should only answer from these documents, not from its general training knowledge.

Which approach enables the model to answer from the company's own documents?

A)Fine-tune the GPT model on the company's documents
B)Use a larger GPT model version with more parameters
C)Retrieval Augmented Generation (RAG) using Azure AI Search + Azure OpenAI
D)Increase the model's temperature setting
24

Responsible Generative AI

An organization is deploying an Azure OpenAI-based chatbot for public use. They need to ensure the model does not generate hate speech, sexually explicit content, or instructions for harmful activities.

Which Azure OpenAI feature automatically filters and blocks harmful content?

A)System message instructions
B)Prompt engineering best practices
C)Model temperature settings
D)Azure AI Content Safety / Content Filters in Azure OpenAI
25

Copilot and AI Assistants

A Microsoft 365 user wants an AI assistant that can summarize long email threads, draft replies, create PowerPoint presentations from a prompt, and generate Excel formulas — all integrated within their existing Microsoft 365 apps.

Which Microsoft product provides this AI assistant capability embedded in Microsoft 365 apps?

A)Microsoft 365 Copilot
B)Azure OpenAI Service API
C)Azure AI Language
D)Microsoft Bing Chat Enterprise

✋ Stop Here Before Scrolling!

Have you answered all 25 questions? Complete the test before checking the answers below.

Pro tip: The real AI-900 exam heavily tests which Azure service matches which AI capability

📝 Answer Key with Detailed Explanations

Review each explanation carefully, even for questions you answered correctly

Quick Answer Reference

Q1
A
Q2
B
Q3
C
Q4
D
Q5
A
Q6
B
Q7
C
Q8
D
Q9
A
Q10
B
Q11
C
Q12
D
Q13
A
Q14
B
Q15
C
Q16
D
Q17
A
Q18
B
Q19
C
Q20
D
Q21
A
Q22
B
Q23
C
Q24
D
Q25
A
1

Question 1: AI Workload Types

✓ Correct Answer: A) Computer vision

Why this is correct:

Computer vision is the AI workload that enables machines to interpret and understand visual information from images and videos. Analyzing X-ray images to detect abnormalities is a classic computer vision (specifically image classification/object detection) use case. The model learns from labeled images.

Why other answers are incorrect:

B: NLP processes text and speech — not images.
C: Knowledge mining extracts insights from large volumes of unstructured text documents — not medical images.
D: Conversational AI handles dialog and question-answering interactions — not image analysis.

💡 Key Concept:

AI workload types: Computer vision (images/video), NLP (text/speech), Knowledge mining (extract from unstructured data), Generative AI (create content), Conversational AI (dialog systems), Anomaly detection, Prediction/ML.

2

Question 2: Responsible AI — Fairness

✓ Correct Answer: B) Fairness

Why this is correct:

Fairness in AI means systems should treat all groups of people equitably. When a model discriminates based on demographic characteristics (race, gender, age) even with identical financial profiles, it violates fairness. This is often caused by historical bias in training data.

Why other answers are incorrect:

A: Reliability and safety refers to the system performing as expected and not causing harm — not demographic discrimination.
C: Transparency/explainability refers to understanding how the AI makes decisions — the violation here is discriminatory outcomes, not opacity.
D: Privacy and security refers to protecting personal data — not equal treatment across groups.

💡 Key Concept:

Microsoft's 6 responsible AI principles: Fairness, Reliability & Safety, Privacy & Security, Inclusiveness, Transparency, Accountability. Know which scenario maps to which principle.

3

Question 3: Responsible AI — Transparency

✓ Correct Answer: C) Explainability (Transparency)

Why this is correct:

Transparency (which includes explainability) means AI systems should be understandable — users should know how decisions are made and be able to challenge those decisions. When employees can't understand why they received a specific AI-generated rating, transparency is lacking.

Why other answers are incorrect:

A: Fairness addresses equal treatment across groups — not the ability to understand decision reasoning.
B: Inclusiveness means AI should benefit everyone and not exclude people — not about explaining decisions.
D: Reliability means consistent, trustworthy performance — not decision explainability.

💡 Key Concept:

Transparency includes: model explainability (why did it decide X?), documentation of model capabilities and limitations, disclosure that AI is being used. Azure AI provides Responsible AI tools in Azure ML for model explanation.

4

Question 4: Machine Learning Types

✓ Correct Answer: D) Regression

Why this is correct:

Regression is a supervised learning task that predicts a continuous numeric output value. Predicting customer lifetime value (a dollar amount) is regression because the output is a number on a continuous scale — not a category. You have labeled training data (historical records with known values) — making it supervised learning.

Why other answers are incorrect:

A: Classification predicts discrete categories (e.g., "high value," "medium value," "low value") — not exact numeric values.
B: Reinforcement learning learns through trial, reward, and penalty — not from labeled historical datasets.
C: Clustering groups similar items together without predefined labels — no labeled training data. The question has labeled data.

💡 Key Concept:

Supervised ML task types: Regression = predict a number (price, value, temperature). Classification = predict a category (yes/no, A/B/C). Clustering = find natural groups in unlabeled data (unsupervised).

5

Question 5: Unsupervised Learning

✓ Correct Answer: A) Clustering (unsupervised learning)

Why this is correct:

Clustering is an unsupervised ML technique that discovers natural groupings in data without predefined labels. Since the retailer has no predefined customer categories and wants the algorithm to find natural groupings, clustering is correct. K-means is a common clustering algorithm.

Why other answers are incorrect:

B: Supervised classification requires labeled training data — customers already categorized into known segments. The retailer has no predefined segments.
C: Regression predicts a numeric value — not groupings.
D: Anomaly detection identifies outliers or unusual patterns — not customer segment groupings.

💡 Key Concept:

Unsupervised vs Supervised: Supervised = labeled training data, predict known output types. Unsupervised = no labels, discover hidden patterns. Clustering, dimensionality reduction, and generative models are unsupervised.

6

Question 6: Azure Machine Learning

✓ Correct Answer: B) Azure Machine Learning

Why this is correct:

Azure Machine Learning is Microsoft's comprehensive ML platform providing: experiment tracking, model registry, automated ML, ML pipelines, managed compute (CPU/GPU clusters), MLOps, model deployment (endpoints), and monitoring. It supports the full ML lifecycle from data prep to production deployment.

Why other answers are incorrect:

A: Azure Databricks is an Apache Spark-based analytics platform — excellent for big data processing but not the primary ML lifecycle management platform.
C: Azure Synapse Analytics is a unified analytics platform for data warehousing and big data — not specifically for ML model training and deployment pipelines.
D: Azure AI Search (Cognitive Search) is a search service that can incorporate ML capabilities — it's not an ML training platform.

💡 Key Concept:

Azure ML key features: Automated ML, Designer (drag-and-drop), Notebooks (Jupyter), Managed Compute, Model Registry, Online Endpoints (real-time), Batch Endpoints (bulk inference), Responsible AI dashboard.

7

Question 7: Automated ML

✓ Correct Answer: C) Automated ML (AutoML)

Why this is correct:

AutoML automatically tries multiple algorithms (logistic regression, random forest, gradient boosting, etc.), performs hyperparameter tuning, and selects the best-performing model based on your chosen metric — all without coding. Perfect for business analysts with no ML coding experience.

Why other answers are incorrect:

A: Azure ML Designer is visual and low-code but still requires the user to manually select and configure ML components — more knowledge needed than AutoML.
B: Azure ML Notebooks require Python or R coding — not suitable for no-code scenarios.
D: Azure ML Endpoints are for deploying already-trained models — not for training from scratch.

💡 Key Concept:

Azure ML no-code path: AutoML (automated algorithm selection) → Designer (visual pipeline building) → Notebooks (full code control). AutoML is the least code required.

8

Question 8: Azure ML Designer

✓ Correct Answer: D) Azure ML Designer

Why this is correct:

Azure ML Designer provides a drag-and-drop visual interface where you can build ML pipelines by connecting pre-built components (data import, data transformation, training algorithms, evaluation modules). You can see the data flow graphically and run the pipeline without writing code.

Why other answers are incorrect:

A: AutoML picks the best algorithm automatically but doesn't provide a visual pipeline builder — you configure it through a wizard, not a graphical canvas.
B: Azure ML CLI is command-line interface — requires coding.
C: Azure ML SDK is the Python/R API for programmatic ML — requires coding.

💡 Key Concept:

Designer vs AutoML: Designer = you choose and connect components visually, full control over pipeline. AutoML = you define the task and target column, system tries everything. Both are low/no-code.

9

Question 9: Model Evaluation

✓ Correct Answer: A) Accuracy is misleading because the dataset is imbalanced. Recall (sensitivity) is better — it measures what percentage of actual fraudulent transactions the model correctly identifies

Why this is correct:

When class distribution is highly imbalanced (99% vs 1%), accuracy is misleading — a model that always predicts "legitimate" gets 99% accuracy without detecting any fraud. Recall (True Positives / (True Positives + False Negatives)) measures how well the model catches actual positives — critical for fraud where missing a fraud is costly.

Why other answers are incorrect:

B: 99% accuracy on an imbalanced dataset means the model detected 0 fraudulent transactions — this is terrible performance despite high accuracy.
C: More data won't help if the evaluation metric is wrong.
D: Precision measures accuracy of positive predictions (when it says fraud, how often is it right?) — but recall (catching all actual fraud) is more important for fraud detection.

💡 Key Concept:

Precision vs Recall: Precision = when you say positive, how often right? Recall = of all actual positives, how many did you catch? For fraud/disease detection, high Recall is critical (don't miss cases). For spam filtering, high Precision is critical (don't block legit emails).

10

Question 10: Feature Engineering

✓ Correct Answer: B) One-hot encoding

Why this is correct:

One-hot encoding converts categorical variables (like zip code, which has no numeric ordering relationship) into binary columns — one column per category value with 1 or 0. This allows ML algorithms that require numeric input to process categorical data without implying a false numeric ordering.

Why other answers are incorrect:

A: Normalization scales numeric values to a standard range (e.g., 0-1) — it's for continuous numeric features, not categorical ones.
C: Imputation fills missing values — the question is about encoding existing values.
D: Feature scaling (standardization/normalization) adjusts numeric ranges — not applicable to categorical data encoding.

💡 Key Concept:

Categorical encoding: One-hot encoding = each category becomes a binary column (no ordering). Label encoding = assign integers (implies ordering — only use for ordinal data like Small/Medium/Large). One-hot for nominal data like zip codes.

11

Question 11: Azure AI Vision — OCR

✓ Correct Answer: C) Azure AI Vision — Optical Character Recognition (OCR)

Why this is correct:

OCR (within Azure AI Vision or Azure AI Document Intelligence) extracts text from images and documents. Azure AI Document Intelligence (Form Recognizer) is the more specialized service for structured form extraction with field mapping, while Azure AI Vision OCR is for general text extraction from images.

Why other answers are incorrect:

A: Image Analysis identifies objects, generates captions, and detects visual features — not for text extraction from documents.
B: Azure AI Document Intelligence (Form Recognizer) is actually the BEST answer for structured forms with field extraction — but the question asks about text from "scanned documents and invoices" which Document Intelligence handles, while Azure AI Vision OCR handles more general text extraction. In practice, Document Intelligence is better for forms.

💡 Key Concept:

OCR in Azure: Azure AI Vision OCR (general text extraction from images), Azure AI Document Intelligence (structured forms, receipts, invoices, IDs — extracts fields with labels). Use Document Intelligence for business documents.

12

Question 12: Image Analysis

✓ Correct Answer: D) Azure AI Vision — Image Analysis (caption and detect objects)

Why this is correct:

Image Analysis in Azure AI Vision can generate natural language captions describing an image, detect and label objects with bounding boxes, identify visual features (colors, brands, adult content), and generate tags. Perfect for auto-captioning product images.

Why other answers are incorrect:

A: OCR extracts text — it doesn't generate descriptive captions or detect objects.
B: Azure AI Face detects and analyzes human faces — not general object detection or captioning.
C: Custom Vision classification categorizes images into user-defined classes — it doesn't generate natural language descriptions.

💡 Key Concept:

Image Analysis features: Smart crops, Auto-captioning (natural language descriptions), Object detection, Tag generation, Brand detection, Color analysis, Background removal, Content moderation.

13

Question 13: Custom Vision

✓ Correct Answer: A) Azure AI Custom Vision

Why this is correct:

Azure AI Custom Vision allows you to train image classification and object detection models with your own labeled images. You don't need ML expertise — upload labeled images, train the model, and deploy it. It's designed for domain-specific visual recognition that general-purpose AI Vision can't do.

Why other answers are incorrect:

B: Azure AI Vision Image Analysis is a general pre-trained model — it recognizes common objects but can't be trained on product-specific defects.
C: Azure ML Designer can train image classification models but requires more ML expertise and setup — Custom Vision is purpose-built for this use case.
D: Document Intelligence is for extracting text/data from documents — not for visual defect inspection.

💡 Key Concept:

Custom Vision vs AI Vision: AI Vision = pre-trained, recognizes thousands of general objects/scenes. Custom Vision = you train it on your specific images and classes. Use Custom Vision when general models don't recognize your specific domain.

14

Question 14: Face Verification

✓ Correct Answer: B) Azure AI Face — Verify (face comparison)

Why this is correct:

Azure AI Face service includes a Verify feature that determines whether two face images are from the same person (1:1 comparison). Given a live camera image and a stored employee photo, Face Verify returns a confidence score indicating if they match. This is exactly the face verification use case.

Why other answers are incorrect:

A: Image Analysis doesn't specifically compare facial identities — it detects faces as objects but doesn't match identities.
C: Custom Vision is for image classification into custom categories — not face identity comparison.
D: Spatial Analysis analyzes people's movements and presence in physical spaces using video — not identity verification.

💡 Key Concept:

Azure AI Face capabilities: Detect (find faces in images), Analyze (age, emotion, attributes), Verify (is this the same person? 1:1), Identify (who is this person? 1:N), Find Similar, Group. Note: some features require Limited Access approval.

15

Question 15: Sentiment Analysis

✓ Correct Answer: C) Sentiment Analysis and Opinion Mining

Why this is correct:

Sentiment Analysis in Azure AI Language classifies text as positive, negative, neutral, or mixed. Opinion Mining (aspect-based sentiment) goes further — it identifies specific aspects being commented on and their sentiment (e.g., "battery life is great" = aspect: battery, sentiment: positive). Perfect for product review analysis.

Why other answers are incorrect:

A: Named Entity Recognition identifies specific entities (people, places, organizations) — not sentiment classification.
B: Key Phrase Extraction identifies the most important phrases and topics in text — not their sentiment.
D: Language Detection identifies what language text is written in — not its sentiment.

💡 Key Concept:

Azure AI Language text analytics features: Sentiment Analysis (positive/negative/neutral), Key Phrase Extraction (important topics), Named Entity Recognition (people/places/orgs), Language Detection, Personally Identifiable Information (PII) extraction.

16

Question 16: Named Entity Recognition

✓ Correct Answer: D) Named Entity Recognition (NER)

Why this is correct:

NER identifies and categorizes named entities in text — people, organizations, locations, dates, monetary amounts, legal case numbers (as custom entities). It provides structured extraction of specific information from unstructured text, exactly what the legal document system needs.

Why other answers are incorrect:

A: Sentiment Analysis measures positive/negative sentiment — not entity extraction.
B: Azure AI Translator translates between languages — not entity extraction.
C: Text Summarization condenses long documents into shorter summaries — doesn't extract specific entity types.

💡 Key Concept:

NER entity types: Person, Organization, Location, DateTime, Quantity, URL, Email, IP Address, Event, Product, Skill. Custom NER allows you to define your own entity types (like legal case numbers).

17

Question 17: Azure AI Translator

✓ Correct Answer: A) Azure AI Translator

Why this is correct:

Azure AI Translator is a neural machine translation service that supports 100+ languages for text translation. It provides real-time translation via REST API, supports language detection, transliteration, and dictionary lookup. Designed exactly for multi-language communication platforms.

Why other answers are incorrect:

B: Language Detection identifies what language text is in — it doesn't translate.
C: Custom Text Classification categorizes text into custom categories — not translation.
D: Azure OpenAI can translate but it's expensive, slower, and less optimized for translation than dedicated Translator service.

💡 Key Concept:

Azure AI Translator: 100+ languages, real-time and batch translation, custom translation (domain-specific vocabulary), language detection included. Use dedicated Translator service rather than GPT for cost-effective high-volume translation.

18

Question 18: Azure AI Speech

✓ Correct Answer: B) Speech to Text (Real-time transcription)

Why this is correct:

Speech to Text (also called Speech Recognition) converts spoken audio into text. Real-time transcription processes audio as it arrives, providing live text output — ideal for call center live monitoring and compliance logging.

Why other answers are incorrect:

A: Text to Speech converts text into spoken audio (opposite direction) — for generating voice responses, not transcribing calls.
C: Speech Translation translates spoken audio directly to text in a different language — the question needs English-to-text transcription, not cross-language translation.
D: Speaker Recognition identifies who is speaking — not transcription of what is being said.

💡 Key Concept:

Azure AI Speech capabilities: Speech to Text (STT), Text to Speech (TTS), Speech Translation (STT + translate), Speaker Recognition (identify/verify speakers), Custom Speech (train for domain vocabulary), Custom Neural Voice.

19

Question 19: Azure Bot Service

✓ Correct Answer: C) Azure Bot Service

Why this is correct:

Azure Bot Service is the orchestration layer for building bots. It handles conversation flow, connects to multiple channels (Teams, web chat, email, Slack, etc.), integrates with Azure AI Language for NLU, and connects to backend APIs for actions like checking account balances. It's the framework that ties everything together.

Why other answers are incorrect:

A: CLU (Conversational Language Understanding) understands user intent and extracts entities from text — it's a component that powers the bot's language understanding, not the bot itself.
B: Question Answering service handles FAQ-style Q&A — it's one capability that can be added to a bot, not the bot framework itself.
D: Azure OpenAI provides generative AI capabilities — it can be a component in a bot but doesn't handle channel connectivity or conversation orchestration.

💡 Key Concept:

Bot architecture: Azure Bot Service (orchestration + channels) + Azure AI Language CLU (intent recognition) + Question Answering (FAQ) + Azure OpenAI (generative responses) + your APIs (backend actions). Bot Service connects all channels.

20

Question 20: Question Answering

✓ Correct Answer: D) Question Answering (formerly QnA Maker)

Why this is correct:

Azure AI Language Question Answering (the successor to QnA Maker) is designed specifically for building FAQ-style Q&A systems. You provide documents or FAQ lists, it creates a knowledge base, and then answers user questions by finding the best matching answer. It's the simplest path for FAQ bots.

Why other answers are incorrect:

A: CLU is for understanding user intents and entities in conversational commands (e.g., "Book me a flight to London") — not for matching questions to FAQ answers.
B: Custom Text Classification assigns documents to categories — not question-answer matching.
C: Azure OpenAI with RAG is more powerful but much more complex to implement than Question Answering for a simple FAQ use case.

💡 Key Concept:

QA vs CLU: Question Answering = FAQ matching, finds best answer from knowledge base. CLU = intent classification (user wants to do X) and entity extraction (extract "London" from "fly to London"). Use QA for FAQ bots, CLU for action-oriented bots.

21

Question 21: Large Language Models

✓ Correct Answer: A) Azure OpenAI Service

Why this is correct:

Azure OpenAI Service provides access to OpenAI's large language models (GPT-4, GPT-4o, DALL-E, Whisper) through a REST API with enterprise-grade security, compliance, and private networking. It supports text generation, question answering, summarization, and code generation use cases.

Why other answers are incorrect:

B: Azure ML AutoML trains ML models for structured prediction tasks — not generative text generation.
C: Azure AI Language Text Summarization uses extractive/abstractive summarization — it's a specific feature, not a general LLM for all generative tasks.
D: Azure AI Search with semantic ranking improves search results — it doesn't generate text or write code.

💡 Key Concept:

Azure OpenAI Service models (2026): GPT-4o, GPT-4o mini (cost-effective), GPT-4 Turbo (128K context), o1/o3 (reasoning models), DALL-E 3 (images), Whisper (speech), text-embedding-3 (embeddings). Accessed via REST API or SDKs.

22

Question 22: Prompt Engineering

✓ Correct Answer: B) In the system message (system prompt)

Why this is correct:

The system message (or system prompt) is a special input in Azure OpenAI's chat completions API that sets the assistant's behavior, persona, and constraints for the entire conversation. It's processed before every user message and shapes how the model responds throughout the conversation — perfect for persistent instructions.

Why other answers are incorrect:

A: Inserting instructions in each user message is redundant, increases token cost, and can be confusing for the model — system message is the right place for persistent instructions.
C: Fine-tuning trains the model on examples of desired behavior — expensive, requires many examples, and not needed for simple behavioral rules.
D: Azure OpenAI deployment configuration controls model settings (temperature, max tokens) but not behavioral instructions like "respond formally."

💡 Key Concept:

ChatCompletion messages: System (instructions, persona) → User (what the user says) → Assistant (model response) → User → Assistant... The system message is sent once but applies to the whole conversation.

23

Question 23: Retrieval Augmented Generation

✓ Correct Answer: C) Retrieval Augmented Generation (RAG) using Azure AI Search + Azure OpenAI

Why this is correct:

RAG combines a search/retrieval system (Azure AI Search) with a generative model (Azure OpenAI). Documents are indexed in Azure AI Search. When a question is asked, relevant document chunks are retrieved and provided as context to GPT, which then generates a grounded answer based only on the provided context — not its general training data.

Why other answers are incorrect:

A: Fine-tuning trains the model on new data but is expensive, doesn't update easily, and the model may still "hallucinate" from general training. RAG is preferred for knowledge-grounded Q&A.
B: A larger model has more general knowledge but still can't answer questions about specific private documents it was never trained on.
D: Temperature controls randomness/creativity — a lower temperature makes output more predictable but doesn't constrain it to specific documents.

💡 Key Concept:

RAG architecture: Chunked documents → Embedded as vectors → Stored in Azure AI Search → Question asked → Search retrieves relevant chunks → Chunks + question sent to GPT → GPT answers based on retrieved context. "Add your data" in Azure OpenAI Studio implements RAG.

24

Question 24: Responsible Generative AI

✓ Correct Answer: D) Azure AI Content Safety / Content Filters in Azure OpenAI

Why this is correct:

Azure OpenAI has built-in content filters that automatically screen both inputs (user messages) and outputs (model responses) for harmful content categories: hate, sexual content, violence, and self-harm. The filters can be configured for different severity thresholds and will block violating content before it reaches users.

Why other answers are incorrect:

A: System message instructions can discourage harmful content but they're not a reliable safety mechanism — they can be circumvented by adversarial prompts.
B: Prompt engineering is important for quality but isn't a technical safety control — users can still try to elicit harmful content.
C: Temperature controls randomness — it doesn't filter harmful content categories.

💡 Key Concept:

Azure OpenAI content filters: 4 categories (Hate, Sexual, Violence, Self-harm), 4 severity levels (Safe, Low, Medium, High). Default filtering is enabled. Can customize thresholds with Azure approval. Azure AI Content Safety provides standalone filtering APIs.

25

Question 25: Microsoft 365 Copilot

✓ Correct Answer: A) Microsoft 365 Copilot

Why this is correct:

Microsoft 365 Copilot is an AI assistant embedded directly in Microsoft 365 applications (Outlook, Word, Excel, PowerPoint, Teams, etc.). It uses GPT-4 models via Azure OpenAI with access to your Microsoft 365 data (emails, documents, calendar) to provide contextual assistance — summarizing emails, drafting documents, creating presentations, generating Excel formulas.

Why other answers are incorrect:

B: Azure OpenAI Service API is the developer API for building applications — it's not a ready-to-use end-user assistant embedded in Microsoft 365 apps.
C: Azure AI Language provides NLP capabilities for developers — it's not an end-user productivity assistant in Microsoft 365.
D: Bing Chat Enterprise (now Microsoft Copilot for Microsoft 365) is a related but separate product for web-based AI search — not embedded in productivity apps.

💡 Key Concept:

Microsoft Copilot ecosystem: Microsoft 365 Copilot (Office apps), GitHub Copilot (coding), Copilot in Windows, Bing Copilot (web search), Copilot Studio (build custom Copilots), Microsoft Copilot for Security (SOC). All powered by Azure OpenAI.

📊 How Did You Score?

23–25
Exam Ready
Excellent! Schedule your exam.
19–22
Almost There
Review generative AI and Azure service mapping.
14–18
Keep Studying
Focus on which Azure AI service does what.
0–13
More Study Needed
Work through the AI-900 study guide.

Ready for More AI-900 Practice?

These 25 questions are just a sample. The actual AI-900 exam has 40–60 questions.

MSCertQuiz AI-900 includes 500 questions covering:

  • ✓ All Azure AI services with service-selection scenarios
  • ✓ Responsible AI principles with scenario matching
  • ✓ Azure OpenAI Service, prompt engineering, and RAG
  • ✓ Computer vision, NLP, and speech services in depth
  • ✓ Latest exam content including Microsoft Copilot ecosystem

$14.99 One-Time Payment

Lifetime access • No subscription • 500 questions