This chapter covers the essential exam strategy for the AI-900 certification, focusing on how to approach scenario-based questions effectively. Understanding this topic is critical because scenario questions make up a significant portion of the exam, testing your ability to apply Azure AI concepts to real-world situations. Approximately 40-50% of AI-900 questions present a business scenario and ask you to choose the appropriate AI service or principle. Mastering the strategy for these questions can mean the difference between passing and failing. This chapter will dissect the question patterns, common traps, and provide a step-by-step approach to maximize your score.
Jump to a section
Think of the AI-900 exam as a pilot preparing for a flight. A pilot doesn't just jump into the cockpit and start pressing buttons; they follow a structured preflight checklist to ensure every system is working and they are prepared for any scenario. Similarly, for AI-900, you need a systematic strategy: first, understand the exam blueprint (the flight plan), then practice with scenario questions (simulators), and finally, review key concepts (the checklist). Just as a pilot checks weather, fuel, and instruments before takeoff, you must check your understanding of Azure AI services, responsible AI principles, and workload types. The exam scenario questions are like emergency procedures—they test your ability to apply knowledge under pressure. Without a strategy, you might panic and choose the wrong answer, just as a pilot without a checklist might forget to deploy landing gear. This analogy emphasizes that exam strategy is not about memorization but about having a repeatable process to evaluate options and select the best solution, mirroring how a pilot methodically ensures a safe flight.
Understanding the AI-900 Exam Structure
The AI-900 exam, officially titled 'Microsoft Azure AI Fundamentals,' is designed for candidates with non-technical backgrounds, such as business stakeholders, sales professionals, or students. It validates foundational knowledge of AI concepts and Azure services. The exam consists of 40-60 questions, with a time limit of 85 minutes. Question types include multiple-choice, drag-and-drop, and scenario-based questions. The passing score is 700 out of 1000.
Scenario-based questions are the most challenging because they require you to analyze a business problem and select the best Azure AI service or principle. These questions often include a paragraph describing a company's need, followed by four options. The key is to map the scenario to the correct service based on its capabilities.
The Scenario Question Strategy Framework
To tackle scenario questions effectively, use the 'S.T.E.P.' framework: - S – Scan the scenario for keywords (e.g., 'sentiment,' 'anomaly,' 'extract text from images'). - T – Translate the business need into a technical requirement (e.g., 'analyze customer feedback' → 'text analytics'). - E – Evaluate each option against the requirement, eliminating mismatches. - P – Pick the best fit, considering cost, complexity, and Azure's specific service boundaries.
Common Scenario Patterns on AI-900
The exam tests specific Azure AI services in scenarios. Here are the most common patterns:
#### Pattern 1: Text Analysis - Keywords: 'sentiment,' 'key phrases,' 'language detection,' 'extract information from text.' - Correct Service: Azure Cognitive Service for Language (formerly Text Analytics). - Trap: Choosing Azure Bot Service (for conversational AI) or Azure Cognitive Search (for indexing).
#### Pattern 2: Image and Video Analysis - Keywords: 'identify objects in images,' 'extract text from images,' 'moderate content.' - Correct Service: Azure Cognitive Service for Vision (Computer Vision) or Azure AI Content Safety. - Trap: Choosing Form Recognizer (now Azure AI Document Intelligence) for structured data extraction from forms, not general images.
#### Pattern 3: Speech Processing - Keywords: 'transcribe speech to text,' 'convert text to speech,' 'speaker recognition.' - Correct Service: Azure Cognitive Service for Speech. - Trap: Choosing Azure Translator for text translation, not speech.
#### Pattern 4: Decision Support - Keywords: 'detect anomalies,' 'moderate content,' 'personalize recommendations.' - Correct Service: Anomaly Detector, Content Safety, or Personalizer (all part of Azure Cognitive Services). - Trap: Choosing a generic service like Azure Machine Learning for a specific pre-built API.
#### Pattern 5: Responsible AI - Keywords: 'fairness,' 'transparency,' 'accountability,' 'inclusiveness,' 'privacy.' - Correct Principle: The six Microsoft responsible AI principles: fairness, reliability and safety, privacy and security, inclusiveness, transparency, and accountability. - Trap: Confusing principles with each other (e.g., 'transparency' vs. 'accountability').
How to Eliminate Wrong Answers
Most wrong answers are plausible but not the best fit. Use these elimination techniques:
Service Mismatch: If the scenario involves images, eliminate speech or text services.
Complexity Overkill: If a pre-built API can solve the problem, avoid options that require custom ML training (e.g., Azure Machine Learning).
Cost Mismatch: The exam sometimes tests cost awareness. For example, using a serverless service like Cognitive Services is cheaper than provisioning dedicated VMs.
Principle Confusion: For responsible AI, remember that 'transparency' means users know when they are interacting with AI, while 'accountability' means the organization takes ownership of AI outcomes.
Detailed Mechanism of Scenario Question Processing
When you read a scenario question, your brain must process the text and map it to Azure services. This involves: - Input: The scenario description (e.g., 'A retail company wants to analyze customer reviews to detect negative feedback.') - Mapping: Identify keywords ('analyze,' 'customer reviews,' 'negative feedback') → 'sentiment analysis.' - Service Selection: Azure Cognitive Service for Language offers sentiment analysis. - Verification: Check if the service is available in the region, pricing tier, and any limitations (e.g., language support).
Exam Blueprint Alignment
The AI-900 exam objectives (from Microsoft) include: - Describe AI workloads and considerations (15-20%): This includes responsible AI principles. - Describe fundamental principles of machine learning on Azure (20-25%): Focuses on ML concepts, not specific services. - Describe features of computer vision workloads on Azure (15-20%): Covers Computer Vision, Custom Vision, Face, and Form Recognizer. - Describe features of Natural Language Processing (NLP) workloads on Azure (15-20%): Covers Language service, Translator, and Speech. - Describe features of conversational AI workloads on Azure (15-20%): Covers Bot Service and QnA Maker (now part of Language service).
Scenario questions can come from any of these domains, but the most common are from NLP, Computer Vision, and Responsible AI.
Common Exam Traps
Choosing Azure Machine Learning when a Cognitive Service works: Many candidates think they need to train a custom model, but the exam tests pre-built services.
Confusing Content Safety with Computer Vision: Content Safety can moderate text and images, but for image moderation specifically, Computer Vision's 'moderate content' feature might be more appropriate.
Mixing up Text Analytics and Translator: Text Analytics detects language and sentiment; Translator translates text. A scenario asking for 'translate customer emails' needs Translator, not Text Analytics.
Forgetting region availability: Some services are not available in all regions. The exam may test this by listing regions.
Step-by-Step Example
Let's walk through a sample question:
Question: A company wants to build a chatbot that answers user questions about their products. They have a list of FAQs in a PDF. Which Azure service should they use? - A. Azure Bot Service - B. QnA Maker (now part of Language service) - C. Azure Cognitive Search - D. Azure Machine Learning
Step 1: Scan keywords: 'chatbot,' 'answers questions,' 'FAQs.' Step 2: Translate: The need is to create a Q&A system from existing FAQs. Step 3: Evaluate:
- A: Bot Service is the framework for building bots, but it needs a knowledge base. - B: QnA Maker (Language service) allows you to create a knowledge base from a PDF and answer questions. - C: Cognitive Search is for indexing documents, not answering questions directly. - D: Machine Learning is overkill. Step 4: Pick B. QnA Maker (now part of Language service) is the correct answer.
Advanced Strategy: Reading Between the Lines
Sometimes the scenario includes subtle clues like 'real-time,' 'pre-built,' 'no coding.' These indicate a Cognitive Service. If the scenario mentions 'custom model' or 'train with own data,' then Custom Vision or Custom Text (Language service) might be appropriate. Also, note 'serverless' or 'pay-as-you-go' hints at Cognitive Services, while 'dedicated compute' hints at Azure Machine Learning.
Time Management
With 85 minutes for 40-60 questions, you have about 1.5-2 minutes per question. For scenario questions, spend up to 2 minutes reading and analyzing. If stuck, eliminate two options and guess. Mark questions for review if time allows.
Summary of Key Services and Their Scenarios
Computer Vision: Extract text from images, identify objects, moderate content.
Custom Vision: Classify images with custom categories.
Face: Detect and recognize faces.
Form Recognizer: Extract data from forms (now Document Intelligence).
Language: Sentiment, key phrases, language detection, QnA, custom text classification.
Translator: Text translation.
Speech: Speech-to-text, text-to-speech, speaker recognition.
Content Safety: Moderate text and images for offensive content.
Anomaly Detector: Detect anomalies in time-series data.
Personalizer: Provide personalized recommendations.
Bot Service: Build conversational bots.
Azure Machine Learning: Build, train, and deploy custom ML models.
Remember, the exam tests your ability to choose the right tool for the job, not your ability to implement it.
Scan the Scenario for Keywords
Read the scenario quickly and underline or mentally note keywords that indicate the type of AI workload. For example, words like 'sentiment,' 'key phrases,' or 'language detection' point to NLP. 'Identify objects,' 'extract text from images,' or 'moderate content' point to Computer Vision. 'Transcribe,' 'speech,' or 'speaker' point to Speech services. 'Anomalies,' 'outliers,' or 'unusual patterns' point to Anomaly Detector. 'Fairness,' 'transparency,' or 'bias' point to responsible AI principles. This step sets the direction for the correct Azure service.
Translate Business Need to Technical Requirement
Convert the business problem into a clear technical requirement. For instance, 'analyze customer feedback to understand sentiment' becomes 'perform sentiment analysis on text.' 'Build a chatbot that answers FAQs' becomes 'create a Q&A knowledge base.' 'Detect offensive content in images' becomes 'image moderation.' This translation helps map the requirement to a specific Azure AI service. Avoid jumping to conclusions; ensure the requirement is precise.
Evaluate Each Option Against Requirement
For each answer choice, ask: 'Does this service directly address the technical requirement?' Eliminate any option that does not match. For example, if the requirement is 'extract text from scanned documents,' Computer Vision (OCR) is correct, but Form Recognizer (Document Intelligence) is also possible if the documents are structured forms. However, if the scenario doesn't mention forms, Computer Vision is more appropriate. Watch for options that are partially correct but not the best fit.
Eliminate Obvious Mismatches First
Quickly cross out options that are clearly wrong based on the workload type. For example, if the scenario is about text, eliminate all Computer Vision and Speech services. If it's about images, eliminate Language and Speech services. This reduces the number of options to consider. Also eliminate any service that requires custom training if the scenario mentions 'pre-built' or 'no coding.'
Select the Best Fit Considering Nuances
From the remaining options, choose the one that best matches the scenario. Consider nuances like: Is the data structured or unstructured? Is real-time processing required? Does the scenario mention specific Azure regions? For example, if the scenario requires 'real-time translation of customer chat,' use Translator (text) or Speech Translation (speech). If the scenario mentions 'custom model,' then Custom Vision or Custom Text may be needed. Validate your choice by ensuring it fully meets the requirement without overcomplicating.
In enterprise environments, AI-900 scenario questions mirror real-world decisions made by solution architects. For example, a large e-commerce company wants to automatically moderate user-generated content (product reviews and images) to ensure compliance with community guidelines. The architect evaluates Azure AI Content Safety for text and images, but also considers Computer Vision for image moderation if more granular object detection is needed. In production, Content Safety is deployed with a REST API endpoint, configured with content categories (e.g., hate speech, violence). The company sets up thresholds for severity levels and integrates with Azure Logic Apps to trigger manual review when content exceeds a certain score. A common pitfall is misconfiguring the severity levels, causing false positives that block legitimate content, or false negatives that allow harmful content. The architect must also consider cost: Content Safety has a pay-per-transaction pricing, so high volume requires cost analysis.
Another scenario: A healthcare provider wants to extract patient data from handwritten forms (e.g., intake forms). They use Azure AI Document Intelligence (formerly Form Recognizer) with custom models trained on their specific form templates. In production, they preprocess scanned images using Computer Vision's OCR to enhance text extraction. The system processes thousands of forms daily, and the architect must ensure the service scales with throughput. A common misconfiguration is not using the correct model ID or failing to handle low-quality images, leading to extraction errors. The responsible AI principle of 'privacy and security' requires that patient data is encrypted in transit and at rest, and that the service is deployed in a compliant region (e.g., Azure Government for HIPAA).
A third scenario: A financial services firm wants to detect anomalies in transaction data to identify potential fraud. They use Azure Anomaly Detector, which is a pre-built API. In production, they ingest time-series data from their database, set the sensitivity level (e.g., 95% confidence), and configure the granularity (e.g., hourly). The architect must handle edge cases like seasonal patterns (e.g., higher spending during holidays) to avoid false alarms. A common mistake is not tuning the 'period' parameter, causing the model to miss seasonal anomalies. The firm integrates Anomaly Detector with Azure Logic Apps to send alerts to fraud analysts. The exam scenario might test your ability to choose Anomaly Detector over custom ML for a pre-built solution.
The AI-900 exam focuses on scenario questions that test your ability to match business needs to Azure AI services. The specific objective codes are: - 1.1 Describe AI workloads and considerations: This includes responsible AI principles and the types of AI workloads (Computer Vision, NLP, etc.). - 1.2 Describe fundamental principles of machine learning on Azure: This is less about scenarios and more about concepts like regression, classification, and clustering. - 2.1 Describe features of computer vision workloads: Scenarios about image classification, object detection, OCR, and face detection. - 3.1 Describe features of NLP workloads: Scenarios about sentiment analysis, key phrase extraction, language detection, translation, and speech. - 4.1 Describe features of conversational AI workloads: Scenarios about chatbots and QnA.
Common wrong answers and why candidates choose them: 1. Choosing Azure Machine Learning instead of Cognitive Services: Candidates think 'AI' always requires custom ML, but the exam tests pre-built services. They overlook keywords like 'pre-built' or 'no coding.' 2. Confusing Text Analytics with Translator: Both deal with text, but one analyzes, the other translates. Candidates miss the keyword 'translate.' 3. Selecting Bot Service for QnA: Bot Service is the framework, but QnA Maker (Language service) provides the knowledge base. Candidates see 'chatbot' and automatically choose Bot Service. 4. Choosing Computer Vision for form extraction: Computer Vision can do OCR, but Form Recognizer (Document Intelligence) is specialized for structured forms. Candidates ignore the word 'forms.'
Specific numbers and terms that appear verbatim: - Six responsible AI principles: fairness, reliability and safety, privacy and security, inclusiveness, transparency, accountability. - Azure Cognitive Services: The term 'Cognitive Services' is often used; know that it includes Vision, Language, Speech, and Decision services. - QnA Maker: Even though it's now part of Language service, the exam may still use the old name. - Custom Vision: For custom image classification, not general object detection.
Edge cases the exam loves: - Language support: Some services support only certain languages. For example, sentiment analysis may not support all languages equally. - Real-time vs. batch: Some services offer both, but the scenario may specify 'real-time' to eliminate batch-only options. - Cost: The exam may ask for the 'most cost-effective' solution, which often points to Cognitive Services over custom ML.
How to eliminate wrong answers using the underlying mechanism:
If the scenario involves 'text' and 'analyze,' eliminate any option that deals with images or speech.
If the scenario mentions 'train with your own data,' consider Custom Vision or Custom Text, not pre-built APIs.
If the scenario mentions 'fairness' or 'bias,' the answer is a responsible AI principle, not a service.
Scenario questions test your ability to match business needs to Azure AI services, not implementation details.
Use the S.T.E.P. framework: Scan, Translate, Evaluate, Pick.
Keywords are critical: 'sentiment' -> Language, 'object detection' -> Computer Vision, 'transcribe' -> Speech.
Pre-built Cognitive Services are usually the correct answer for common AI tasks.
Custom ML (Azure Machine Learning) is only needed for unique problems not solved by pre-built APIs.
Responsible AI principles: fairness, reliability and safety, privacy and security, inclusiveness, transparency, accountability.
Eliminate wrong answers by checking workload type (text, image, speech, etc.) and complexity (pre-built vs. custom).
The exam may test cost-awareness: prefer serverless Cognitive Services over custom ML for cost efficiency.
Be aware of service limits: e.g., Language service supports 120+ languages for detection but fewer for sentiment.
QnA Maker is now part of Language service but the exam may still use the old name.
These come up on the exam all the time. Here's how to tell them apart.
Azure Cognitive Services
Pre-built AI APIs, no training required
Pay-per-use pricing, serverless
Best for common AI tasks like sentiment analysis, OCR, object detection
Limited customization (e.g., Custom Vision allows training)
Fewer lines of code, faster time to value
Azure Machine Learning
Custom ML model training and deployment
Requires data, compute, and ML expertise
Best for unique or complex AI problems
Full control over algorithms and data
Higher cost and complexity
Computer Vision
General image analysis: object detection, OCR, description
Works with any image or video
Pre-built models for common tasks
No understanding of form structure
Can extract text from images but not key-value pairs from forms
Form Recognizer (Document Intelligence)
Specialized for extracting data from forms and documents
Understands form fields, tables, and signatures
Requires training on custom forms (or pre-built models)
Outputs structured data (key-value pairs)
Better for invoice, receipt, and form processing
Mistake
AI-900 scenario questions require deep technical knowledge of Azure implementation.
Correct
The exam tests foundational knowledge, not implementation details. You need to understand what each service does at a high level, not how to deploy it. Focus on mapping scenarios to services.
Mistake
All AI services on Azure are part of 'Cognitive Services'.
Correct
Cognitive Services is a specific category of pre-built AI APIs. Other AI services include Azure Machine Learning, Bot Service, and Cognitive Search. Not all AI services are Cognitive Services.
Mistake
If a scenario mentions 'custom model,' you should always choose Custom Vision or Custom Text.
Correct
Sometimes 'custom model' refers to training a model in Azure Machine Learning, not necessarily using Custom Vision. Read carefully: if the scenario involves images, Custom Vision is correct; if it involves text, Custom Text (Language service) is correct; if it involves general ML, Azure Machine Learning is correct.
Mistake
The exam always expects the most advanced or expensive solution.
Correct
The exam often tests cost-effectiveness. Pre-built Cognitive Services are usually the cheapest and easiest. Avoid over-engineering with custom ML when a pre-built API suffices.
Mistake
Responsible AI principles are only tested in dedicated questions about ethics.
Correct
Responsible AI principles can appear in scenario questions where you must choose the correct principle for a given situation. For example, a scenario about explaining AI decisions tests 'transparency'.
Reveal each answer, then mark whether you got it right. Score 60%+ to unlock the next chapter.
First, scan the scenario for keywords that indicate the type of AI workload (e.g., 'sentiment' for NLP, 'object detection' for Computer Vision). Then, translate the business need into a technical requirement. Next, evaluate each answer option against that requirement, eliminating those that don't match. Finally, pick the best fit, considering cost and complexity. Practice with sample questions to build speed.
Cognitive Services are pre-built AI APIs that require no training. They are ready to use via REST API or SDK. Azure Machine Learning is a platform for building, training, and deploying custom ML models. Use Cognitive Services for common tasks (sentiment analysis, OCR) and Azure ML for unique problems that need custom models. The exam often favors Cognitive Services for scenario questions.
Use Azure Cognitive Service for Vision (Computer Vision) for general OCR from images. If the images are documents like invoices or forms, use Azure AI Document Intelligence (formerly Form Recognizer) for structured extraction. The exam may test the distinction: Computer Vision for simple text extraction, Document Intelligence for form fields.
You may be given a scenario describing an AI system and asked which principle it aligns with. For example, a scenario about explaining how an AI model makes decisions tests 'transparency.' Another about ensuring the system works for all users tests 'inclusiveness.' Memorize the six principles and their definitions.
The most common mistake is choosing Azure Machine Learning for tasks that can be done with pre-built Cognitive Services. Many candidates think all AI requires custom training. Also, confusing similar services like Text Analytics and Translator, or Computer Vision and Form Recognizer. Practice eliminating options based on keywords.
Azure Bot Service is a framework for building bots, but it needs a knowledge base to answer questions. The correct service for creating a Q&A knowledge base from FAQs is QnA Maker (now part of Language service). The bot can then use that knowledge base. So the direct answer is QnA Maker, not Bot Service.
Anomaly detection is identifying unusual patterns in data that deviate from the norm. Azure Anomaly Detector is a pre-built service that applies to time-series data. It is used for fraud detection, equipment monitoring, etc. The exam may present a scenario about detecting unusual transactions or sensor readings.
You've just covered AI-900 Exam Strategy and Scenario Questions — now see how well it sticks with free AI-900 practice questions. Full explanations included, no account needed.
Done with this chapter?