AI-900 — Microsoft Azure AI Fundamentals Scenario Practice Guide

The Microsoft Azure AI Fundamentals (AI-900) exam tests whether you can recognize AI workloads, match business needs to Azure AI capabilities, and reason about responsible AI concepts at a fundamental level. Scenario questions are less about memorizing every feature and more about selecting the most appropriate service, workload type, or design principle from a short set of facts.

Use this guide to slow down, identify the actual decision point, and choose the most defensible answer when a scenario includes extra details.

What an AI-900 scenario is usually asking you to decide

AI-900 scenarios often describe a business goal, data type, application requirement, or ethical concern. Your job is to translate that scenario into the right AI concept or Azure AI service family.

Common decision areas include:

• AI workload type
  • Prediction, classification, regression, clustering, anomaly detection, computer vision, natural language processing, generative AI, speech, translation, or document processing.
• Azure AI capability
  • Azure AI Vision, Azure AI Language, Azure AI Speech, Azure AI Translator, Azure AI Document Intelligence, Azure Machine Learning, Azure AI Search, or Azure OpenAI-related capabilities where relevant to the tested objective.
• Machine learning task
  • Supervised versus unsupervised learning, classification versus regression, training versus inferencing, model evaluation, or automated ML concepts.
• Responsible AI principle
  • Fairness, reliability and safety, privacy and security, inclusiveness, transparency, or accountability.
• Implementation approach
  • Use a prebuilt AI service, train a custom model, build a search solution, analyze text, extract document fields, generate content, or classify images.

The best answer is usually the one that directly satisfies the stated requirement with the simplest suitable Azure AI capability.

Start with the final sentence

In scenario questions, the final sentence often contains the actual instruction:

• “Which service should you use?”
• “Which type of machine learning model is most appropriate?”
• “Which responsible AI principle is demonstrated?”
• “Which feature should be used to extract the required data?”
• “Which workload does the scenario describe?”

Before evaluating the answer choices, restate the question in your own words.

Example:

A company wants to analyze customer comments and determine whether each comment is positive, neutral, or negative. Which capability should it use?

Restated decision:

This is asking for sentiment detection from text, so I should look for a natural language capability, not image analysis, speech recognition, or document extraction.

That short restatement helps prevent you from being pulled toward services mentioned elsewhere in the scenario.

Identify the environment and input type

For AI-900, the type of input data is one of the strongest clues. Before choosing an answer, ask: What kind of data is being processed?

Text input

Text scenarios often point to natural language processing.

Look for phrases such as:

• Customer reviews
• Support tickets
• Emails
• Chat messages
• Key phrases
• Named entities
• Sentiment
• Language detection
• Summarization or content generation

Likely service area:

• Azure AI Language for text analytics tasks such as sentiment analysis, key phrase extraction, entity recognition, and language understanding.
• Azure AI Translator when the core requirement is translation between languages.
• Azure OpenAI-related capabilities when the requirement is generative content, conversational responses, summarization, or natural language generation, depending on the options provided.

Images or video frames

Visual input points toward computer vision.

Look for phrases such as:

• Identify objects in photos
• Detect faces or visual features
• Read printed or handwritten text from an image
• Generate captions for images
• Moderate visual content

Likely service area:

• Azure AI Vision for image analysis, object detection, image classification, OCR-style image text extraction, and visual description scenarios.

Audio input

Audio points toward speech.

Look for phrases such as:

• Convert speech to text
• Convert text to speech
• Translate spoken audio
• Transcribe calls
• Recognize spoken commands

Likely service area:

• Azure AI Speech for speech recognition, speech synthesis, and speech translation scenarios.

Forms, invoices, receipts, or structured documents

Documents with fields point toward document intelligence rather than general image analysis.

Look for phrases such as:

• Extract invoice number, vendor, total, and due date
• Process receipts
• Read forms
• Extract key-value pairs
• Analyze structured or semi-structured documents

Likely service area:

• Azure AI Document Intelligence when the goal is to extract fields, tables, and structured information from documents.

Business records or tabular data

Rows and columns often point toward machine learning.

Look for phrases such as:

• Predict sales
• Estimate price
• Classify customers
• Detect unusual transactions
• Group similar records
• Forecast demand

Likely service area:

• Azure Machine Learning or a machine learning task type, depending on the answer choices.

Find the goal, not just the topic

A scenario may mention several technologies, but only one goal matters. Identify the verb that describes what the system must do.

Use this quick mapping:

• Predict a numeric value: regression
• Predict a category or label: classification
• Group similar items without predefined labels: clustering
• Find unusual behavior: anomaly detection
• Understand or extract meaning from text: natural language processing
• Analyze images: computer vision
• Convert speech and text: speech service
• Translate language: translation
• Extract fields from documents: document intelligence
• Generate text, answer prompts, or summarize content: generative AI
• Search across indexed content with AI enrichment: AI-powered search

Mini example: classify or predict?

Scenario:

A retailer wants to use past customer data to determine whether a customer is likely to cancel a subscription.

Decision:

• The output is a category: likely to cancel or not likely to cancel.
• This is a classification scenario.

If the scenario instead asks to estimate the amount a customer will spend next month, the output is numeric, so it is a regression scenario.

Separate constraints from preferences

AI-900 scenarios often include business constraints. Some are essential. Others are background details.

A constraint is something the answer must satisfy:

• Must analyze text, not audio.
• Must translate between languages.
• Must extract specific fields from invoices.
• Must use a prebuilt model rather than building a custom model.
• Must ensure users understand why a decision was made.
• Must protect personal data.
• Must reduce bias across demographic groups.
• Must provide a conversational interface.

A preference may be useful context but does not define the answer:

• The company is global.
• The app is used by sales employees.
• The data is stored in the cloud.
• The team wants to improve productivity.
• The solution will be used by customers.

Do not choose an answer just because it relates to the company background. Choose the one that satisfies the must-have requirement.

Practical reading habit

Mark each scenario fact mentally as one of these:

• Input: What data is provided?
• Output: What result is required?
• Constraint: What must the solution do or avoid?
• Principle: Is this about responsible AI?
• Distractor context: What is interesting but not decisive?

Match the Azure AI service to the requirement

AI-900 is a fundamentals exam, so service selection questions usually reward clear service-to-use-case matching.

Azure AI Vision

Choose a vision-related answer when the scenario is about images or visual content.

Typical clues:

• Detecting objects in images
• Categorizing images
• Describing image content
• Reading text from images
• Identifying visual features

Reasoning pattern:

The input is an image, and the required output is information about the image. This is a computer vision workload.

Azure AI Language

Choose a language-related answer when the scenario is about understanding text.

Typical clues:

• Sentiment analysis
• Key phrase extraction
• Entity recognition
• Language detection
• Text classification
• Conversational language understanding, when the scenario focuses on interpreting user intent

Reasoning pattern:

The input is written text, and the required output is meaning, sentiment, entities, or intent. This is a natural language processing workload.

Azure AI Speech

Choose a speech-related answer when the scenario involves spoken audio.

Typical clues:

• Speech-to-text transcription
• Text-to-speech voice output
• Real-time speech translation
• Call center audio analysis at a basic recognition level

Reasoning pattern:

The input or output is spoken audio. This is a speech workload, even if the speech is later analyzed as text.

Azure AI Translator

Choose translation when the core requirement is converting content from one language to another.

Typical clues:

• Translate documents or messages
• Support users in multiple languages
• Convert text from English to another language
• Translate spoken content when paired with speech requirements

Reasoning pattern:

The main decision is language conversion. If the scenario only asks to identify sentiment, translation is not the main answer.

Azure AI Document Intelligence

Choose document intelligence when the scenario is about extracting structured information from forms and documents.

Typical clues:

• Invoices
• Receipts
• Tax forms
• Purchase orders
• Key-value pairs
• Tables
• Fields such as total, date, vendor, or account number

Reasoning pattern:

The input is a document, and the required output is structured data extracted from the document.

Azure Machine Learning

Choose machine learning platform concepts when the scenario is about building, training, evaluating, deploying, or managing custom models.

Typical clues:

• Train a model using historical data
• Use automated machine learning
• Evaluate model performance
• Deploy a model endpoint
• Track experiments
• Use datasets and compute for model development

Reasoning pattern:

The scenario is not just consuming a prebuilt AI capability. It is about creating or managing a predictive model.

Azure AI Search

Choose search when the scenario focuses on indexing and retrieving information from content.

Typical clues:

• Search across documents
• Index content
• Enrich search results with AI
• Find relevant documents by query
• Build a knowledge mining solution

Reasoning pattern:

The goal is not just extracting one field or analyzing one text item. The goal is discoverability across a searchable content set.

Use the simplest suitable service

For AI-900 scenarios, avoid overbuilding the solution in your head. If the requirement can be met by a prebuilt Azure AI service, that is often more appropriate than training a custom model.

Ask:

1. Is the task a common AI capability, such as sentiment analysis, OCR, translation, speech-to-text, or invoice extraction?
2. Does the scenario require custom training, custom labels, or a specialized model?
3. Are the answer choices asking for a service family or a machine learning task?

If the scenario says:

A company wants to identify whether support messages are positive or negative.

A prebuilt language capability is more defensible than creating a custom deep learning model, unless the scenario specifically says the company needs a custom model for domain-specific labels.

Read responsible AI scenarios as principle-identification questions

Some AI-900 scenarios are not asking for a service. They are asking which responsible AI principle best applies.

When you see words about ethics, trust, user impact, bias, data protection, or governance, pause before choosing a technical service.

Fairness

Choose fairness when the scenario focuses on avoiding unequal treatment or biased outcomes.

Clues:

• The model performs worse for one demographic group.
• Loan approvals differ unfairly across groups.
• Hiring recommendations disadvantage certain applicants.
• The team tests outcomes across user populations.

Decision question:

Is the concern about equitable treatment and reducing bias?

Reliability and safety

Choose reliability and safety when the scenario focuses on dependable behavior, error handling, and preventing harm.

Clues:

• The system must operate safely under unexpected conditions.
• Incorrect predictions could harm users.
• The team monitors performance before deployment.
• The model should fail gracefully.

Decision question:

Is the concern about whether the AI system behaves safely and consistently?

Privacy and security

Choose privacy and security when the scenario focuses on protecting data and controlling access.

Clues:

• Personal data must be protected.
• Access should be restricted.
• Data should be secured.
• Sensitive information must not be exposed.

Decision question:

Is the concern about safeguarding data and user information?

Inclusiveness

Choose inclusiveness when the scenario focuses on designing for people with different abilities, languages, environments, or access needs.

Clues:

• The system should support users with disabilities.
• The interface should work for diverse users.
• The solution should accommodate different accessibility needs.

Decision question:

Is the concern about ensuring the system can be used by as many people as possible?

Transparency

Choose transparency when the scenario focuses on making AI behavior understandable.

Clues:

• Users should know they are interacting with AI.
• The organization should explain how the model makes decisions.
• The team should document model limitations.
• Stakeholders need understandable information about the system.

Decision question:

Is the concern about explainability, disclosure, or understanding?

Accountability

Choose accountability when the scenario focuses on human responsibility, governance, oversight, and ownership.

Clues:

• A team is assigned to review AI decisions.
• Humans remain responsible for outcomes.
• The organization defines governance processes.
• Someone must be answerable for system behavior.

Decision question:

Is the concern about who is responsible for the AI system and its impact?

Decide the machine learning task from the output

Machine learning task questions become easier when you ignore the business story and focus on the predicted output.

Classification

Use classification when the output is a label or category.

Examples:

• Fraudulent or not fraudulent
• High risk, medium risk, or low risk
• Approved or denied
• Product category
• Customer churn or no churn

Question to ask:

Is the model choosing from predefined categories?

Regression

Use regression when the output is a number.

Examples:

• Price
• Revenue
• Temperature
• Delivery time
• Number of units sold

Question to ask:

Is the model predicting a continuous numeric value?

Clustering

Use clustering when the system groups items without predefined labels.

Examples:

• Group customers by similar behavior
• Find natural segments in purchase history
• Organize documents by similarity without prior categories

Question to ask:

Are there no labels, and is the goal to discover groups?

Anomaly detection

Use anomaly detection when the goal is to find unusual patterns.

Examples:

• Unusual credit card transactions
• Abnormal machine sensor readings
• Unexpected network traffic
• Outlier behavior in time-series data

Question to ask:

Is the goal to identify what does not look normal?

Understand training, validation, and inferencing language

AI-900 scenarios may use model lifecycle terms. Keep them distinct.

• Training: The model learns patterns from historical data.
• Validation or evaluation: The model is tested to estimate performance.
• Inferencing: The trained model is used to make predictions on new data.
• Features: Input variables used by the model.
• Labels: Known answers used during supervised training.
• Dataset: The collection of data used for training, testing, or analysis.

Example:

A model has already been trained to predict delivery delays. A new order is submitted, and the system predicts whether it will be late.

Decision:

• The model is being used on new data.
• This is inferencing, not training.

Interpret generative AI scenarios carefully

Generative AI scenarios often describe systems that create or transform content.

Look for requirements such as:

• Generate a response to a user prompt
• Summarize a long document
• Draft an email
• Create natural language answers
• Build a chat experience
• Use prompts and completions
• Retrieve relevant source content before generating an answer, if the scenario includes search or grounding

Decision habit:

1. Identify whether the goal is to create content, extract facts, classify data, or search content.
2. If the goal is content generation or conversational response, think generative AI.
3. If the goal is exact field extraction from forms, think document intelligence.
4. If the goal is sentiment, entities, or language detection, think language analytics.
5. If the goal is retrieving documents, think search.

Do not label every text scenario as generative AI. Sentiment analysis, entity recognition, translation, and document extraction are different tasks.

Use answer choices as a second-pass filter

After you understand the scenario, evaluate each answer choice against the facts.

For each option, ask:

• Does this option process the input type described?
• Does it produce the required output?
• Does it satisfy the stated constraint?
• Is it more complex than necessary?
• Is it solving a different problem?
• Is it a concept answer when the question asks for a service, or a service answer when the question asks for a concept?

A good AI-900 answer should align with the scenario at all three levels:

• Workload: vision, speech, language, document, machine learning, search, or generative AI
• Task: classify, extract, translate, transcribe, predict, group, detect, generate, or search
• Constraint: prebuilt, custom, responsible, secure, explainable, or accessible

Short scenario walkthroughs

Scenario 1: Customer feedback

A company collects thousands of customer comments. It wants to identify whether each comment expresses a positive, neutral, or negative opinion.

Reasoning:

• Input: written comments
• Goal: determine opinion
• Output: sentiment category
• Best match: natural language processing, sentiment analysis, Azure AI Language

Defensible answer:

• Use a language/text analytics capability for sentiment analysis.

Scenario 2: Invoice extraction

An accounting team scans supplier invoices and wants to extract invoice numbers, dates, vendor names, and totals into a database.

Reasoning:

• Input: invoices
• Goal: extract structured fields
• Output: key document fields
• Best match: Azure AI Document Intelligence

Defensible answer:

• Use document intelligence rather than general image classification or text sentiment analysis.

Scenario 3: Product demand

A retailer wants to predict the number of units of a product that will be sold next week based on historical sales data.

Reasoning:

• Input: historical sales records
• Goal: predict a number
• Output: numeric quantity
• Best match: regression or forecasting-style prediction, depending on options

Defensible answer:

• Choose regression if the answer choices are machine learning task types.

Scenario 4: Accessibility

A company designs an AI application and wants it to be usable by people with visual, hearing, and motor impairments.

Reasoning:

• Goal: support diverse user needs
• Topic: responsible AI
• Best match: inclusiveness

Defensible answer:

• Choose inclusiveness, not privacy or transparency.

Scenario 5: Spoken customer calls

A support center records customer calls and wants to convert the spoken words into text for later analysis.

Reasoning:

• Input: audio
• Goal: speech-to-text
• Output: transcript
• Best match: Azure AI Speech

Defensible answer:

• Use a speech service for transcription. Text analytics may be used later, but the first required step is speech recognition.

Build a repeatable decision sequence

Use this sequence during practice until it becomes automatic.

1. Read the final question first

   • Identify whether you need a service, workload, ML task, or responsible AI principle.

2. Find the input

   • Text, image, audio, document, tabular data, or user prompt.

3. Find the required output

   • Label, number, extracted fields, translated text, transcript, generated response, search result, or explanation.

4. Identify the action verb

   • Predict, classify, group, detect, translate, transcribe, extract, summarize, generate, search, secure, explain, or govern.

5. Apply constraints

   • Prebuilt versus custom, privacy, accessibility, explainability, human oversight, or safety.

6. Map to the simplest suitable concept

   • Choose the service or AI workload that directly matches the input and output.

7. Eliminate answers solving a different task

   • Do not choose a service just because it is generally related to AI.

8. Select the most defensible answer

   • Prefer the option that uses the scenario facts with the fewest assumptions.

Final-review checklist for AI-900 scenarios

Before exam day, make sure you can quickly answer these questions:

• Can I distinguish classification, regression, clustering, and anomaly detection from the output?
• Can I identify when a scenario is about text, vision, speech, translation, document extraction, search, or generative AI?
• Can I match common business goals to Azure AI service families?
• Can I tell when a scenario is asking for a responsible AI principle rather than a technical service?
• Can I separate required facts from background context?
• Can I choose a prebuilt AI capability when custom machine learning is unnecessary?
• Can I explain why the correct answer is better than the other plausible options?

Practice method for the final week

For each scenario practice question, do not jump straight to the answer. Write or say one sentence in this format:

The input is ___, the required output is ___, so the best match is ___.

Examples:

• The input is customer text, the required output is sentiment, so the best match is Azure AI Language sentiment analysis.
• The input is scanned invoices, the required output is structured fields, so the best match is Azure AI Document Intelligence.
• The input is historical tabular data, the required output is a numeric prediction, so the best match is regression.
• The scenario is about biased outcomes across groups, so the best match is fairness.

Then review the explanations for both correct and incorrect choices. Your goal is not only to know the answer, but to recognize the decision pattern quickly.

Next, combine targeted topic drills with full-length mock exams. Use topic drills to strengthen weak areas such as responsible AI, machine learning task types, or Azure AI service selection, then use mock exams to practice reading scenarios under timed conditions.