AI-900 — Microsoft Azure AI Fundamentals Exam Blueprint
Last revised: June 18, 2026
Practical AI-900 exam blueprint for Microsoft Azure AI Fundamentals candidates reviewing Azure AI concepts, services, workloads, responsible AI, and scenario readiness.
How to Use This Exam Blueprint
This independent Exam Blueprint is for candidates preparing for Microsoft Azure AI Fundamentals (AI-900). Use it as a practical readiness map: identify what you can explain, what you can recognize in scenarios, and where you still confuse Azure AI services, machine learning concepts, responsible AI principles, or generative AI capabilities.
AI-900 is a fundamentals exam, so readiness is not about memorizing every Azure setting. You should be able to:
Recognize common AI workload types.
Match business problems to appropriate Azure AI services.
Understand core machine learning concepts.
Identify computer vision, natural language processing, knowledge mining, and generative AI use cases.
Apply responsible AI thinking to realistic scenarios.
Interpret basic model evaluation and deployment ideas without overengineering.
Distinguish document extraction from search and knowledge mining
Generative AI
Prompts, completions, embeddings, grounding, copilots, large language models, responsible generation
Explain how generative AI differs from predictive AI and how to reduce risk
Azure AI service selection
Azure AI Services, Azure AI Search, Azure OpenAI Service, Azure Machine Learning, Azure Bot Service
Choose the most fitting Azure option from scenario requirements
Security and governance basics
Keys, endpoints, managed identity concepts, access control, data privacy, monitoring
Recognize secure configuration and operational concerns at a fundamentals level
AI Workload Fundamentals Checklist
Review the basic AI workload categories before going deep into services.
Workload type
Scenario clue
Readiness check
Classification
“Assign each item to a known category”
Can you identify binary vs multiclass classification?
Regression
“Predict a numeric value”
Can you distinguish price, demand, or temperature prediction from classification?
Clustering
“Group similar items without predefined labels”
Can you recognize unsupervised grouping scenarios?
Anomaly detection
“Find unusual activity or outliers”
Can you identify fraud, equipment failure, or abnormal telemetry scenarios?
Computer vision
“Analyze images or video”
Can you tell image classification, object detection, and OCR apart?
Natural language processing
“Analyze, understand, or generate language”
Can you map sentiment, translation, entities, and summarization to NLP?
Speech
“Convert speech to text or text to speech”
Can you identify transcription, synthesis, and translation use cases?
Generative AI
“Create text, images, code, summaries, or answers”
Can you explain prompts, responses, grounding, and content risk?
Knowledge mining/search
“Extract insights from large document collections”
Can you distinguish search indexing from model training?
Can You Do This?
Explain the difference between artificial intelligence, machine learning, and deep learning.
Identify whether a scenario is classification, regression, clustering, or anomaly detection.
Recognize when a prebuilt AI service is more appropriate than building a custom model.
Explain why data quality affects model quality.
Distinguish supervised, unsupervised, and generative AI at a high level.
Recognize AI use cases that require human review or governance.
Responsible AI Readiness
AI-900 commonly expects candidates to reason about responsible AI in plain language. Do not just memorize principle names; practice applying them to scenarios.
Principle area
What it means in exam scenarios
Example decision cue
Fairness
Avoiding unjust bias or unequal outcomes
“Does the model perform differently for different groups?”
Reliability and safety
The system should behave consistently and safely
“What happens if the model is wrong?”
Privacy and security
Protect data and access to AI systems
“Is sensitive data being collected, stored, or exposed?”
Inclusiveness
The solution should work for diverse users
“Can people with different abilities or contexts use it?”
Transparency
Users and stakeholders should understand system behavior
“Can users know when AI is being used?”
Accountability
Humans and organizations remain responsible
“Who monitors, approves, and responds to failures?”
Responsible AI Scenario Checks
Scenario
Strong answer pattern
A hiring model ranks applicants using historical hiring data
Check for bias, fairness, explainability, and human oversight
A chatbot answers medical or legal questions
Add grounding, disclaimers, safety controls, escalation, and content review
A facial analysis solution is proposed for a public setting
Consider privacy, consent, security, accuracy, and appropriate use
A model performs well overall but poorly for one demographic group
Investigate fairness and dataset representation before deployment
A generative AI app exposes confidential customer data in responses
Review data handling, grounding sources, access control, and prompt protections
A prediction model is used without explaining limitations to users
Improve transparency, documentation, and accountability
Can You Do This?
Identify bias risks in training data.
Explain why transparency matters when AI influences decisions.
Recognize privacy risks in speech, image, document, and generative AI systems.
Choose human review when AI output has high impact.
Explain why monitoring is still required after an AI solution is deployed.
Recognize that responsible AI applies to both custom models and prebuilt AI services.
Machine Learning Concepts Checklist
You do not need to be a data scientist for AI-900, but you should know the vocabulary and lifecycle.
Concept
What to know
Ready when you can…
Dataset
Collection of examples used for analysis or training
Explain why representative data matters
Feature
Input variable used by a model
Identify features in a sample scenario
Label
Known output value used in supervised learning
Identify the target value being predicted
Training
Process of learning patterns from data
Explain that training creates a model
Validation/testing
Data used to evaluate model performance
Explain why evaluation should use data not used for training
Algorithm
Method used to learn patterns
Recognize that different algorithms fit different tasks
Model
Trained artifact used to make predictions
Explain that a model is the result of training
Inference
Using a trained model to generate predictions
Recognize real-time vs batch prediction concepts
Overfitting
Model performs well on training data but poorly on new data
Identify when a model memorizes instead of generalizes
Evaluation metric
Measure of model performance
Match common metrics to broad task types
Common ML Task Types
Task
Input
Output
Example
Binary classification
Features
One of two classes
Approve or deny a loan
Multiclass classification
Features
One of several classes
Classify support ticket category
Regression
Features
Numeric value
Predict monthly sales
Clustering
Unlabeled data
Groups of similar items
Segment customers
Anomaly detection
Normal behavior data or telemetry
Unusual events
Detect abnormal network traffic
Model Evaluation Readiness
Metric or idea
Best associated with
What to understand
Accuracy
Classification
Overall proportion of correct predictions
Precision
Classification
Of predicted positives, how many were correct
Recall
Classification
Of actual positives, how many were found
Confusion matrix
Classification
Counts of true/false positives and negatives
Mean absolute error
Regression
Average size of numeric prediction errors
Root mean squared error
Regression
Error measure that penalizes larger errors more
R-squared
Regression
How much variance the model explains
Silhouette score concept
Clustering
How well-separated clusters are, at a high level
Can You Do This?
Identify features and labels from a business example.
Explain training versus inference.
Distinguish classification from regression.
Explain why splitting data helps evaluate a model.
Recognize overfitting in a scenario.
Interpret a confusion-matrix-style description at a basic level.
Know when clustering is used instead of classification.
Azure Machine Learning Readiness
For AI-900, focus on what Azure Machine Learning is used for and how it differs from prebuilt Azure AI services.
Area
What to review
Scenario cue
Custom model development
Building, training, evaluating, and deploying ML models
“The organization has historical labeled data and needs a custom predictor”
Automated machine learning
Selecting algorithms and tuning models with automation
“Users want help finding a good model without manually coding every algorithm”
Designer-style workflows
Visual pipeline creation for ML tasks
“Low-code drag-and-drop model training workflow”
Model management
Registering, versioning, and deploying models
“Track models and use them for predictions”
Endpoints
Exposing a model for inference
“Applications need to call the model for predictions”
Responsible ML
Explainability, fairness, data handling, monitoring
“Understand why the model made predictions or whether it is biased”
Azure Machine Learning vs Azure AI Services
Choose this
When the scenario says…
Azure Machine Learning
Build a custom model from your own data, control training, compare algorithms, deploy custom inference
Azure AI Services
Use prebuilt capabilities such as vision, speech, language, translation, or document extraction
Azure OpenAI Service
Use large language models for generation, summarization, chat, embeddings, or copilots
Azure AI Search
Index and search content, often with AI enrichment or retrieval for generative AI apps
Can You Do This?
Recognize when a custom ML model is needed.
Explain what automated machine learning helps with.
Understand that model deployment enables applications to use the model.
Identify training data, validation data, and test data in a lifecycle.
Explain why model monitoring is important after deployment.
Avoid choosing Azure Machine Learning when a simple prebuilt AI capability is enough.
Computer Vision Checklist
Computer vision questions often describe images, scanned documents, videos, or visual detection requirements.
Capability
What it does
Example scenario
Image classification
Assigns a category to an image
Classify product images by type
Object detection
Finds and locates objects in an image
Detect vehicles, tools, or safety equipment
Image analysis
Describes visual features or content
Generate tags or captions for images
Optical character recognition
Extracts printed or handwritten text from images
Read text from signs, receipts, or scanned pages
Face-related detection concepts
Detects or analyzes faces depending on service capability and appropriate use
Identify whether images contain faces for workflow routing
Spatial/video analysis concepts
Understand movement or visual events
Monitor a physical space or video stream
Vision Service Selection
Requirement
Likely Azure direction
Extract text from images
Azure AI Vision OCR capability or document-focused service depending on document complexity
Analyze general image content
Azure AI Vision
Extract structured fields from invoices, receipts, forms, or documents
Azure AI Document Intelligence
Search images or documents by extracted content
Azure AI Search with enrichment concepts
Build a custom image classification or detection model
Custom vision or custom ML approach, depending on the requirement
Can You Do This?
Distinguish image classification from object detection.
Recognize OCR scenarios.
Identify when a document extraction service is more appropriate than general image analysis.
Explain why face-related scenarios require privacy and responsible AI review.
Match visual business problems to Azure AI Vision or Azure AI Document Intelligence.
Recognize when custom training is needed because prebuilt categories are not enough.
Natural Language Processing Checklist
Natural language processing includes text analysis, language understanding, translation, and speech-related workloads.
Capability
What it does
Example scenario
Sentiment analysis
Determines positive, negative, neutral, or mixed tone
Analyze customer feedback
Key phrase extraction
Identifies important phrases
Summarize survey themes
Named entity recognition
Finds entities such as people, places, organizations, dates, or quantities
Extract entities from support emails
Language detection
Identifies the language of text
Route multilingual content
Translation
Converts text from one language to another
Translate product documentation
Question answering
Provides answers from a knowledge base or content source
FAQ bot for internal policies
Conversational language understanding
Interprets user intents and entities
Route chatbot requests
Speech to text
Converts audio into text
Transcribe calls or meetings
Text to speech
Converts text into spoken audio
Voice response system
Speech translation
Translates spoken language
Multilingual live communication
NLP Service Selection
Requirement
Likely Azure direction
Analyze sentiment, key phrases, entities, or language
Azure AI Language
Translate text between languages
Azure AI Translator
Convert speech to text or text to speech
Azure AI Speech
Build a conversational bot experience
Azure Bot Service with AI language capabilities
Answer questions from enterprise documents
Question answering, Azure AI Search, or generative AI with retrieval depending on complexity
Summarize or generate text
Azure OpenAI Service or other generative AI capability, with responsible AI controls
Can You Do This?
Identify sentiment analysis versus key phrase extraction.
Recognize entity extraction scenarios.
Distinguish translation from transcription.
Explain speech to text versus text to speech.
Identify intent and entity concepts in conversational systems.
Choose between a rules/knowledge-base answer and a generative answer pattern.
Recognize multilingual requirements.
Document Intelligence, Search, and Knowledge Mining
Document and search scenarios often combine OCR, enrichment, indexing, and retrieval.
Topic
What to know
Scenario cue
Document extraction
Pulls structured fields from forms and documents
“Extract invoice number, date, vendor, and total”
OCR
Extracts text from images or scanned files
“Read printed or handwritten text”
Prebuilt document models
Target common document types
“Receipts, invoices, IDs, forms”
Custom document extraction
Trained for organization-specific layouts or fields
“Extract fields from a proprietary form”
Search indexing
Makes content searchable
“Users need to search across PDFs and documents”
AI enrichment
Adds extracted text, entities, key phrases, or other metadata
“Improve search with AI-generated fields”
Retrieval for generative AI
Finds relevant source content for generated responses
“Ground chatbot answers in company documents”
Can You Do This?
Distinguish extracting fields from searching documents.
Recognize when OCR alone is not enough.
Identify when a prebuilt document model might fit.
Identify when custom extraction is needed.
Explain how AI enrichment can improve search.
Understand why search grounding can reduce unsupported generative AI answers.
Generative AI Checklist
Generative AI is a major readiness area because scenarios test both capability and judgment.
Concept
What to know
Readiness cue
Large language model
Model that can understand and generate language-like outputs
Can explain chat, completion, summarization, and reasoning-style tasks
Prompt
User or system instruction provided to the model
Can improve vague prompts with context and constraints
Completion/response
Model output
Can identify that output may require validation
System message/instruction
Higher-priority guidance for model behavior
Can explain why apps define behavior and tone
Grounding
Providing trusted source data to guide responses
Can explain why grounding reduces unsupported answers
Retrieval-augmented generation
Retrieve relevant documents, then generate an answer using them
Can identify enterprise document chatbot scenarios
Embeddings
Numeric representation of meaning for similarity search
Can recognize semantic search and retrieval use cases
Token concept
Units of text processed by a model
Can reason that longer prompts/responses consume more capacity without needing exact limits
Content filtering
Detects or reduces harmful content
Can identify safety controls for generated output
Copilot pattern
AI assistant embedded in a workflow
Can describe productivity assistance without assuming full automation
Generative AI Scenario Decision Points
Scenario
What to choose or watch
“Summarize long customer emails and suggest replies”
Generative AI with review, privacy controls, and tone guidance
“Answer employee questions using HR policies”
Retrieval-grounded generative AI or question answering over trusted content
“Generate SQL queries from natural language”
Generative AI with validation, least privilege, and human approval for risky actions
“Create marketing copy in many languages”
Generative AI plus translation/language review and brand/safety checks
“Use AI to make final legal decisions”
High-risk; require accountability, human oversight, and domain review
“Chatbot gives confident answers not found in source documents”
Improve grounding, retrieval, prompt instructions, and evaluation
Can You Do This?
Explain generative AI versus traditional predictive ML.
Identify prompts, responses, grounding data, and system instructions.
Recognize when Azure OpenAI Service is relevant.
Explain why generated content may be incorrect or unsupported.
Identify where content filtering and human review are needed.
Recognize retrieval-augmented generation from a scenario.
Explain embeddings at a high level as useful for similarity search.
Avoid treating generative AI output as automatically authoritative.
Azure AI Service Selection Matrix
Use this table for rapid final review. Many fundamentals questions are service-selection questions.
If the requirement is…
Review this Azure area
Build a custom predictive model from historical data
Azure Machine Learning
Use automated model training and comparison
Azure Machine Learning automated ML concepts
Analyze images, tags, captions, or visual content
Azure AI Vision
Extract text from images
Azure AI Vision OCR or document-focused services
Extract structured fields from forms, invoices, receipts, or documents
Azure AI Document Intelligence
Analyze sentiment, key phrases, language, or entities
Azure AI Language
Translate text
Azure AI Translator
Convert speech to text or text to speech
Azure AI Speech
Build a chat or conversational interface
Azure Bot Service plus language/generative AI capabilities
Search enterprise content
Azure AI Search
Enrich searchable content with AI
Azure AI Search enrichment concepts
Generate, summarize, chat, or create natural-language answers
Azure OpenAI Service
Ground a chatbot in enterprise documents
Azure AI Search plus generative AI pattern
Detect unusual behavior or outliers
Anomaly detection or ML-based approach
Analyze fairness, explainability, or responsible deployment
Responsible AI and Azure ML responsible ML concepts
Security, Privacy, and Operations Basics
AI-900 is not a deep security administration exam, but you should recognize secure and responsible implementation patterns.
Area
What to know at a fundamentals level
Authentication and access
AI services require controlled access; avoid exposing secrets
Keys and endpoints
Applications typically call service endpoints using authorized credentials
Managed identity concept
Helps avoid hardcoded secrets where supported
Role-based access control concept
Assign only needed access to users and services
Data privacy
Know whether input data contains personal, confidential, or regulated information
Network/security posture
Sensitive AI workloads may require stricter access paths and monitoring
Logging and monitoring
AI services and apps should be monitored for failures, usage, and quality issues
Cost awareness
Model choice, request volume, indexing, training, and generation can affect cost
Human review
Important for high-impact or uncertain AI outputs
Lifecycle management
AI solutions should be tested, deployed, monitored, and improved
Can You Do This?
Recognize why hardcoding keys in application code is risky.
Explain why confidential data should be handled carefully in prompts and documents.
Identify when least privilege should be applied.
Explain why monitoring AI outputs matters, not just uptime.
Recognize that cost and latency can influence service and architecture choices.
Choose human approval for high-impact AI-generated recommendations.
Scenario and Decision-Point Practice
Use these prompts to test whether you can reason through AI-900-style decisions.
Scenario prompt
Ask yourself
Likely readiness outcome
A retailer wants to predict next month’s revenue
Is the output numeric?
Regression / custom ML concept
A bank wants to approve or reject applications
Is the output one of two categories? Are fairness concerns present?
Binary classification plus responsible AI
A manufacturer wants to detect unusual machine telemetry
Is the goal to find outliers?
Anomaly detection
A support team wants to route tickets into categories
Are categories predefined?
Multiclass classification or language classification
A website wants to translate articles into multiple languages
Is text translation the core need?
Azure AI Translator
A call center wants searchable transcripts
Is speech converted to text first?
Azure AI Speech plus search
An insurance company wants to extract values from claim forms
Are structured fields needed from documents?
Azure AI Document Intelligence
A legal team wants a chatbot answering from internal policies
Must answers be grounded in trusted content?
Search/retrieval plus generative AI or question answering
A media library wants to find images containing specific objects
Is visual object identification needed?
Azure AI Vision
A hospital wants AI to make final treatment decisions automatically
Is this high impact and safety-critical?
Responsible AI, human oversight, caution
Common Weak Areas and Traps
Trap
Why it causes mistakes
What to do instead
Confusing classification and regression
Both are predictive ML, but outputs differ
Ask: category or number?
Treating clustering as classification
Clustering has no predefined labels
Ask whether labels exist before training
Choosing custom ML for every problem
Prebuilt services may already solve vision, speech, language, or documents
Match simple AI tasks to Azure AI Services
Choosing OCR for full document understanding
OCR reads text; it may not extract structured business fields
Use document intelligence concepts for forms and fields
Assuming generative AI is always correct
LLMs can produce unsupported or inaccurate output
Use grounding, evaluation, and human review
Ignoring responsible AI in technical scenarios
Ethics, privacy, fairness, and accountability are testable concepts
Add responsible AI checks to every AI system
Mixing translation and transcription
Translation changes language; transcription changes speech to text
Identify the input and output format
Confusing entities and key phrases
Entities are specific recognized items; key phrases are important phrases
Look for people, places, organizations, dates, or quantities
Assuming high accuracy means fair performance
Aggregate metrics can hide group-level issues
Check performance across relevant populations
Forgetting monitoring after deployment
AI models and inputs can change over time
Include monitoring, retraining, and review concepts
Overlooking data sensitivity in prompts
Prompts and documents may contain confidential information
Apply privacy and access controls
Treating search and generation as the same thing
Search retrieves; generative AI creates responses
Understand retrieval-grounded generation patterns
Quick “Can You Do This?” Master Checklist
Core AI Concepts
Define AI, machine learning, and deep learning at a high level.
Identify supervised and unsupervised learning scenarios.
Identify classification, regression, clustering, and anomaly detection.
Explain training, testing, validation, model, feature, label, and inference.
Recognize overfitting and poor generalization.
Match common metrics to classification or regression.
Azure AI Workloads
Choose Azure Machine Learning for custom model training scenarios.
Choose Azure AI Vision for image analysis scenarios.
Choose OCR or document intelligence appropriately.
Choose Azure AI Language for text analytics scenarios.
Choose Azure AI Translator for translation scenarios.
Choose Azure AI Speech for speech-to-text and text-to-speech scenarios.
Choose Azure AI Search for indexing and retrieval scenarios.
Choose Azure OpenAI Service for generative AI scenarios.
Recognize chatbot and conversational AI patterns.
Responsible AI and Governance
Apply fairness, reliability and safety, privacy and security, inclusiveness, transparency, and accountability.
Identify bias and representation problems.
Recognize high-impact scenarios requiring human oversight.
Explain why explainability and transparency matter.
Identify privacy risks in documents, speech, images, and prompts.
Recognize monitoring and review needs after deployment.
Generative AI
Explain prompts, completions, grounding, embeddings, and retrieval-augmented generation.
Identify when generative AI should use trusted source content.
Recognize harmful content, hallucination, and data leakage risks.
Explain why content filters, evaluation, and human review may be needed.
Distinguish generative AI from search, translation, and traditional ML.
Final-Week Review Checklist
Timeframe
Focus
Checklist
5-7 days before
Rebuild the topic map
Review each major workload area and write one example scenario for each
4-5 days before
Service selection
Drill “which Azure service fits this requirement?” questions
3-4 days before
ML concepts
Recheck classification, regression, clustering, features, labels, metrics, and overfitting
2-3 days before
Responsible AI
Practice applying principles to hiring, healthcare, finance, surveillance, and chatbot scenarios
1-2 days before
Generative AI
Review prompts, grounding, embeddings, retrieval, content safety, and human review
Final day
Light review
Revisit weak tables, avoid cramming product minutiae, and practice scenario reasoning
Final Readiness Questions
Before exam day, you should be able to answer these without notes:
What makes a problem classification instead of regression?
When would you use Azure Machine Learning instead of a prebuilt Azure AI service?
What is the difference between OCR and document field extraction?
What service area handles sentiment and entity recognition?
What service area handles speech-to-text?
What service area handles text translation?
What is grounding in a generative AI solution?
Why can a generative AI answer be fluent but wrong?
What responsible AI risks appear in a hiring, lending, healthcare, or surveillance scenario?
Why should AI systems be monitored after deployment?
Practical Next Step
Turn this checklist into active practice: for each unchecked item, create a short scenario, choose the Azure AI workload or concept that fits, and explain the responsible AI concern in one sentence. Then use original AI-900 practice questions to verify that you can apply the checklist under exam-style wording, not just recognize definitions.