AI-900 — Microsoft Azure AI Fundamentals Quick Reference

Last revised: June 18, 2026

Compact AI-900 reference for Microsoft Azure AI Fundamentals: AI workloads, Azure AI services, machine learning, vision, language, and generative AI.

How to Use This Quick Reference

This independent Quick Reference is for candidates preparing for the real Microsoft Azure AI Fundamentals (AI-900) exam. Focus on recognizing workloads, choosing the right Azure AI service, and understanding core AI concepts at a fundamentals level.

AI-900 is not primarily a coding exam. Expect scenario-style questions such as:

Which Azure service fits a workload?
Is the problem classification, regression, clustering, NLP, vision, or generative AI?
Which responsible AI principle is involved?
Which metric or ML concept best matches the scenario?
How do Azure AI services, Azure Machine Learning, Azure AI Search, and Azure OpenAI Service differ?

High-Yield Azure AI Service Selection

Scenario	Best-fit Azure capability	Key exam cues	Common trap
Predict a category, value, or cluster from data	Azure Machine Learning	Train, evaluate, deploy ML models; AutoML; designer; notebooks	Do not choose Azure AI Vision or Language unless the input is image/text-specific
Build an ML model without writing much code	Azure Machine Learning automated ML / designer	Low-code model training, pipelines, drag-and-drop workflow	AutoML chooses models; it is not the same as generative AI
Analyze images	Azure AI Vision	Tags, captions, object detection, OCR, image analysis	OCR extracts text; it does not understand document structure by itself
Train custom image classification or object detection	Custom vision capabilities	Your own labeled images; custom tags/classes	General image analysis uses prebuilt models
Extract fields from invoices, receipts, IDs, forms	Azure AI Document Intelligence	Key-value pairs, tables, structured document extraction	Do not use simple OCR if the question asks for structured fields
Search large document collections with AI enrichment	Azure AI Search	Indexes, indexers, skillsets, knowledge mining, semantic search	Search is retrieval; it is not primarily model training
Detect sentiment, entities, key phrases, language	Azure AI Language	Text analytics, NER, sentiment, PII, summarization	Translator is specifically for translation
Build intent-based conversational understanding	Azure AI Language - Conversational Language Understanding	Intents, utterances, entities	A chatbot channel is not the same as language understanding
Build FAQ-style answers from a knowledge base	Azure AI Language - question answering	Questions and answers, knowledge base, FAQ	Not the same as open-ended generative chat
Translate text between languages	Azure AI Translator	Text translation, language pairs	Speech translation uses speech services
Convert speech to text or text to speech	Azure AI Speech	Transcription, synthesis, speech translation	Speech-to-text is not OCR
Generate text, summarize, chat, reason over prompts	Azure OpenAI Service / Azure AI Foundry	Large language models, prompts, completions, chat, embeddings	Generative AI can hallucinate; grounding may be needed
Detect harmful AI content	Azure AI Content Safety	Hate, sexual, violence, self-harm, harmful content moderation	Content filtering is not the same as model accuracy
Build a bot that connects to channels	Azure Bot Service	Web chat, Teams, channels, bot conversations	Bot Service hosts orchestration; language models handle understanding

AI Workload Types

Workload	What it does	Example exam scenario
Machine learning	Learns patterns from data to make predictions	Predict customer churn or house prices
Computer vision	Interprets images and video	Detect products in shelf images
Natural language processing	Understands or generates human language	Extract entities from support tickets
Document intelligence	Extracts structured data from documents	Pull invoice totals and vendor names
Knowledge mining	Extracts searchable insights from content	Enrich PDFs and make them searchable
Generative AI	Creates new content from prompts	Draft answers, summarize reports, generate chat responses
Conversational AI	Enables user interactions through natural language	Customer support chatbot
Anomaly detection	Finds unusual patterns	Identify abnormal sensor readings
Speech AI	Processes spoken language	Transcribe meeting audio

Responsible AI Principles

Microsoft emphasizes responsible AI concepts throughout Azure AI workloads. Know the principle and the scenario cue.

Principle	Meaning	Exam cue	Practical controls
Fairness	AI systems should avoid unfair bias or discrimination	Model performs worse for one demographic group	Representative data, bias testing, human review
Reliability and safety	AI systems should work reliably and safely under expected conditions	Incorrect prediction could cause harm	Testing, monitoring, fallback paths, safe deployment
Privacy and security	Protect data and systems	Personal data in prompts, training data, or logs	Access control, encryption, data minimization
Inclusiveness	AI should work for people with diverse abilities and needs	Application excludes users with disabilities or language needs	Accessibility, multilingual support, inclusive design
Transparency	Users and stakeholders should understand AI behavior and limitations	Need to explain why or how AI is used	Explainability, disclosures, model documentation
Accountability	People and organizations remain responsible for AI outcomes	Who approves model use or handles errors?	Governance, auditability, human oversight

Responsible AI Traps

If the question says…	Think…
“The model works well overall but poorly for one group”	Fairness
“Users should know they are interacting with AI”	Transparency
“Sensitive customer data is used in prompts”	Privacy and security
“A human must approve high-impact recommendations”	Accountability
“The system must avoid dangerous behavior under edge cases”	Reliability and safety
“The app should support people with different abilities”	Inclusiveness

Core AI and ML Terms

Term	Exam-ready meaning
Artificial intelligence	Broad field of systems that perform tasks associated with human intelligence
Machine learning	AI technique where models learn patterns from data
Deep learning	ML using neural networks with many layers; often used for vision, speech, and language
Model	Trained artifact that maps inputs to predictions or outputs
Feature	Input variable used by a model
Label	Known answer used during supervised training
Training data	Data used to fit the model
Validation data	Data used during model selection/tuning
Test data	Held-out data used to estimate final model performance
Inference	Using a trained model to make predictions
Algorithm	Method used to train a model
Overfitting	Model memorizes training data and performs poorly on new data
Underfitting	Model is too simple and performs poorly even on training data
Bias in data	Skewed or unrepresentative data that may lead to unfair or inaccurate outcomes
Feature engineering	Selecting or transforming inputs to improve model performance
Hyperparameters	Training settings chosen before or during training, not learned directly from data

Machine Learning Task Decision Table

Task	Prediction/output	Supervision type	Example
Binary classification	One of two classes	Supervised	Fraud or not fraud
Multiclass classification	One of more than two classes	Supervised	Classify support ticket category
Multilabel classification	Multiple labels can apply	Supervised	Tag an image as outdoor, vehicle, daytime
Regression	Numeric value	Supervised	Predict price, revenue, temperature
Clustering	Groups with similar characteristics	Unsupervised	Segment customers into groups
Anomaly detection	Normal vs unusual pattern	Usually unsupervised or semi-supervised	Detect unusual machine telemetry
Forecasting	Future numeric values over time	Supervised time-series	Predict next month’s demand
Ranking/recommendation	Ordered results or suggested items	Often supervised or hybrid	Recommend products

Quick Distinctions

Distinction	Remember
Classification vs regression	Classification predicts categories; regression predicts numbers
Binary vs multiclass	Binary has two classes; multiclass has three or more
Clustering vs classification	Clustering has no known labels during training
Training vs inference	Training creates the model; inference uses it
Validation vs test data	Validation helps tune; test estimates final performance
Overfitting vs underfitting	Overfitting is too tailored to training data; underfitting is too simple

Model Evaluation Metrics

Classification Confusion Matrix Terms

Term	Meaning
True positive	Predicted positive and actually positive
True negative	Predicted negative and actually negative
False positive	Predicted positive but actually negative
False negative	Predicted negative but actually positive

\[ \text{Accuracy} = \frac{TP + TN}{TP + TN + FP + FN} \]\[ \text{Precision} = \frac{TP}{TP + FP} \]\[ \text{Recall} = \frac{TP}{TP + FN} \]\[ \text{F1 Score} = 2 \times \frac{\text{Precision} \times \text{Recall}}{\text{Precision} + \text{Recall}} \]

Metric	Use when…	Watch for
Accuracy	Classes are balanced and all errors have similar cost	Misleading with imbalanced data
Precision	False positives are costly	“When the model says yes, how often is it right?”
Recall	False negatives are costly	“How many real positives did the model find?”
F1 score	Need balance between precision and recall	Useful for imbalanced classification
ROC/AUC	Need overall separation ability across thresholds	Higher generally means better separation
MAE	Regression error in original units	Easier to interpret
RMSE	Penalizes large regression errors more	Sensitive to outliers
R-squared	Proportion of variance explained	Can be misleading if used alone

Metric Decision Cues

Scenario	Prefer
Cancer screening: missing a positive case is very bad	High recall
Spam filter: wrongly blocking valid email is very bad	High precision
Fraud detection with rare fraud cases	Precision, recall, F1; not accuracy alone
Predicting sales amount	Regression metrics such as MAE/RMSE
Comparing classification thresholds	ROC/AUC, precision-recall tradeoff

Azure Machine Learning Quick Reference

Concept	What it is	Exam cue
Workspace	Top-level Azure Machine Learning resource	Organizes experiments, jobs, models, compute, data
Data asset	Registered dataset or data reference	Reusable training data
Datastore	Connection to storage	Where data is stored
Compute instance	Development workstation in the cloud	Notebooks, interactive development
Compute cluster	Scalable training compute	Runs jobs at scale
Experiment/job	Training or evaluation run	Track metrics and outputs
Environment	Runtime dependencies	Python packages, Docker image, reproducibility
Model	Registered trained model	Deploy for inference
Endpoint	Hosted model access point	Real-time or batch predictions
Automated ML	Tries algorithms/preprocessing automatically	Low-code model creation
Designer	Visual drag-and-drop ML pipelines	No-code/low-code workflow
Responsible AI dashboard	Model insights and fairness/explainability tools	Evaluate model behavior

ML Lifecycle

    flowchart LR
	    A[Define problem] --> B[Prepare data]
	    B --> C[Train model]
	    C --> D[Evaluate metrics]
	    D --> E{Good enough?}
	    E -- No --> B
	    E -- Yes --> F[Deploy endpoint]
	    F --> G[Monitor performance]
	    G --> H[Retrain when needed]

Azure ML Decision Points

Need	Choose
No-code visual ML pipeline	Designer
Automatically try multiple algorithms	Automated ML
Full control with code	Notebooks / SDK / CLI
Reproducible training environment	Environment
Scale-out training jobs	Compute cluster
Interactive development	Compute instance
Track training metrics	Experiment/job history
Serve predictions to apps	Managed endpoint

Computer Vision Reference

Capability	What it does	Example
Image classification	Assigns one or more labels to an image	“This image contains a dog”
Object detection	Locates objects with bounding boxes	Find vehicles in traffic images
Image tagging	Adds descriptive tags	Outdoor, building, person
Image captioning	Generates a natural language description	“A person riding a bicycle”
OCR / Read	Extracts printed or handwritten text	Read text from a sign or scanned page
Face detection	Detects faces and attributes depending on configuration	Locate faces in an image
Spatial analysis	Understands people movement/presence in spaces	Occupancy or distancing scenarios
Custom vision	Trains a custom classifier or detector	Identify company-specific product defects

Vision Service Selection

Requirement	Choose
General image analysis using prebuilt models	Azure AI Vision
Extract text from images	Azure AI Vision OCR / Read
Extract fields from forms and documents	Azure AI Document Intelligence
Custom image labels or object detection	Custom vision capabilities
Face-related detection or recognition scenarios	Azure AI Face capabilities
Search image-enriched documents	Azure AI Search with AI enrichment

Vision Traps

Trap	Correct idea
OCR vs Document Intelligence	OCR extracts text; Document Intelligence extracts structured fields
Classification vs object detection	Classification labels the whole image; detection locates objects
Prebuilt vs custom vision	Prebuilt works for common objects; custom uses your labeled data
Image analysis vs generative AI	Vision analyzes images; generative AI creates content or responses

Document Intelligence

Capability	What it extracts	Example
Prebuilt document models	Common structured fields	Invoices, receipts, IDs, tax forms
Custom document models	Fields specific to your forms	Internal order forms
Layout extraction	Text, tables, selection marks, structure	Convert scanned form layout to data
Key-value extraction	Named fields and values	Invoice number, total, due date
Table extraction	Rows and columns	Line items on an invoice

If the question asks for…	Choose
“Read text from an image”	OCR
“Extract invoice total, vendor, and line items”	Document Intelligence
“Search thousands of enriched PDFs”	Azure AI Search plus enrichment
“Classify custom product images”	Custom vision capabilities

Natural Language Processing Reference

Capability	Azure service/capability	What it does
Sentiment analysis	Azure AI Language	Positive, negative, neutral sentiment
Opinion mining	Azure AI Language	Sentiment about specific aspects
Key phrase extraction	Azure AI Language	Important phrases in text
Named entity recognition	Azure AI Language	People, locations, organizations, dates, quantities
PII detection	Azure AI Language	Detects personally identifiable information
Language detection	Azure AI Language	Identifies the language of text
Summarization	Azure AI Language or generative AI, depending on scenario	Condenses text
Custom text classification	Azure AI Language	Assigns custom categories
Conversational language understanding	Azure AI Language	Maps utterances to intents/entities
Question answering	Azure AI Language	Answers from a defined knowledge base
Text translation	Azure AI Translator	Translates text between languages

NLP Decision Cues

Scenario	Choose
“Is this review positive or negative?”	Sentiment analysis
“Find all company and person names”	Named entity recognition
“Detect credit card numbers or emails”	PII detection
“Find the most important words or phrases”	Key phrase extraction
“Determine whether text is English, French, or Spanish”	Language detection
“Route user request to BookFlight intent”	Conversational Language Understanding
“Answer FAQs from a knowledge base”	Question answering
“Translate a document from German to English”	Azure AI Translator
“Generate a new paragraph from a prompt”	Azure OpenAI Service

Speech and Conversational AI

Requirement	Choose	Notes
Convert spoken audio to text	Azure AI Speech - speech to text	Transcription
Convert text to natural-sounding audio	Azure AI Speech - text to speech	Voice synthesis
Translate spoken language	Azure AI Speech translation	Speech input to translated output
Translate written text	Azure AI Translator	Text-only translation
Build bot channel integration	Azure Bot Service	Connect to Teams, web chat, and other channels
Understand user intent in a bot	Conversational Language Understanding	Intents and entities
Answer predefined user questions	Question answering	FAQ/knowledge-base style

Bot Architecture at a Glance

Layer	Role
Bot application	Orchestrates conversation flow
Channel	Where users interact, such as web chat or Teams
Language understanding	Detects intent and entities
Knowledge base / data	Provides factual answers
Generative model	Creates flexible natural language responses
Human handoff	Escalates when automation is insufficient

Azure AI Search and Knowledge Mining

Azure AI Search is used to make information searchable. It can combine search indexing with AI enrichment.

Concept	Meaning
Data source	Where content comes from, such as storage or a database
Index	Searchable structure containing fields
Indexer	Crawls data and populates an index
Skillset	AI enrichment steps, such as OCR or entity extraction
Enrichment	Adds AI-generated metadata to content
Knowledge store	Stores enriched outputs for downstream use
Semantic search/ranking	Improves relevance using semantic understanding
Vector search	Retrieves content by similarity using embeddings

Search vs Other AI Services

Requirement	Best match
“Make a large document repository searchable”	Azure AI Search
“Extract entities during indexing”	Azure AI Search skillset with Azure AI Language
“Read text from scanned documents before indexing”	OCR skill in enrichment pipeline
“Extract invoice fields into structured records”	Document Intelligence
“Generate an answer grounded in search results”	Azure AI Search plus Azure OpenAI Service

Generative AI and Azure OpenAI Service

Term	Exam-ready meaning
Generative AI	AI that creates new content such as text, images, code, or summaries
Foundation model	Large pretrained model adaptable to many tasks
Large language model	Model specialized in language understanding and generation
Prompt	User or system input that guides model output
Completion	Model-generated response
Token	Unit of text processed by a model
Context	Information supplied to the model for a request
Grounding	Providing relevant source data to reduce unsupported answers
RAG	Retrieval-augmented generation; retrieve relevant data, then generate
Embedding	Numeric representation of text or content for similarity search
Fine-tuning	Further training a model on task-specific examples
Hallucination	Plausible-sounding but unsupported or incorrect output
Content filter	Control that detects or blocks harmful content
Temperature	Setting that influences randomness/creativity
System message	Instruction that defines assistant behavior or constraints

Generative AI Service Selection

Scenario	Choose
Build chat or text generation with Microsoft-hosted OpenAI models	Azure OpenAI Service
Explore, build, and manage AI apps and model deployments	Azure AI Foundry
Add enterprise document grounding to chat	Azure AI Search plus Azure OpenAI Service
Compare or select models for AI apps	Azure AI Foundry model catalog
Moderate harmful user or model content	Azure AI Content Safety
Detect sentiment or entities using prebuilt NLP	Azure AI Language
Translate text	Azure AI Translator

Retrieval-Augmented Generation Pattern

    flowchart LR
	    U[User question] --> A[App orchestration]
	    A --> S[Retrieve relevant chunks<br/>Azure AI Search]
	    S --> A
	    A --> P[Prompt with question + retrieved context]
	    P --> M[Azure OpenAI model]
	    M --> R[Answer with grounded context]

Generative AI Traps

Trap	Correct idea
“The model knows all company data automatically”	Private data usually must be supplied through grounding, retrieval, or integration
“Generative AI is always deterministic”	Outputs can vary depending on model settings and prompt
“Fine-tuning is the only way to use private data”	RAG is often used to ground responses without retraining
“A fluent answer is necessarily correct”	Generated content can be unsupported or incorrect
“Content filtering guarantees perfect safety”	Filtering helps reduce risk but does not remove the need for governance
“Embeddings generate final answers”	Embeddings support similarity search; generation uses a generative model

Azure AI Resource, Security, and Access Basics

Concept	What to know for AI-900
Azure AI services resource	Azure resource that provides access to AI APIs
Single-service resource	Resource scoped to one service, such as Speech or Language
Multi-service resource	One resource that can access multiple Azure AI services
Endpoint	URL used by applications to call the service
Key	Secret credential used to authenticate API calls
Microsoft Entra ID	Identity platform used for role-based access and managed identities
RBAC	Grants users/services permissions to Azure resources
Managed identity	Lets Azure resources authenticate without storing secrets in code
Key Vault	Secure storage for secrets, keys, and certificates
Private networking	Can restrict access paths for sensitive workloads
Monitoring	Track availability, errors, latency, and usage patterns

Security Decision Cues

Requirement	Prefer
Avoid hard-coded API keys in application code	Managed identity or secure secret storage
Store API keys securely	Azure Key Vault
Grant least-privilege access to Azure resources	RBAC
Track service errors and performance	Azure monitoring/logging tools
Protect sensitive prompts and outputs	Privacy controls, access control, data minimization
Govern generated content risks	Content filtering, human review, Responsible AI practices

Common AI-900 Scenario Cues

Question cue	Likely answer
“Predict whether a loan will default”	Binary classification
“Predict next month’s revenue”	Regression or forecasting
“Group customers by purchasing behavior”	Clustering
“Detect unusual machine behavior”	Anomaly detection
“Identify objects and locations in an image”	Object detection
“Read handwritten text from a scanned page”	OCR / Read
“Extract fields from invoices”	Azure AI Document Intelligence
“Find names and addresses in text”	Named entity recognition / PII detection
“Determine whether feedback is positive”	Sentiment analysis
“Translate support articles”	Azure AI Translator
“Transcribe call center audio”	Azure AI Speech
“Create a chatbot for Teams”	Azure Bot Service plus language/generative capabilities
“Answer from company documents with citations”	RAG with Azure AI Search and Azure OpenAI Service
“Automatically try multiple ML algorithms”	Azure Machine Learning automated ML
“Build ML pipeline visually”	Azure Machine Learning designer
“Need explanation of model behavior”	Responsible AI / model interpretability tools
“Need to reduce biased outcomes”	Fairness
“Need user disclosure that AI is used”	Transparency
“Need protect personal data”	Privacy and security

Final Review Checklist

Before practice questions, make sure you can:

Distinguish AI, ML, deep learning, and generative AI.
Identify classification, regression, clustering, anomaly detection, and forecasting scenarios.
Choose between Azure Machine Learning, Azure AI services, Azure AI Search, and Azure OpenAI Service.
Match vision tasks to image analysis, OCR, object detection, custom vision, and Document Intelligence.
Match language tasks to sentiment, NER, PII detection, translation, CLU, and question answering.
Explain the six Microsoft Responsible AI principles.
Interpret accuracy, precision, recall, F1, MAE, RMSE, and R-squared at a basic level.
Recognize when RAG, grounding, embeddings, content filtering, and prompt design apply.
Avoid service traps: OCR vs Document Intelligence, Translator vs Speech, Search vs model training, classification vs regression.

Next step: use targeted AI-900 practice questions to drill service selection, workload identification, responsible AI scenarios, and metric interpretation under exam-style wording.

Scenario Guide

AI Workloads