AI-901 — Microsoft Azure AI Fundamentals Quick Reference

Last revised: June 18, 2026

Compact AI-901 quick reference for Microsoft Azure AI workloads, Azure AI services, machine learning, vision, language, speech, search, 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-901) exam from Microsoft. Use it to connect scenario keywords to the right AI concept, Azure service, workload type, or responsible AI concern.

Focus on these recurring exam decisions:

Prebuilt AI service vs custom machine learning
Classification vs regression vs clustering
Computer vision vs OCR vs document intelligence
Language vs speech vs translation
Search index vs generative AI model vs RAG solution
Responsible AI principle vs technical mitigation
Real-time prediction vs batch processing

AI Workload and Azure Service Selection

Scenario keyword	Workload type	Best Azure fit	Watch for
Predict house price, sales amount, temperature	Regression	Azure Machine Learning	Numeric output
Predict churn, fraud yes/no, product category	Classification	Azure Machine Learning	Discrete class output
Group similar customers without labels	Clustering	Azure Machine Learning	Unsupervised learning
Detect unusual login, abnormal sensor reading	Anomaly detection	Azure AI services or Azure Machine Learning	“Unusual” does not always mean fraud
Identify objects in photos	Computer vision	Azure AI Vision or Custom Vision	Object detection returns locations
Read printed or handwritten text from images	OCR	Azure AI Vision	Text extraction, not semantic understanding
Extract fields from invoices, receipts, forms	Document intelligence	Azure AI Document Intelligence	Structured documents and key-value pairs
Detect faces in images	Face analysis	Azure AI Face	Identification/verification may have access and responsible-use constraints
Analyze sentiment, entities, key phrases	Natural language processing	Azure AI Language	Text input
Translate text between languages	Translation	Azure AI Translator	Text-to-text translation
Convert speech to text	Speech recognition	Azure AI Speech	Audio input, text output
Convert text to realistic speech	Speech synthesis	Azure AI Speech	Text input, audio output
Build chatbot or conversational assistant	Conversational AI / generative AI	Azure AI Foundry, Azure OpenAI, Azure Bot Service, Azure AI Language	Choose based on whether answers are scripted, grounded, or generated
Search internal documents with AI enrichment	Knowledge mining / search	Azure AI Search	Index, query, and enrich content
Summarize, draft, classify, reason over prompts	Generative AI	Azure OpenAI Service in Azure AI Foundry	Requires safety, grounding, prompt design
Detect harmful user-generated content	Content moderation	Azure AI Content Safety	Safety filter, not a full governance program

Core AI Concepts

Term	Exam meaning	Common trap
Artificial intelligence	Software performing tasks associated with human intelligence	AI is broader than machine learning
Machine learning	Models learn patterns from data	Not all AI uses ML
Deep learning	Neural-network-based ML, often with many layers	Often used for vision, speech, and language
Generative AI	Produces new text, images, code, or other content	Generated output can be incorrect
Model	Learned function used for prediction or generation	Model is not the same as dataset
Algorithm	Method used to train or run a model	Algorithm is the process, model is the result
Feature	Input variable used by a model	Poor features can limit model quality
Label	Known target value used in supervised training	Clustering usually has no labels
Training data	Data used to learn model patterns	Can contain bias or leakage
Validation data	Data used to tune and compare models	Do not treat as final proof of performance
Test data	Held-out data used for final evaluation	Should represent real-world use
Inference	Using a trained model to produce output	Different from training
Batch inference	Run predictions over many records asynchronously	Good for scheduled scoring
Real-time inference	Low-latency prediction for one request or small set	Good for apps and APIs
Grounding	Supplying source data to guide generated answers	Helps reduce hallucination but does not guarantee correctness
Prompt	Instruction or input to a generative model	Prompt quality strongly affects output
Token	Unit of text processed by a model	Not always the same as a word
Embedding	Numeric vector representing semantic meaning	Used for similarity search and RAG
Vector search	Finds items with similar embeddings	Different from exact keyword match
RAG	Retrieval-augmented generation	Combines search/retrieval with generation

Responsible AI Principles

Microsoft AI exams frequently test whether you can match a risk or mitigation to the right responsible AI concept.

Principle	Means	Scenario clues	Practical mitigation
Fairness	AI should treat people equitably	Unequal outcomes across groups	Representative data, bias testing, human review
Reliability and safety	AI should work dependably and avoid harm	Model fails in edge cases or unsafe conditions	Testing, monitoring, fail-safe design
Privacy and security	AI should protect data and systems	Sensitive data, identity, access, encryption	Data minimization, access control, encryption
Inclusiveness	AI should empower diverse users	Accessibility, different abilities, languages	Inclusive design, accessibility testing
Transparency	Users should understand AI behavior and limits	Explainability, disclosure, confidence	Documentation, explanations, model cards
Accountability	People remain responsible for AI systems	Ownership, audit, governance	Human oversight, audit trails, review process

Responsible AI Traps

Trap	Correct exam interpretation
“The model is accurate overall, so it is fair.”	Overall accuracy can hide subgroup harm.
“Content filters make generative AI safe.”	Filters help, but safety also needs design, testing, monitoring, and human oversight.
“An AI system decides, so nobody is accountable.”	People and organizations remain accountable.
“Explainability is the same as privacy.”	Explainability supports transparency; privacy protects data.
“More data always reduces bias.”	More biased data can reinforce bias.
“Grounding eliminates hallucinations.”	Grounding reduces risk but outputs still require validation.

Machine Learning Task Reference

Task	Input	Output	Uses labels?	Example
Binary classification	Features	One of two classes	Yes	Fraud/not fraud
Multiclass classification	Features	One of many classes	Yes	Product category
Regression	Features	Numeric value	Yes	Forecast revenue
Clustering	Features	Group assignment	No	Customer segments
Anomaly detection	Features or time series	Normal/unusual	Sometimes	Detect abnormal device behavior
Object detection	Image	Labels plus bounding boxes	Yes	Locate cars in traffic images
Image classification	Image	Image-level label	Yes	Cat, dog, or bird
Text classification	Text	Category label	Yes	Support ticket routing
Named entity recognition	Text	Entity spans and types	Prebuilt or custom	Person, location, date
Question answering	Question plus knowledge base	Answer	Often grounded in content	FAQ bot

Supervised, Unsupervised, and Reinforcement Learning

Learning type	Description	Typical exam examples
Supervised learning	Learns from labeled examples	Classification, regression
Unsupervised learning	Finds patterns without target labels	Clustering, some anomaly detection
Reinforcement learning	Learns actions from rewards	Robotics, game playing, optimization scenarios

Evaluation Metrics

Use metric names carefully. AI-901 usually tests concepts, not deep math, but the distinctions matter.

Metric	Used for	High-yield meaning
Accuracy	Classification	Share of predictions that are correct
Precision	Classification	Of predicted positives, how many were truly positive
Recall	Classification	Of actual positives, how many were found
F1 score	Classification	Balance between precision and recall
Confusion matrix	Classification	Counts true positives, false positives, true negatives, false negatives
Mean absolute error	Regression	Average absolute prediction error
Mean squared error	Regression	Penalizes larger errors more heavily
Root mean squared error	Regression	Error measure in original target units
R-squared	Regression	How much variance the model explains
AUC	Classification	Ability to separate classes across thresholds

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

Precision vs Recall Decision Points

Scenario priority	Prefer	Why
Avoid false alarms	Higher precision	Positive predictions should be trustworthy
Catch as many true cases as possible	Higher recall	Missing positives is costly
Balance both	F1 score	Useful when classes are imbalanced
Medical screening, safety alerting	Recall is often emphasized	False negatives can be dangerous
Automatic blocking, fraud accusation	Precision is often emphasized	False positives can harm users

Azure Machine Learning Reference

Use Azure Machine Learning when the scenario requires building, training, evaluating, managing, or deploying custom machine learning models.

Component	Purpose	Exam clue
Workspace	Top-level Azure ML resource for assets and collaboration	Central place for ML projects
Data asset	Versioned reference to data	Reuse training data
Compute instance	Managed development workstation	Notebook authoring and experimentation
Compute cluster	Scalable training compute	Train jobs at scale
Notebooks	Code-first experimentation	Python, SDK, data science
Designer	Visual drag-and-drop ML pipelines	Low-code pipeline creation
Automated ML	Automatically tries algorithms and settings	Quickly find a good model
Jobs	Training or processing runs	Track experiments
Models	Registered trained artifacts	Version and deploy trained models
Online endpoint	Real-time inference	App needs immediate prediction
Batch endpoint	Batch scoring	Large scheduled prediction job
Pipelines	Repeatable multi-step workflow	Data prep, training, evaluation, deployment
Responsible AI dashboard	Model assessment and explanation support	Error analysis, fairness, explainability

When to Choose Azure Machine Learning

Choose Azure Machine Learning when…	Prefer prebuilt Azure AI services when…
You need a custom predictive model	The workload is common, such as OCR or translation
You have domain-specific labeled data	You want fast integration with minimal training
You must control training, evaluation, and deployment	You do not need to design an algorithm
You need MLOps lifecycle management	The scenario asks for a ready API
You need custom regression/classification/clustering	The scenario asks for vision, speech, language, or document extraction

Common ML Traps

Trap	Correct answer pattern
Using Azure AI Vision to predict sales	Use Azure Machine Learning regression
Using clustering when target labels exist	Use supervised learning if labels exist
Treating validation data as training data	Keep validation/test sets separate
Deploying a batch endpoint for an interactive app	Use an online endpoint
Assuming AutoML removes evaluation needs	You still evaluate, validate, and monitor

Azure AI Services Overview

Azure AI services provide prebuilt APIs for common AI workloads. They are usually the best choice when the exam scenario says “use a prebuilt model,” “minimize custom training,” or “call an API.”

Service family	Use for	Examples
Azure AI Vision	Image analysis and OCR	Caption images, detect objects, read text
Azure AI Face	Face detection and analysis	Detect faces, estimate attributes where supported
Azure AI Document Intelligence	Structured document extraction	Invoices, receipts, forms, IDs
Azure AI Language	Text analytics and language understanding	Sentiment, entities, key phrases, PII, custom classification
Azure AI Translator	Text translation	Translate text between languages
Azure AI Speech	Audio and voice workloads	Speech-to-text, text-to-speech, speech translation
Azure AI Search	Search and knowledge mining	Index documents, enrich content, vector search
Azure OpenAI Service	Generative AI models in Azure	Chat, summarization, generation, embeddings
Azure AI Content Safety	Detect harmful content	Text/image safety classification

Resource and Authentication Concepts

Concept	Meaning	Exam note
Endpoint	Service URL used by client apps	Required for API calls
Key	Secret used to authenticate requests	Protect and rotate keys
Microsoft Entra ID	Identity-based authentication	Prefer for enterprise access control when supported
Managed identity	Azure-managed service identity	Avoids storing credentials in code
Region	Azure location for the resource	Some capabilities vary by region
Multi-service resource	One resource for multiple Azure AI services	Simplifies access for supported services
Single-service resource	Dedicated resource for one service	Useful for isolated management or service-specific needs

Computer Vision Quick Reference

Requirement	Choose	Output
Describe contents of an image	Azure AI Vision image analysis	Captions, tags, objects
Classify entire image into categories	Image classification / Custom Vision	Image-level label
Detect and locate multiple items	Object detection / Custom Vision	Labels plus bounding boxes
Read text in image	OCR with Azure AI Vision	Extracted text and location
Extract fields from structured business docs	Azure AI Document Intelligence	Fields, tables, confidence values
Analyze video content	Azure AI Video Indexer	Transcripts, scenes, insights
Detect faces in images	Azure AI Face	Face detection and related analysis

Vision Distinctions

Pair	Difference
Image classification vs object detection	Classification labels the whole image; detection locates objects with bounding boxes.
OCR vs Document Intelligence	OCR extracts text; Document Intelligence extracts structured fields from documents.
Vision vs Machine Learning	Vision is prebuilt image analysis; Azure ML is for custom model training.
Face detection vs face identification	Detection finds faces; identification attempts to match identity and is more sensitive.
Tags vs captions	Tags are labels; captions are natural-language descriptions.

Document Intelligence Reference

Document task	Service capability	Example
Extract text from pages	OCR/read	Scan to text
Extract key-value pairs	Document analysis	“Invoice Date: 2026-06-01”
Extract tables	Layout analysis	Rows and columns from forms
Use prebuilt models	Prebuilt document models	Receipts, invoices, identity documents
Train for a custom form type	Custom document model	Company-specific contract format
Automate document workflow	Document Intelligence plus downstream app	Validate, route, store extracted data

Document Intelligence Traps

Trap	Correct interpretation
“Read all text from a photo of a sign”	OCR may be enough.
“Extract vendor, total, tax from invoices”	Use Document Intelligence.
“Understand document meaning and answer questions”	May need Azure AI Search and generative AI/RAG.
“Translate a document’s text”	Use Translator after extraction, not Document Intelligence alone.

Natural Language Processing Reference

Requirement	Choose	Notes
Detect sentiment	Azure AI Language	Positive, neutral, negative; may include opinion mining
Extract key phrases	Azure AI Language	Identifies main topics
Identify entities	Azure AI Language	People, locations, organizations, dates, quantities
Detect PII	Azure AI Language	Finds sensitive personal information in text
Detect language	Azure AI Language	Determines input language
Classify support tickets	Custom text classification	Requires labeled examples
Extract domain-specific entities	Custom named entity recognition	Example: part numbers, policy IDs
Build intent-based conversational app	Conversational language understanding	Intents and entities
Create FAQ-style answers	Question answering	Grounded in provided knowledge source
Translate text	Azure AI Translator	Text-to-text translation
Summarize or generate text	Azure OpenAI Service	Generative AI, not basic text analytics

Language Service Distinctions

Pair	Difference
Sentiment analysis vs key phrase extraction	Sentiment detects emotion/opinion; key phrases detect important topics.
Entity recognition vs key phrases	Entities are typed items; key phrases are important terms.
PII detection vs entity recognition	PII focuses on sensitive personal data.
Language detection vs translation	Detection identifies language; translation converts text.
Question answering vs generative chat	Question answering returns grounded answers from a knowledge source; generative chat can produce broader responses.
Intent vs entity	Intent is what the user wants; entity is data needed to fulfill it.

Speech and Translation Reference

Requirement	Choose	Input	Output
Transcribe audio	Azure AI Speech speech-to-text	Audio	Text
Generate spoken audio	Azure AI Speech text-to-speech	Text	Audio
Translate spoken language	Azure AI Speech translation	Audio	Translated text or speech
Translate written text	Azure AI Translator	Text	Text
Enable voice command app	Azure AI Speech plus language understanding	Audio	Intent/action
Create accessible audio content	Text-to-speech	Text	Natural-sounding voice

Speech Traps

Trap	Correct answer
Text translation from English to French	Azure AI Translator
Audio transcription from meeting recording	Azure AI Speech speech-to-text
Spoken translation during conversation	Azure AI Speech translation
Voice response from chatbot	Azure AI Speech text-to-speech
Understanding user intent after transcription	Use language understanding after speech-to-text

Azure AI Search and Knowledge Mining

Use Azure AI Search when the scenario involves indexing, querying, enriching, or retrieving information from content repositories.

Concept	Meaning	Exam clue
Index	Searchable structure containing fields	Query documents quickly
Indexer	Pulls data from supported data sources into an index	Automate ingestion
Data source	Where content comes from	Blob storage, databases, documents
Skillset	AI enrichment pipeline	OCR, entity extraction, key phrases
Cognitive skills	Built-in AI enrichments	Extract text, translate, detect entities
Custom skill	Your own enrichment logic	Call custom API/function
Analyzer	Controls text processing for search	Tokenization, language-aware search
Semantic ranking	Improves relevance using semantic understanding	Better result ordering
Vector search	Finds semantically similar content	Embedding similarity
Hybrid search	Combines keyword and vector search	Better retrieval coverage
RAG	Search retrieves grounding data for a generative model	“Answer from company documents”

Azure AI Search vs Generative AI

Requirement	Use Azure AI Search	Use Generative AI
Retrieve matching documents	Yes	Not by itself
Rank search results	Yes	Not primary role
Enrich content during indexing	Yes	Sometimes
Generate natural-language answer	Often with RAG	Yes
Ground answers in internal documents	Retrieval layer	Generation layer
Store searchable vectors	Yes	No, model creates embeddings but search stores/queries them

Generative AI Reference

Generative AI scenarios on AI-901 usually test vocabulary, architecture, prompt roles, safety, and when to use Azure OpenAI Service or Azure AI Foundry.

Concept	Meaning	Exam relevance
Large language model	Model trained to process and generate language	Chat, summarization, classification, reasoning
Foundation model	General-purpose model adaptable to many tasks	Used as base model
Prompt engineering	Designing inputs to guide model behavior	Improves relevance and format
System message	High-priority instruction that sets behavior	“You are a support assistant…”
User message	End-user request	“Summarize this article…”
Assistant message	Model response in conversation history	Maintains context
Temperature	Controls randomness/creativity	Lower is more deterministic, higher is more varied
Context window	Amount of input/output context the model can consider	Long prompts and documents may exceed capacity
Embeddings	Numeric semantic representation	Similarity search and RAG
Fine-tuning	Further training a model on task examples	Different from grounding
Grounding	Adding retrieved source content to prompt	Helps answer from trusted data
Hallucination	Plausible but incorrect generated output	Key generative AI risk
Content filtering	Detecting/blocking unsafe content	Safety control, not truth verification

Common Generative AI Workloads

Workload	Example prompt goal	Azure capability
Summarization	“Summarize this transcript in 5 bullets”	Azure OpenAI Service
Drafting	“Write a response to this customer email”	Azure OpenAI Service
Classification	“Classify this ticket as billing, technical, or sales”	Azure OpenAI Service or Azure AI Language
Extraction	“Extract product names and quantities”	Azure OpenAI Service, Language, or Document Intelligence depending on source
Code generation	“Create a function that validates input”	Azure OpenAI Service
Embeddings	“Create vectors for documents”	Azure OpenAI Service embeddings
RAG assistant	“Answer using only these indexed documents”	Azure AI Search plus Azure OpenAI
Image generation	“Create an illustration from this description”	Generative image model where available

Prompt Message Pattern

[
  {
    "role": "system",
    "content": "Answer only from the provided policy excerpt. If the answer is not present, say you do not know."
  },
  {
    "role": "user",
    "content": "What is the reimbursement deadline?"
  }
]

RAG Decision Flow

Step	Purpose	Azure service often used
Ingest documents	Bring enterprise content into a searchable system	Azure AI Search
Chunk content	Split long documents into retrievable sections	App logic / indexing pipeline
Create embeddings	Represent text semantically	Azure OpenAI Service
Store vectors	Enable similarity retrieval	Azure AI Search
Retrieve relevant chunks	Find grounding content	Azure AI Search
Build prompt	Combine instruction, question, and retrieved context	Application / orchestration
Generate answer	Produce natural-language response	Azure OpenAI Service
Cite sources and monitor	Improve trust and governance	App logic, Azure monitoring tools

Generative AI Traps

Trap	Correct interpretation
“Fine-tuning is required to answer from company documents.”	Often RAG/grounding is the better first choice.
“A chatbot always requires generative AI.”	Simple FAQ or intent bots can use question answering or language understanding.
“Embeddings generate final answers.”	Embeddings support retrieval and similarity; generation is separate.
“Content filters ensure factual accuracy.”	Filters target unsafe content, not correctness.
“More creative output is always better.”	Business workflows often need deterministic, grounded responses.
“The model remembers private uploaded documents automatically.”	Apps must manage retrieval, context, storage, and security.

Conversational AI and Bots

Requirement	Choose	Notes
Menu-like bot with fixed paths	Azure Bot Service / Bot Framework	Rule-based or scripted conversation
FAQ bot over known content	Azure AI Language question answering	Grounded answers
Bot that detects intent and entities	Conversational language understanding	“Book flight” plus date/location
Generative assistant over enterprise documents	Azure AI Foundry / Azure OpenAI + Azure AI Search	RAG pattern
Voice-enabled assistant	Speech + language/generative AI	Speech-to-text and text-to-speech
Multi-channel bot	Azure Bot Service	Connect to chat channels

Intent and Entity Example

User utterance	Intent	Entities
“Book a table for four tomorrow at 7 PM”	Make reservation	Party size: four; date: tomorrow; time: 7 PM
“Cancel my order 12345”	Cancel order	Order ID: 12345
“What is the refund policy?”	Ask policy question	Policy topic: refund

Security, Privacy, and Governance Basics

Control	What it protects	Exam scenario
Authentication	Verifies caller identity	App calling an AI API
Authorization	Controls allowed actions	Restrict who can manage models or indexes
Managed identity	Avoids embedded secrets	Azure service securely calls another service
Key rotation	Reduces risk from exposed keys	AI service key compromised
Network controls	Limits service exposure	Private enterprise workloads
Encryption	Protects data at rest/in transit	Sensitive documents or prompts
Data minimization	Reduces privacy exposure	Do not send unnecessary PII
Logging and monitoring	Detects issues and supports audit	Track failures, latency, unsafe outputs
Human review	Adds oversight for high-impact decisions	Approvals, escalations, appeals
Content safety	Detects unsafe content	User-generated or generated content

Exam Scenario Shortcuts

If the question says…	Think…
“No labeled data”	Clustering or unsupervised approach
“Predict a number”	Regression
“Predict a category”	Classification
“Find unusual values”	Anomaly detection
“Bounding boxes”	Object detection
“Extract text from image”	OCR
“Extract totals from invoices”	Document Intelligence
“Translate text”	Translator
“Transcribe audio”	Speech-to-text
“Generate spoken response”	Text-to-speech
“Detect sentiment”	Azure AI Language
“Find people, locations, organizations”	Named entity recognition
“Remove or detect personal data”	PII detection
“Index and search documents”	Azure AI Search
“Answer from internal documents”	RAG with Azure AI Search and Azure OpenAI
“Train a custom model”	Azure Machine Learning
“Low-code ML pipeline”	Azure Machine Learning designer
“Automatically select ML algorithm”	Automated ML
“Deploy for immediate app calls”	Online endpoint
“Score many records overnight”	Batch endpoint
“Unsafe text/image detection”	Azure AI Content Safety
“Explain why a model predicted something”	Transparency/explainability
“Unequal impact across demographic groups”	Fairness
“Who is responsible for the AI decision?”	Accountability

High-Yield Comparison Tables

Azure AI Services vs Azure Machine Learning vs Azure OpenAI

Need	Azure AI services	Azure Machine Learning	Azure OpenAI Service
Prebuilt vision, speech, language APIs	Best fit	Not primary	Sometimes
Train custom predictive model	Not primary	Best fit	Not primary
Manage ML experiments and deployments	No	Best fit	No
Generate natural-language text	Limited	No	Best fit
Create embeddings	No	Not primary	Best fit
Analyze documents with forms/fields	Document Intelligence	Possible custom approach	Possible but not primary
Build RAG answer system	Search/Language components	Not primary	Generation component
Need minimal custom model training	Best fit	Less likely	Best fit for generative tasks

Azure AI Language vs Azure OpenAI

Requirement	Better fit
Deterministic sentiment analysis	Azure AI Language
Named entity recognition	Azure AI Language
PII detection	Azure AI Language
Custom text classification with labels	Azure AI Language
Open-ended summarization	Azure OpenAI
Drafting new content	Azure OpenAI
Conversational reasoning over context	Azure OpenAI
FAQ over curated source content	Azure AI Language question answering or RAG, depending on complexity

OCR vs Document Intelligence vs Azure AI Search

Requirement	Best fit
Read text from image	OCR
Extract fields from forms	Document Intelligence
Search many documents	Azure AI Search
Enrich documents during indexing	Azure AI Search skillset
Answer questions over documents	Azure AI Search + generative AI
Convert scanned documents into searchable text	OCR plus Azure AI Search

Compact Review Checklist

Before exam day, confirm you can:

Match classification, regression, clustering, and anomaly detection to scenarios.
Explain features, labels, training, validation, testing, and inference.
Choose Azure Machine Learning for custom model lifecycle tasks.
Choose the right prebuilt Azure AI service for vision, language, speech, translation, and documents.
Distinguish OCR from Document Intelligence.
Distinguish image classification from object detection.
Distinguish Speech from Translator.
Distinguish Azure AI Search from Azure OpenAI Service.
Explain RAG, embeddings, grounding, tokens, and prompt roles.
Identify responsible AI principles from scenario clues.
Recognize security basics: endpoint, key, managed identity, authorization, privacy, and monitoring.
Avoid assuming that generative AI outputs are factual, unbiased, private, or safe without controls.

Practical Next Step

Use this Quick Reference as a checklist, then practice with scenario-based AI-901 questions that force you to choose the correct Azure service, AI workload type, metric, and responsible AI principle under exam-style wording.

Scenario Guide

Identify AI Concepts and Capabilities