Cheat sheet

Azure AI-102 Cheat Sheet

Plan + Manage Solution

20-25%of exam

Generative AI

15-20%of exam

Azure OpenAI + FoundryPrompt + TuningRAG groundingFine-tuningDALL-E images

Agentic Solution

5-10%of exam

AgentsToolsFoundry Agent ServiceAgent FrameworkMulti-agent orchestration

Computer Vision

10-15%of exam

Vision ServicesImage Analysis + OCRCustom VisionVideo IndexerSpatial Analysis

Natural Language Processing

15-20%of exam

Language (Text)Speech + CLUSentiment + PIITranslatorQuestion answering

Knowledge Mining

15-20%of exam

AI SearchDoc IntelligenceVector + hybrid searchContent UnderstandingSkillsets

Quick Facts

Exam
AI-102
Credential
Azure AI Engineer
Time
100 min
Pass
700/1000
Questions
40-60
Fee
$165 USD
Blueprint
Dec 23 2025
Retired
June 30 2026

Six Skill Areas

Plan | GenAI | Agent | Vision | NLP | Mining

Plan: 20-25%GenAI: 15-20%Agent: 5-10%Vision: 10-15%NLP: 15-20%Mining: 15-20%

Multi vs Single Service

Multi-service

  • One endpoint/key
  • Shared billing
  • Fast start

Single-service

  • Dedicated endpoint
  • Per-service RBAC
  • Network isolation

Simple vs isolated

Service Picker

  1. Generate text/imagesAzure OpenAI(Generative)
  2. Analyze imagesAzure AI Vision(Prebuilt)
  3. Own image classesCustom Vision(Trained)
  4. Extract form fieldsDocument Intelligence(Structured)
  5. Search your contentAzure AI Search(Retrieval)
  6. Transcribe audioAzure AI Speech(STT)
  7. Detect user intentCLU(Language)
  8. Multimodal extractionContent Understanding(Foundry Tools)

Resources + Endpoints

Multi-service resource
One key many services
Single-service resource
Dedicated endpoint per service
Azure OpenAI resource
Own kind OpenAI
Endpoint
Service base URL
Key pair
Two rotatable keys
Region
Fixed by resource
kind
Resource type flag

Responsible AI Layers

Filters | Blocklists | Shields | Harm detection

Filters: severity levelsBlocklists: exact termsShields: jailbreak blockHarm: category scores

Key vs Managed Identity

API key

  • Shared secret
  • Manual rotation
  • Leak risk

Managed identity

  • No secrets
  • Entra tokens
  • RBAC scoped

Secret vs identity

Auth Picker

  1. Azure-hosted appManaged identity(No secrets)
  2. Local dev + prodDefaultAzureCredential(Chained)
  3. Grant resource accessAzure RBAC(Roles)
  4. Private clients onlyPrivate endpoint(Disable public)
  5. Getting 429Backoff retry(Check quota)

Auth + Security

Managed identity
No stored secrets
DefaultAzureCredential
Chained auth methods
Microsoft Entra
Token-based identity
Azure RBAC
Role-based authorization
Private endpoint
Private network only
Disable public access
Close public endpoint
429 throttle
Retry exponential backoff

Content Safety vs Filters

Content Safety

  • Standalone service
  • Any content
  • Text and images

OpenAI filters

  • Built into deployment
  • Prompt/completion only
  • Not external content

Anywhere vs model-only

Deploy + Ops

IaC
Repeatable resource templates
CI/CD
Automated deployment pipeline
Container
Local/edge deployment
Azure Monitor
Metrics and logs
Diagnostic settings
Route resource logs
Cost management
Track and budget
Quota
Rate and token limits

Responsible AI

Content Safety
Standalone content moderation
Content filters
Model prompt/completion filter
Blocklists
Exact banned terms
Harm categories
Hate sexual violence self-harm
Severity thresholds
Per-category block levels
Prompt Shields
Block jailbreak attacks
Model card
Intended use limitations

Generative Flow

Deploy | Prompt | RAG | Evaluate | Optimize

Deploy: modelPrompt: templateRAG: groundEvaluate: scoreOptimize: tune

Fine-tune vs RAG

Fine-tune

  • Change behavior/style
  • Labeled JSONL
  • Slower iteration

RAG

  • Add fresh grounding
  • No retraining
  • Cheaper iteration

Behavior vs knowledge

Azure OpenAI + Foundry

Foundry hub
Shared compute and connections
Foundry project
One solution workspace
Azure OpenAI
GPT DALL-E embeddings
Standard deployment
Pay per token
Provisioned throughput
Reserved predictable capacity
DALL-E
Image generation model
Embedding model
Text to vectors

Prompt + Tuning

Prompt flow
Reusable multi-step pipeline
Prompt template
Reusable parameterized prompt
System message
Durable role instructions
Temperature
Controls randomness
Fine-tuning
Train on examples
Evaluation
Score models and flows
Grounding
Add your data

Agents

Agent
Reasoning plus tools
Tool
Callable agent capability
Foundry Agent Service
Managed agent hosting
Agent Framework
Code-first complex agents
Supervisor pattern
Routes specialist agents
Orchestration
Coordinate multi-agent workflow
Guardrails
Max turns limits

Classification vs Detection

Classification

  • What is present
  • Whole-image label
  • No boxes

Object detection

  • What and where
  • Bounding boxes
  • Location matters

What vs where

Vision Services

Image Analysis
Prebuilt tags and captions
OCR / Read
Extract printed handwritten text
Custom Vision
Train your images
Classification
What is present
Object detection
What and where
Compact domain
Exportable edge model
Video Indexer
Video insights and transcripts
Spatial Analysis
People movement in video

CLU vs Question Answering

CLU

  • Intents + entities
  • Bounded actions
  • Trained utterances

Question answering

  • Knowledge base
  • FAQ + docs
  • Best answer match

Action vs answer

Language (Text)

Sentiment analysis
Positive negative neutral mixed
Key phrase extraction
Main text points
NER
Detect named entities
PII detection
Find redact sensitive data
Language detection
Identify text language
Translator
Text and document translation

Speech + CLU

Speech-to-text
Transcribe spoken audio
Text-to-speech
Synthesize spoken output
SSML
Control pronunciation and prosody
Custom Speech
Adapt domain vocabulary
CLU
Intents and entities
Question answering
Knowledge base answers
Custom Translation
Domain-tuned translator model

Search Stack

Index | Indexer | Skillset | Query

Index: storeIndexer: ingestSkillset: enrichQuery: retrieve

Keyword vs Vector Search

Keyword

  • Exact terms
  • Lexical BM25
  • Literal match

Vector

  • Embedding similarity
  • Semantic meaning
  • Handles synonyms

Literal vs meaning

RAG + Search Picker

  1. Ground model answersRAG pattern(Retrieve + prompt)
  2. Exact keyword matchLexical search(BM25)
  3. Meaning similarityVector search(Embeddings)
  4. Best default searchHybrid + semantic(Combined)
  5. Persist enrichmentsKnowledge store(Projections)
  6. Enrich while indexingSkillset(OCR + phrases)

Doc Intelligence

Prebuilt model
Invoice receipt layout ID
Custom model
Train your forms
Composed model
Route multiple submodels
Layout model
Tables structure and text
Content Understanding
Multimodal schema extraction
Knowledge mining
Search across documents

Common Traps

Resource type

OpenAI needs own Others share multi-service

Auth method

Keys are secrets Managed identity tokenless

Moderation scope

Content Safety anywhere Filters model-only

Vision task

Classification labels image Detection locates objects

Language task

CLU classifies intent QnA matches answer

Search type

Keyword matches terms Vector matches meaning

Grounding vs training

RAG adds context Fine-tune changes behavior

Last Minute

  1. 1.Pass = 700 of 1000
  2. 2.100 minutes, 40-60 questions
  3. 3.Six weighted skill areas
  4. 4.OpenAI needs own resource
  5. 5.Managed identity over keys
  6. 6.Content Safety moderates anywhere
  7. 7.RAG grounds, fine-tune shapes
  8. 8.Classification labels; detection locates
  9. 9.CLU intents; QnA answers
  10. 10.Hybrid plus semantic ranker
  11. 11.Prompt Shields block jailbreaks
  12. 12.Prebuilt before custom models
Same family resources

Explore More Microsoft Azure Certifications

Continue into nearby exams from the same family. Each card keeps practice questions, study guides, flashcards, videos, and articles in one place.