Study Notes

Appearance

AI-900: Exam Guide ​

<- Overview | Cheatsheet ->

How the Exam Wants You to Think ​

AI-900 is a fundamentals exam. It checks whether you can recognize AI workloads, basic ML techniques, responsible AI principles, and Azure service fit.

Answer Philosophy ​

  1. Identify the workload first: vision, language, speech, document processing, ML, or generative AI.
  2. Pick the managed Azure service: Azure AI Vision, Azure AI Language, Azure AI Speech, Azure Machine Learning, Azure AI Foundry, or Azure OpenAI.
  3. Apply responsible AI by default: fairness, reliability, privacy, inclusiveness, transparency, and accountability are design requirements, not optional add-ons.

Keyword Detection Table ​

If you see...Look for this in the answer...
"predict a number" / "forecast amount"Regression
"predict a category" / "fraud or not fraud"Classification
"group similar customers" / "no labels"Clustering
"features and labels"Supervised learning dataset terms
"train and deploy custom ML models"Azure Machine Learning
"try many algorithms automatically"Automated ML
"extract text from image"OCR / Azure AI Vision
"draw boxes around objects"Object detection
"find faces"Azure AI Face
"important terms from text"Key phrase extraction
"people, places, dates, organizations"Entity recognition
"positive or negative feedback"Sentiment analysis
"audio to text"Speech recognition / Azure AI Speech
"text to audio"Speech synthesis / Azure AI Speech
"translate between languages"Translation
"generate answers, summaries, images, or code"Generative AI
"ground answers in company documents"RAG with retrieval context
"OpenAI models on Azure"Azure OpenAI Service
"build AI apps and select models"Azure AI Foundry / model catalog

Exam Traps ​

Watch out for these common mistakes!

  • Classification vs regression: Classification predicts labels. Regression predicts numbers.
  • Classification vs object detection: Image classification labels the image. Object detection locates objects with bounding boxes.
  • OCR vs document processing: OCR reads text. Document processing extracts structured data like fields and tables.
  • Language vs Speech: Azure AI Language handles text analytics. Azure AI Speech handles audio input/output.
  • RAG vs fine-tuning: RAG retrieves fresh context at runtime. Fine-tuning changes model behavior through additional training examples.
  • Foundry vs Azure OpenAI: Azure AI Foundry is the broader AI app platform. Azure OpenAI Service provides OpenAI model access in Azure.
  • Responsible AI keywords are literal: Fairness, reliability, privacy, inclusiveness, transparency, and accountability are frequently tested by definition.

Decision Quick Reference ​

Which ML technique? ​

text
Numeric prediction              -> Regression
Category prediction             -> Classification
Discover natural groups         -> Clustering
Images, speech, text at scale   -> Deep learning
Language and generative AI      -> Transformer-based models

Which Azure AI service? ​

text
General image analysis or OCR   -> Azure AI Vision
Face-specific detection         -> Azure AI Face
Text analytics                  -> Azure AI Language
Speech to text / text to speech -> Azure AI Speech
Custom ML lifecycle             -> Azure Machine Learning
OpenAI models through Azure     -> Azure OpenAI Service
AI app/model platform           -> Azure AI Foundry

Which responsible AI principle? ​

text
Avoid bias across groups              -> Fairness
Perform consistently and safely       -> Reliability and safety
Protect data and access               -> Privacy and security
Serve diverse users and abilities     -> Inclusiveness
Explain use, limits, and behavior     -> Transparency
Define ownership and human oversight  -> Accountability

Study Priority ​

PriorityWhy
Generative AI workloadsHighest domain weight at 20-25%
Responsible AI principlesEasy points if definitions are memorized
ML technique matchingCommon fundamentals trap area
Vision/NLP service matchingMany scenario questions are service selection

Final Strategy ​

  • Memorize the service mapping table before practice tests.
  • Treat every scenario as "input -> desired output -> workload -> Azure service".
  • Spend extra time on Domain 5 because generative AI carries the highest weight.
  • Book before the June 30, 2026 retirement date only if Microsoft still offers AI-900 in your region and language.

<- Overview | Cheatsheet ->