6.1 Domain Summary and Key Comparisons
Key Takeaways
- The AI-900 covers five domains: AI Workloads (15-20%), ML Principles (20-25%), Computer Vision (15-20%), NLP (15-20%), and Generative AI (15-20%).
- Know the Azure AI service for each task: Vision = images, Language = text analysis, Speech = audio, OpenAI = generation, Search = retrieval, Document Intelligence = forms.
- Distinguish ML types: regression = predict numbers, classification = predict categories, clustering = discover groups (unsupervised).
- Responsible AI has six principles: fairness, reliability/safety, privacy/security, inclusiveness, transparency, accountability — tested across ALL domains.
- Key modern concepts: RAG reduces hallucinations, Copilot uses RAG + Microsoft Graph, Azure AI Foundry is the unified AI portal, content filters scan prompts AND responses.
Last updated: March 2026
Domain Summary and Key Comparisons
Quick Answer: This section consolidates the most important concepts, comparisons, and Azure service mappings from all five AI-900 domains into a single review. Use this as your final study reference before the exam.
Domain 1: AI Workloads and Considerations (15-20%)
Key Concepts to Remember
- AI is the broadest umbrella → ML is a subset → Deep Learning is a subset of ML → Generative AI is a subset of DL
- Seven AI workloads: prediction, classification, anomaly detection, computer vision, NLP, conversational AI, generative AI
- Six responsible AI principles: fairness, reliability/safety, privacy/security, inclusiveness, transparency, accountability
- Responsible AI appears across ALL five domains, not just Domain 1
Common Exam Questions
- "Which AI workload is this?" → Identify by input/output type
- "Which responsible AI principle?" → Match scenario to principle
- "What does Azure AI Content Safety do?" → Detects harmful content in four categories
Domain 2: Machine Learning Principles (20-25%)
Key Concepts to Remember
| Concept | Key Point |
|---|---|
| Features vs. Labels | Features = inputs; Labels = outputs (what you predict) |
| Training vs. Validation vs. Test | Training teaches, validation tunes, test evaluates |
| Overfitting | Great on training data, poor on new data |
| Supervised learning | Uses labeled data (regression, classification) |
| Unsupervised learning | Uses unlabeled data (clustering) |
ML Type Selection
| Question | Output Type | ML Type |
|---|---|---|
| "How much?" "How many?" | Continuous number | Regression |
| "Which category?" "Is it X?" | Discrete class | Classification |
| "What groups exist?" | Unknown groups | Clustering |
Evaluation Metrics
| ML Type | Metrics | Higher/Lower Is Better |
|---|---|---|
| Regression | MAE, RMSE, R-squared | MAE/RMSE: lower; R²: higher |
| Classification | Accuracy, Precision, Recall, F1 | All: higher |
| Clustering | Silhouette score | Higher (closer to 1) |
Azure ML Features
| Feature | What It Does | Code Required |
|---|---|---|
| AutoML | Automatically selects best algorithm | No |
| Designer | Drag-and-drop pipeline builder | No |
| Notebooks | Code-based ML development | Yes (Python/R) |
Domain 3: Computer Vision (15-20%)
Vision Task Selection
| Task | Output | When to Use |
|---|---|---|
| Image classification | Single label | "What is this image?" |
| Object detection | Objects + bounding boxes | "Where are objects?" |
| Semantic segmentation | Per-pixel classification | "What is every pixel?" |
| OCR | Extracted text | "What text is in this image?" |
| Facial detection | Face locations + attributes | "Where are faces?" |
Azure Vision Services
| Service | Purpose | Training Required |
|---|---|---|
| Azure AI Vision | Pre-built image analysis (caption, tag, detect) | No |
| Azure AI Custom Vision | Train custom classification/detection models | Yes (your images) |
| Azure AI Face | Face detection, verification, identification | No (but restricted access) |
| Azure AI Document Intelligence | Extract structured data from documents | No (pre-built) or Yes (custom) |
Domain 4: NLP (15-20%)
NLP Task Selection
| Task | What It Extracts | Example |
|---|---|---|
| Sentiment analysis | Positive/negative/neutral | "Great product!" → Positive |
| Key phrase extraction | Main topics | Article → ["AI", "Azure", "cloud"] |
| Named entity recognition | People, places, dates | "John in Paris on March 5" → Person, Location, Date |
| Language detection | Which language | "Bonjour" → French |
| PII detection | Sensitive personal data | SSN, phone numbers, emails |
Azure NLP Services
| Service | Purpose |
|---|---|
| Azure AI Language | Text analysis (sentiment, NER, CLU, Q&A) |
| Azure AI Speech | Audio processing (STT, TTS, translation) |
| Azure AI Translator | Text translation (100+ languages) |
Key Terminology Updates
| Old Name (Retired) | New Name (Current) |
|---|---|
| LUIS | Conversational Language Understanding (CLU) |
| QnA Maker | Custom Question Answering |
| Form Recognizer | Document Intelligence |
| Cognitive Search | Azure AI Search |
| Cognitive Services | Azure AI Services |
| Azure AI Studio | Azure AI Foundry |
On the Exam: Microsoft has renamed many services. Know the CURRENT names. If you see an old name in a question, recognize what it refers to.
Domain 5: Generative AI (15-20%)
Key Concepts
| Concept | Key Point |
|---|---|
| LLMs | Large Language Models that predict the next token |
| Transformers | Architecture using self-attention for context understanding |
| GPT | Generative Pre-trained Transformer |
| Hallucination | Model generates false but plausible information |
| RAG | Retrieve documents → include in prompt → generate grounded response |
| Grounding | Providing factual data to reduce hallucinations |
| Prompt engineering | Crafting effective prompts to control model behavior |
| Content filters | Scan prompts AND responses for harmful content |
| Temperature | 0 = deterministic; 1 = creative |
| Embeddings | Vector representations of text for semantic search |
Prompt Engineering Techniques
| Technique | What It Does | When to Use |
|---|---|---|
| Zero-shot | No examples provided | Simple, well-defined tasks |
| Few-shot | Provide examples | Pattern following, format control |
| Chain-of-thought | "Think step by step" | Math, logic, complex reasoning |
| Grounding | Provide context documents | Factual accuracy, reduce hallucinations |
| System message | Set persona and constraints | Control AI behavior |
Master Comparison: All Azure AI Services
| Azure Service | Category | What It Does |
|---|---|---|
| Azure AI Vision | Vision | Pre-built image analysis, OCR, spatial analysis |
| Azure AI Custom Vision | Vision | Train custom image models |
| Azure AI Face | Vision | Face detection, verification, identification |
| Azure AI Language | NLP | Sentiment, NER, CLU, Q&A, PII detection |
| Azure AI Speech | NLP | STT, TTS, speech translation, speaker recognition |
| Azure AI Translator | NLP | Text translation (100+ languages) |
| Azure AI Document Intelligence | Document | Extract structured data from forms |
| Azure AI Search | Search | Vector search, semantic ranking, RAG retrieval |
| Azure AI Content Safety | Safety | Detect harmful content in text/images |
| Azure OpenAI Service | Generative | GPT models, embeddings, image generation |
| Azure Machine Learning | ML | AutoML, Designer, model training/deployment |
| Azure AI Foundry | Portal | Build, evaluate, deploy AI applications |
Test Your Knowledge
A company wants to analyze customer feedback emails to identify the main topics being discussed, determine if the feedback is positive or negative, and identify any people or products mentioned by name. Which THREE Azure AI Language features should they use?
A
B
C
D
Test Your Knowledge
Which of the following correctly describes the relationship between Azure AI Vision and Azure AI Custom Vision?
A
B
C
D
Test Your KnowledgeMatching
Match each machine learning type to its correct description:
Match each item on the left with the correct item on the right
1
Regression
2
Classification
3
Clustering