AI-900 study guide
Key facts grouped by workload. AI-900 mostly tests whether you can match a scenario to the right kind of AI workload and Azure service.
Key facts by topic
Azure AI Services - Which Does What?
| Service | Capabilities |
|---|---|
| Vision | Image analysis, OCR, object detection, face detection (attributes) |
| Face | Face recognition (verify, identify), face detection |
| Language | CLU, sentiment, NER, key phrases, PII detection |
| Speech | Speech-to-text, text-to-speech |
| Translator | Text translation, document translation |
| Content Safety | Content moderation |
| Document Intelligence | Extract data from forms/documents |
Vision Service Limitations
| CAN Do | CANNOT Do |
|---|---|
| Detect adult content | Moderate content |
| Extract text (OCR) | Translate text |
| Detect faces | Recognize faces (use Face service) |
| Analyze images | Train custom models (use Custom Vision) |
Face Detection vs Face Recognition
| Face Detection/Analysis | Face Recognition |
|---|---|
| Age estimation | Verification (1:1) |
| Glasses detection | Identification (1:many) |
| 3D orientation | - |
| Gender | - |
| NOT emotion | - |
Microsoft Responsible AI Principles
| Principle | Covers |
|---|---|
| Fairness | Treat all people equally |
| Reliability & Safety | Works correctly, safe operation |
| Privacy & Security | Data protection, secure systems |
| Inclusiveness | Empower everyone, accessible |
| Transparency | Understandable, explainable |
| Accountability | PEOPLE responsible for AI |
Responsible AI Lifecycle Stages
- Identify - Detect and prioritize potential harms
- Measure - Assess frequency and severity
- Mitigate - Implement safeguards (content filtering, prompt engineering)
- Operate - Readiness plan, monitoring, incident response
Azure OpenAI Fine-Tuning Facts
- REST API: Supported
- Data format: JSONL (not CSV)
- Billing: Hourly hosting charge (not per-token)
- Storage only (not deployed): Free
Azure AI Foundry
| Concept | Purpose |
|---|---|
| Account | Infrastructure, policies, security |
| Project | Development workspace, deploy agents |
| Evaluations | Assess model quality, safety, metrics |
| Playground | Interactive experimentation |
| Prompt flow | Visual workflow design tool for orchestrating multi-step applications |
| Agents | Intelligent entities built to perform tasks |
Deployment Options
| Type | Use Case |
|---|---|
| Standard | Variable workloads, pay-per-use |
| Batch | Async processing, large data, non-urgent |
| Provisioned | Mission-critical, low-latency, high volume |
ML Fundamentals
| Concept | Definition |
|---|---|
| Feature | Input variable (what you know) |
| Label | Output variable (what you predict) |
| Training set | Fit the model |
| Validation set | Tune hyperparameters |
| Testing set | Final evaluation |
AutoML Support
| Supported | NOT Supported |
|---|---|
| Classification | Clustering |
| Regression | - |
| Time-series forecasting | - |
Sentiment Analysis Details
- Scores at BOTH document and sentence levels
- Confidence scores: 0 to 1 (not -1 to 1)
- All three scores (positive, neutral, negative) add up to 1
Transformer Architecture
- Faster than RNNs due to parallel processing
- Uses self-attention (not convolutions)
- Positional encodings preserve token order
- No sequential dependencies
NLP Workload Types
| Workload | Example |
|---|---|
| Language modeling | Command to smart action |
| Speech recognition | Phone call transcript |
| Sentiment analysis | Mining customer opinions |
| Entity recognition | Finding people/places |
| Key phrase extraction | Main talking points |
| Translation | Language conversion |