This cheat sheet is a fast, exam-focused review of everything on AI-900 (Microsoft Azure AI Fundamentals). It is built for your final days of prep — skim it, test yourself, and drill anything that feels unfamiliar. It is a memory aid, not a replacement for understanding: AI-900 questions describe a use case and ask which Azure AI service or ML approach fits, so know when to use each one.
Exam Snapshot
Passing Score
700 / 1000
Cost
$99 USD
Questions
40–60
Time
45 minutes
AI Workloads & ML
30–40%
Computer Vision & NLP
30–40%
Generative AI
20–25%
1. AI Workloads & Responsible AI
Common AI workload types
| Workload | What it does |
|---|---|
| Machine Learning | Predict values or categories from data (foundation for the others). |
| Computer Vision | Interpret images and video (objects, faces, text). |
| Natural Language Processing | Understand and generate written and spoken language. |
| Document Intelligence | Extract fields and data from forms and documents. |
| Knowledge Mining | Index large content stores to make them searchable (Azure AI Search). |
| Generative AI | Create new content — text, images, code — from prompts. |
The 6 Responsible AI principles — memorize these
| Fairness | Treat all groups equitably; avoid bias. |
| Reliability & Safety | Perform consistently and handle unexpected conditions safely. |
| Privacy & Security | Protect data and respect user privacy. |
| Inclusiveness | Empower everyone, including people with disabilities. |
| Transparency | Make systems understandable; explain decisions. |
| Accountability | People remain responsible for AI systems they build/operate. |
Exam tip: Match the scenario to the principle. "The model denies loans to one demographic more often" = Fairness. "Users should know they're talking to a bot" = Transparency. "A human must approve the AI's decision" = Accountability.
2. Machine Learning Fundamentals
ML approaches — pick by the scenario
| Type | Predicts | Example |
|---|---|---|
| Regression | A continuous number | House price, temperature |
| Classification | A category / label | Spam vs not spam |
| Clustering | Groups in unlabeled data | Customer segments |
Labeled data → supervised (regression + classification). Unlabeled data → unsupervised (clustering).
Key ML terms
- Features: Input variables (columns) used to make a prediction.
- Label: The value the model predicts.
- Training vs validation data: Train on one split, evaluate on a held-out split.
- Model: The trained output that makes predictions on new data.
Classification evaluation metrics
| Accuracy | Overall correct predictions / total. |
| Precision | Of predicted positives, how many were right. |
| Recall | Of actual positives, how many were found. |
| F1 score | Balance of precision and recall. |
Azure Machine Learning tools
- Automated ML (AutoML): Tries many algorithms automatically to find the best model.
- Designer: Drag-and-drop, no-code pipeline builder.
- Data labeling: Create labeled datasets for training.
- Compute: Managed resources to train and deploy models.
Test the concepts you just reviewed
Try 40 Free AI-900 Practice Questions
Scenario-based questions with detailed explanations. No credit card required.
Start Free Practice →3. Computer Vision
| Capability | What it does |
|---|---|
| Image classification | Assign a label to a whole image. |
| Object detection | Locate and label multiple objects with bounding boxes. |
| OCR | Read printed and handwritten text from images. |
| Facial detection / recognition | Detect and identify faces (Azure AI Face). |
| Document Intelligence | Extract key/value pairs and tables from forms and invoices. |
Service to know: Azure AI Vision covers analysis, OCR, and tagging. Custom Vision lets you train your own image classification / object detection models.
4. Natural Language Processing
| Capability | What it does |
|---|---|
| Key phrase extraction | Pull the main talking points from text. |
| Sentiment analysis | Score text as positive, negative, or neutral. |
| Entity recognition (NER) | Identify people, places, dates, organizations. |
| Language detection | Identify the language of text. |
| Translation | Translate text/speech between languages (Azure AI Translator). |
| Speech to text / text to speech | Transcribe and synthesize spoken language (Azure AI Speech). |
| Conversational language understanding | Map utterances to intents and entities for bots. |
Service to know: Azure AI Language handles text analytics, NER, sentiment, and conversational understanding. Azure AI Speech handles speech-to-text, text-to-speech, and translation.
5. Generative AI & Azure OpenAI
Core concepts
| Large Language Model (LLM) | Model trained on huge text corpora that generates language. |
| Prompt | The input/instruction you give the model. |
| Completion | The model's generated response. |
| Token | Chunk of text the model processes; billing/limits are token-based. |
| Embedding | Numeric vector representing meaning; powers search and RAG. |
| Grounding / RAG | Supply your own data so answers are based on it, reducing hallucination. |
| Copilot | An AI assistant embedded in an app to help users complete tasks. |
Where it runs: Azure OpenAI Service provides GPT models for text/code, plus image and embedding models — with Azure security, regional control, and Responsible AI content filters. Azure AI Foundry is the portal/platform for building and deploying generative AI solutions.
Exam tip: Generative AI is the newest and most heavily weighted domain. Know prompts, tokens, grounding/RAG, Copilots, and that content filters enforce Responsible AI.
6. Azure AI Services Quick Map
| Need… | Use… |
|---|---|
| Analyze / tag images, OCR | Azure AI Vision |
| Train your own image model | Custom Vision |
| Sentiment, NER, key phrases | Azure AI Language |
| Speech-to-text, TTS, translation | Azure AI Speech |
| Extract data from forms/invoices | Azure AI Document Intelligence |
| Search large content stores | Azure AI Search |
| Generate text/code/images | Azure OpenAI Service |
| Build and deploy AI solutions | Azure AI Foundry |
| Train/deploy custom ML models | Azure Machine Learning |
7. Acronym Quick List
AI — Artificial Intelligence
ML — Machine Learning
NLP — Natural Language Processing
NER — Named Entity Recognition
OCR — Optical Character Recognition
LLM — Large Language Model
RAG — Retrieval-Augmented Generation
AutoML — Automated Machine Learning
CLU — Conversational Language Understanding
TTS / STT — Text-to-Speech / Speech-to-Text
F1 — Balance of precision & recall
RAI — Responsible AI
Reviewed the cheat sheet? Now prove you're ready.
Take the free 5-minute AI-900 readiness quiz — no signup required — or jump into 40 free practice questions calibrated harder than the real exam.
Common Questions
Is a cheat sheet enough to pass AI-900?
A cheat sheet is a fast review and memory aid, not a substitute for understanding. Use it to consolidate concepts after working through Microsoft Learn and practice questions. AI-900 questions describe a use case and ask which Azure AI service or ML approach fits, so you need to know when to use each one — not just recognize its name.
What is the difference between supervised and unsupervised learning?
Supervised learning trains on labeled data where the correct answer is known — that covers regression (predict a number) and classification (predict a category). Unsupervised learning finds patterns in unlabeled data, such as clustering customers into segments. If the scenario gives historical outcomes to learn from, it's supervised.
What is the passing score for AI-900?
700 out of 1000 on a scaled scoring system (roughly 70%). The exam has 40–60 questions, a 45-minute time limit, and costs $99 USD.
What is the difference between Azure AI Vision and Custom Vision?
Azure AI Vision is a prebuilt service that analyzes, tags, and reads text from images out of the box. Custom Vision lets you train your own image classification or object detection model on your own labeled images when the prebuilt models don't recognize what you need.
About MSCertQuiz
MSCertQuiz provides 500 practice questions per certification, calibrated harder than the real exam so test day feels easier. Questions are built by certified professionals and updated for 2026 exam objectives. Start with 40 free questions — no credit card required.