Azure AI Fundamentals Cheat Sheet (2026): Key Concepts + Exam Traps

Core Framework

AI-900 Domain Areas

AI-900 tests conceptual understanding of AI and Azure AI services — not programming or model development skills. Service selection and responsible AI principles are the most tested areas.

01
AI Workloads & Considerations — Types of AI, responsible AI principles
02
Machine Learning — Supervised, unsupervised, reinforcement learning; Azure ML
03
Computer Vision — Image classification, object detection, Azure Computer Vision
04
Natural Language Processing — Text analytics, translation, LUIS, Azure OpenAI
05
Document Intelligence & Knowledge Mining — Form Recognizer, Cognitive Search

Scenario Traps

Wrong instinct vs correct approach

A company wants to automatically categorize customer support emails

✕ Wrong instinct

Build a custom ML model to classify the emails

✓ Correct approach

Use Azure Cognitive Services Language (Text Analytics) for sentiment and classification, or Azure OpenAI with prompt engineering — pre-built NLP services handle common text classification tasks without custom model development

An AI hiring system consistently scores certain demographic groups lower

✕ Wrong instinct

This is a data quality issue — improve the training data

✓ Correct approach

This is a fairness concern under responsible AI principles. The training data likely contains historical bias — require a fairness audit, apply bias mitigation techniques, and establish human review for high-stakes decisions

A system needs to identify objects in images without any existing training data

✕ Wrong instinct

Use Azure Custom Vision to train a model from scratch

✓ Correct approach

Azure Computer Vision can detect common objects out of the box — Custom Vision is only needed for domain-specific objects the pre-built model wasn't trained on

Quick Rules

Know these cold

▸ Responsible AI — airness, Reliability, Privacy, Inclusiveness, Transparency, Accountability
▸ Supervised = labeled training data; Unsupervised = pattern finding without labels
▸ Regression = continuous output; Classification = categorical output
▸ Pre-built Cognitive Services for common tasks; Custom Vision/Azure ML for domain-specific training
▸ Azure OpenAI = generative AI with LLMs; Cognitive Services = specialized AI tasks
▸ Computer Vision for images; NLP/Language for text; Speech for audio
▸ Bias in AI comes from biased training data — fairness auditing is a design requirement

Self Check

Can you answer these without checking your notes?

In this scenario: "A company wants to automatically categorize customer support emails" — what should you do first?

Use Azure Cognitive Services Language (Text Analytics) for sentiment and classification, or Azure OpenAI with prompt engineering — pre-built NLP services handle common text classification tasks without custom model development

In this scenario: "An AI hiring system consistently scores certain demographic groups lower" — what should you do first?

This is a fairness concern under responsible AI principles. The training data likely contains historical bias — require a fairness audit, apply bias mitigation techniques, and establish human review for high-stakes decisions

In this scenario: "A system needs to identify objects in images without any existing training data" — what should you do first?

Azure Computer Vision can detect common objects out of the box — Custom Vision is only needed for domain-specific objects the pre-built model wasn't trained on

Failure Patterns

Common Exam Mistakes — What candidates get wrong

Confusing Azure AI service capabilities

Azure Cognitive Services covers Vision (image analysis), Speech (text-to-speech/STT), Language (NLP, translation), Decision (personalizer, anomaly detector). Candidates who pick the wrong service for a given AI use case fail service selection questions.

Misidentifying supervised vs. unsupervised learning

Supervised learning trains on labeled data (classification, regression). Unsupervised learning finds patterns in unlabeled data (clustering). Reinforcement learning learns through reward signals. Misidentifying the learning type for a described problem is a fundamental error.

Ignoring responsible AI principles in scenario questions

Microsoft's 6 responsible AI principles: Fairness, Reliability & Safety, Privacy & Security, Inclusiveness, Transparency, Accountability. Scenarios that test whether an AI system design is ethically sound require applying these principles.

Confusing regression with classification

Regression predicts a continuous numerical value (price, temperature). Classification predicts a category (spam/not-spam, disease/no-disease). Selecting the wrong algorithm type for a prediction scenario is a fundamental ML concept error.

Misidentifying when to use custom vs. pre-built models

Pre-built models (Cognitive Services) are used when the task is common (face detection, sentiment analysis). Custom models (Custom Vision, Azure ML) are used when domain-specific training is required.