AI-900 Exam Guide: Azure AI Fundamentals

AI-900 (Azure AI Fundamentals) is Microsoft's entry-level AI certification — ~40–60 questions in about 60 minutes, $99 USD, passing 700/1000. It's vocabulary-and-concepts only: no coding, no math, no model training. It validates that you can name the right Azure AI service for a scenario and understand core ML/GenAI concepts. Think of it as the Azure counterpart to AWS's AIF-C01 AI Practitioner.

Exam format

• ~40–60 questions in ~60 minutes
• $99 USD (regional pricing applies)
• Passing score: 700/1000
• Question types: multiple choice, multiple response, matching, drag-and-drop, scenario blocks
• No prerequisites; pairs naturally after AZ-900

The five domains

• AI workloads and considerations (~15–20%): workload types + the six responsible AI principles
• Fundamental ML principles on Azure (~20–25%): model types, features/labels, evaluation, Azure ML
• Computer vision (~15–20%): Azure AI Vision, Custom Vision, Face, Document Intelligence, Video Indexer
• Natural language processing (~15–20%): Language service, Speech, Translator, CLU, question answering
• Generative AI (~15–20%): Azure OpenAI, prompts, RAG/grounding, AI Studio, Content Safety

Domain 1 — Workloads & responsible AI

Know the six workload families (ML prediction, vision, NLP, document intelligence, knowledge mining, generative AI) and match scenarios to them. Then the six responsible AI principles — fairness, reliability & safety, privacy & security, inclusiveness, transparency, accountability — with scenario recognition: biased outcomes = fairness, edge-case robustness = reliability, explainability docs = transparency.

Domain 2 — ML principles

The 'math' of the exam — but only conceptually: regression vs classification vs clustering; features vs labels; supervised vs unsupervised; train/validation/test splits; overfitting; and basic metrics (precision/recall, RMSE). Azure-side: Azure Machine Learning workspace concepts, AutoML, designer (visual pipelines), and the idea of deploying models as endpoints. You will not write code — you will match concepts to definitions.

Domain 3 — Computer vision

The task hierarchy: classification (label per image) → object detection (bounding boxes) → semantic segmentation (per-pixel). Service mapping is the tested skill: Azure AI Vision = prebuilt analysis + OCR; Custom Vision = train your own classifier/detector; Face = detection/verification/identification; Document Intelligence = structured fields from forms/invoices; Video Indexer = searchable insights from video.

Domain 4 — NLP & speech

Azure AI Language: sentiment, key phrases, NER, entity linking, PII detection, summarization, language detection. Conversational AI: CLU (intent + entities from utterances, the LUIS successor) vs question answering (FAQ-style knowledge base). Azure AI Speech: speech-to-text vs text-to-speech vs speaker recognition vs speech translation. Azure AI Translator: text/document translation.

Domain 5 — Generative AI

The newest and most-updated domain. Azure OpenAI Service = OpenAI models inside your Azure subscription (data not used for training). Concepts: LLMs predict tokens (why they hallucinate), prompt engineering (zero-shot vs few-shot), temperature/top-p/max tokens, RAG/grounding on your data, embeddings and vector search, and Azure AI Studio as the build platform. Responsible GenAI: content safety filters, transparency, mitigating hallucination with grounding.

AI-900 vs AWS AI Practitioner (AIF-C01)

Both are foundational AI vocabulary certs — overlapping heavily on ML/GenAI concepts. The differences are vendor surface: AIF centers on Amazon Bedrock (models, agents, guardrails, knowledge bases); AI-900 spreads across the Azure AI services portfolio plus Azure OpenAI. If you hold one, the other is a fast second credential — the conceptual 60% transfers; only the service names change.