AI-900 Cheat Sheet: Azure AI Fundamentals Quick Reference

AI-900 is a service-mapping exam: given a scenario, name the right Azure AI service or concept. This sheet is organized for exactly that. Work the AI-900 question bank and mock exam for applied practice.

The six responsible AI principles (memorize verbatim)

• Fairness — treat all groups equitably (biased outcomes = fairness violation)
• Reliability & safety — dependable under varied conditions, fails gracefully
• Privacy & security — protect data end-to-end
• Inclusiveness — works for everyone regardless of ability/background
• Transparency — understandable workings, disclosed limitations
• Accountability — humans remain responsible for the system

AI workload families → scenario keywords

• ML prediction — forecasting, churn, pricing, risk scoring
• Computer vision — anything image/video
• NLP — text/speech understanding
• Document intelligence — extracting fields/tables from forms & invoices
• Knowledge mining — searchable insights from unstructured corpora (Azure AI Search)
• Generative AI — creating text/images/code
• Anomaly detection — outliers vs learned normal patterns (fraud flagging)

ML concepts

• Regression = continuous number | Classification = category | Clustering = unlabeled groups
• Features (x) = inputs | Label (y) = the predicted output
• Supervised = labeled data (regression/classification); unsupervised = unlabeled (clustering)
• Train → validation → test; test set = final unbiased evaluation on unseen data
• Overfitting = memorized training data, fails on new data (great train score, poor production)
• Precision = of my positive predictions, how many real; recall = of real positives, how many caught
• AutoML = automated algorithm/hyperparameter search; designer = visual drag-drop pipelines; endpoints = deployed model access

Computer vision services

• Azure AI Vision = prebuilt analysis: captions, tags, objects, OCR/Read API
• Custom Vision = train YOUR classifier/detector on labeled images
• Azure AI Face = detection, verification (same person?), identification, liveness
• Document Intelligence = structured fields/tables from forms, invoices, IDs
• Video Indexer = faces, speech, OCR, scenes, topics — searchable video metadata
• Task ladder: classification (whole image) → object detection (boxes) → segmentation (per-pixel)

NLP & speech services

• Azure AI Language: sentiment, key phrases, NER, entity linking, PII detection, summarization
• CLU = utterance → intent + entities (task-driven bots; LUIS successor)
• Question answering = FAQ knowledge base matching (QnA Maker successor)
• Speech: speech-to-text (transcription) | text-to-speech (neural voices) | speaker recognition (who) | translation (audio→other language)
• Azure AI Translator = text/document machine translation

Generative AI

• Azure OpenAI Service = GPT/DALL-E/Whisper/embeddings in your subscription, enterprise security, no training on your data
• Azure AI Studio = GenAI build platform (model catalog, prompt flow, eval, deploy)
• Zero-shot = instructions only; few-shot = examples in prompt
• Temperature = randomness dial; top-p = nucleus sampling; max tokens = length
• RAG/grounding = retrieve your docs → inject into prompt → factual answers (the 'chat on your data' pattern)
• Hallucination = fluent-but-false output; mitigations: grounding, prompt constraints, human review
• Azure AI Content Safety = harm-category filtering (hate/violence/self-harm/sexual) for prompts + completions
• Embeddings = vectors capturing semantic meaning; enable semantic/vector search

The exam traps

• Language (analyze text) vs Translator (convert language) vs Speech (audio) — three different services
• CLU (intent/entity for tasks) vs question answering (FAQ retrieval) — different bot types
• Vision (prebuilt) vs Custom Vision (train yours) vs Document Intelligence (structured forms) vs Face (identity)
• Grounding (RAG = retrieve+inject) vs fine-tuning (retrain behavior) — different mechanisms for different goals
• Sentiment ≠ safety — Content Safety is the harm filter, not Language