How to Pass AI-900: The 2-Week AI Fundamentals Plan

AI-900 is a vocabulary exam wearing a technical costume. There's no code, no math beyond concept-level precision/recall, and no architecture design — but it covers a WIDE surface: five domains, ~15 services to know by name, and six responsible-AI principles to recognize by scenario. Most candidates with any cloud or IT background can prepare in 1–2 weeks. If you already hold AIF-C01 or AZ-900, a single intensive week is realistic.

Week 1 — Workloads, ML, and vision (days 1–7)

• Day 1: workload families + the six responsible AI principles — learn to map scenarios to principles (biased loan model = fairness; unpredictable edge cases = reliability)
• Days 2–3: ML concepts — model types, features/labels, supervised vs unsupervised, train/val/test, overfitting, precision/recall. Drill with our ML question set
• Day 4: Azure ML surface — workspace, AutoML, designer, endpoints (know what each is, not how to operate)
• Days 5–6: vision — classification/detection/segmentation ladder + the service map (Vision vs Custom Vision vs Face vs Document Intelligence vs Video Indexer)
• Day 7: first practice set — expect vision/NLP service-name confusion, that's normal

Week 2 — NLP, GenAI, and consolidation (days 8–14)

• Day 8: Azure AI Language — sentiment, key phrases, NER, PII, summarization + CLU vs question answering
• Day 9: Speech & Translator — direction matters: speech-to-text vs text-to-speech vs speaker recognition vs translation
• Days 10–11: GenAI — Azure OpenAI, prompting (zero/few-shot), temperature, RAG/grounding, hallucination, AI Studio, Content Safety
• Day 12: full timed mock — identify which domain bleeds points
• Day 13: targeted review — redo every wrong answer; make your own service-mapping flashcards for whatever confused you
• Day 14: second mock; book when consistently 80%+

The memorization strategy that works

Don't study services alphabetically — study them as DECISION TREES. 'Images?' → Vision vs Custom (prebuilt vs yours) vs Document Intelligence (structured fields) vs Face (identity). 'Text?' → Language (analyze) vs Translator (convert) vs QnA (FAQ matching). 'Audio?' → Speech (transcribe/synthesize/identify/translate). 'Generate?' → Azure OpenAI + Studio. 'Harmful content?' → Content Safety. Every AI-900 question is ultimately one of these branches.

If you're coming from AWS AIF-C01

• The conceptual material transfers ~60%: ML types, GenAI mechanics, responsible AI, hallucination/RAG concepts
• Bedrock ↔ Azure OpenAI: both are 'models as a managed service'; agents/knowledge bases ↔ AI Studio + RAG patterns
• The pure-Azure surface to learn fresh: the AI services portfolio (Vision/Language/Speech/Translator/Face/Document Intelligence) — AWS has no single matching catalog
• Microsoft's six responsible-AI principles have specific names — don't rely on 'close enough' recall

The concepts that trip people up

• Object detection vs segmentation vs classification — the granularity ladder
• CLU (intent for tasks) vs question answering (retrieve answers) — two different bot mechanisms
• Document Intelligence vs plain OCR — structured fields vs raw text
• Grounding vs fine-tuning — inject data at prompt-time vs retrain the model
• Sentiment vs safety — Language analyzes opinion; Content Safety filters harm
• Speaker recognition (who is speaking) vs speech recognition (what was said)

What 'ready' looks like

You're ready when service-name questions are reflex: see 'extract invoice fields' → Document Intelligence; see 'verify the person's identity from their face' → Face service; see 'answer from company docs only' → RAG grounding. Two weeks is genuinely enough — this is the friendliest exam in the Azure catalog after AZ-900.