Workload Taxonomies and the Microsoft Six
Two vocabularies the exam tests that you conceptually know but haven't memorized in Microsoft's phrasing: the workload taxonomies (how the exam categorizes AI problems) and Microsoft's six Responsible AI principles (asked by name, frequently).
AI workload categories
When a question says "identify the workload type," it means one of these:
- Machine learning — predictions from data: regression (numbers), classification (categories), clustering (groupings). Your Module 13 lesson 5 depth is exactly exam depth; no more is needed.
- Computer vision — image classification (what is this picture), object detection (what's in it and where, with bounding boxes), OCR (text in images), facial detection/analysis. Know those four apart — "classification vs. detection" is a perennial question.
- Natural language processing — sentiment analysis, key-phrase extraction, entity recognition, translation, speech-to-text/text-to-speech, language understanding.
- Document intelligence / knowledge mining — extracting structured data from documents and making content searchable (Document Intelligence + AI Search).
- Generative AI — creating content: text, code, images; chat; summarization; RAG- grounded answering; agents. The heaviest domain on AI-901.
Drill: hear a scenario, name the workload, then name the Azure service (lesson 2's table). Two-step, every time.
The Microsoft six Responsible AI principles
Microsoft's official six — the exam asks you to match principle to scenario, by name:
- Fairness — the system treats comparable people comparably; watch for training-data bias. (Scenario cue: loan approvals differ by demographic.)
- Reliability and safety — the system behaves as intended, including under unexpected conditions; tested and monitored. (Cue: autonomous system failure modes.)
- Privacy and security — data is protected, consented, and access-controlled throughout. (Cue: personal data in training or prompts.)
- Inclusiveness — the system works for people across abilities and contexts; accessibility. (Cue: speech interface for users who can't type.)
- Transparency — people understand what the system does and its limits; explainable outputs. (Cue: users should know they're talking to AI / why it decided.)
- Accountability — humans remain answerable for the system; governance and oversight. (Cue: who is responsible when the AI is wrong.)
Memory hook: F-R-P-I-T-A — "Fair, Reliable, Private, Inclusive, Transparent, Accountable." Note how cleanly these map onto vocabulary you already use: transparency ≈ your citations/explainability doctrine; accountability ≈ HITL and audit logging; reliability ≈ fail-closed design. You've been practicing the six — now you can name them on demand.
Classic fundamentals still on the exam
A quick checklist of older-style items that survived into AI-901 — all crosswalk, no new study: features and labels; training vs. validation data; overfitting; confusion-matrix intuition (true/false positives and negatives — you drilled precision/recall well past exam depth); when clustering beats classification (no labels). If any of those feel soft, the fix is Module 12 lesson 6 and Module 13 lesson 5, not new material.