Use this cheat sheet for Microsoft Certified: Azure AI Fundamentals (AI-901) when you need fundamentals-level recall without overbuilding the answer. AI-901-style preparation should separate workload recognition, responsible AI, basic implementation awareness, and Microsoft Foundry vocabulary.
AI-901 answer sequence
Use this when the stem mixes workload recognition, responsible AI, data readiness, or service fit.
flowchart TD
S["Scenario"] --> W["Classify the workload"]
W --> R["Check responsible AI or ML fit"]
R --> D["Check data readiness and source governance"]
D --> F["Pick the correct Azure service family"]
Fundamentals triage
- Identify the AI workload before naming a service.
- Decide whether the question is conceptual, responsible-AI, Azure-resource, or light implementation.
- Keep the answer fundamentals-level unless the stem clearly asks for engineering depth.
- Prefer service recognition and safe configuration over custom architecture.
- Do not treat generative AI as the answer to every AI scenario.
Workload chooser
| Scenario clue |
Likely workload |
| predict a category or value from historical examples |
machine learning |
| identify objects, faces, text, or visual features |
computer vision |
| transcribe, synthesize, or translate spoken language |
speech |
| classify sentiment, extract entities, translate text, or summarize language |
natural language processing |
| generate text, code, images, summaries, or conversational responses |
generative AI |
| extract fields from forms, invoices, contracts, or receipts |
document intelligence or content extraction |
| answer questions from company documents |
retrieval grounded generative AI |
Responsible AI map
| Principle |
Exam meaning |
| fairness |
reduce harmful bias or unequal treatment |
| reliability and safety |
make behavior dependable, tested, and safe for intended use |
| privacy and security |
protect data, identity, access, and sensitive outputs |
| inclusiveness |
design for diverse users and accessibility needs |
| transparency |
make system behavior, limitations, and AI involvement understandable |
| accountability |
assign human ownership for outcomes, monitoring, and correction |
Foundry and generative AI basics
| Term |
What to remember |
| model |
system that produces predictions, classifications, embeddings, or generated output |
| prompt |
instruction or input supplied to a generative model |
| system instruction |
higher-priority guidance that shapes assistant behavior |
| grounding |
connecting output to trusted external information |
| embedding |
vector representation used for semantic similarity and retrieval |
| agent |
AI workflow that can use instructions and tools to complete tasks |
| evaluation |
repeatable check of quality, safety, groundedness, relevance, or latency |
Service-fit instincts
| If the question asks for… |
Avoid this mistake |
| a deterministic field extraction |
do not default to open-ended chat output |
| safe AI deployment |
do not ignore content safety, privacy, and human oversight |
| basic Python/API use |
do not answer only with portal clicks if the stem mentions code |
| model output quality |
do not assume model size is the only lever |
| enterprise data answers |
do not ignore retrieval permissions and source freshness |
| fundamentals-level concept |
do not choose a complex custom ML pipeline unless required |
Machine learning basics
| Concept |
Fast recall |
| feature |
input variable used by a model |
| label |
target value predicted in supervised learning |
| training data |
examples used to learn model behavior |
| test data |
held-out data used to evaluate generalization |
| classification |
predicts a category |
| regression |
predicts a numeric value |
| clustering |
groups similar records without predefined labels |
| overfitting |
model performs well on training data but poorly on new data |
Python and Azure resource awareness
| Area |
Fundamentals-level expectation |
| variables and functions |
know simple code shape and what value is passed or returned |
| API or SDK call |
recognize endpoint, credential, request, response, and error handling basics |
| identity |
know that credentials and access should be managed securely |
| resource |
understand that Azure services are created, configured, secured, monitored, and billed |
| keys and secrets |
do not hard-code or expose them |
| monitoring |
know that deployed AI behavior needs logs, metrics, and review |
Common traps
| Trap |
Better instinct |
| Treating AI-901 as only old AI-900 service matching |
Expect more implementation, Foundry, and lightweight Python awareness. |
| Calling every workload generative AI |
Separate prediction, classification, extraction, vision, speech, NLP, and generation. |
| Skipping responsible AI |
Tie principles to concrete design choices and user impact. |
| Choosing advanced engineering too early |
Fundamentals questions usually reward clean recognition and safe defaults. |
| Forgetting data privacy |
Prompts, files, training examples, logs, and outputs can all contain sensitive data. |
Final 15-minute review
| If the stem says… |
Start here |
| fairness, privacy, transparency, accountability |
responsible AI principle and matching control |
| predict, classify, cluster, or train |
machine learning concept and data split |
| image, audio, document, language |
specialized AI workload family |
| prompt, assistant, generated answer |
generative AI, grounding, safety, and evaluation |
| Python, SDK, endpoint, key |
basic implementation flow and secure credential handling |
| company documents or knowledge base |
retrieval grounding and source permission awareness |
Practice fit
Use IT Mastery for the exact product route, practice status, spaced review when available, and close-answer explanation practice as coverage expands.
One-line decision rule
AI-901 answers should be simple but precise: identify the workload, apply responsible AI, recognize the Azure or Foundry building block, and avoid turning a fundamentals question into a professional architecture design.