AI-901 — Microsoft Azure AI Fundamentals Quick Review
Quick Review for Microsoft Azure AI Fundamentals (AI-901): high-yield AI workloads, Azure AI services, machine learning basics, responsible AI, and practice focus.
Quick Review Purpose
This Quick Review is for candidates preparing for the real Microsoft Azure AI Fundamentals (AI-901) exam from Microsoft. It focuses on the high-yield ideas you should be able to recognize quickly before moving into topic drills, mock exams, and detailed explanations.
AI-901 is a fundamentals exam. Expect conceptual and scenario-based questions more than deep implementation tasks. You should be able to:
- Match common AI workloads to the right Azure AI service.
- Distinguish machine learning, computer vision, natural language processing, speech, search, and generative AI scenarios.
- Recognize responsible AI considerations.
- Understand basic model training, evaluation, deployment, and inferencing language.
- Avoid confusing similar Azure services and similar AI concepts.
This page is IT Mastery review support and is not affiliated with Microsoft.
High-Yield Exam Map
| Area | What to know quickly | Common exam angle |
|---|---|---|
| AI workloads | Prediction, classification, anomaly detection, vision, NLP, speech, search, generative AI | “Which workload is this scenario?” |
| Responsible AI | Fairness, reliability and safety, privacy and security, inclusiveness, transparency, accountability | “Which principle is being addressed?” |
| Machine learning | Features, labels, training, validation, regression, classification, clustering, metrics | “Which ML type or metric applies?” |
| Azure Machine Learning | Workspaces, compute, data assets, automated ML, designer, pipelines, endpoints | “Which Azure ML capability helps?” |
| Computer vision | Image analysis, object detection, OCR, document extraction, custom vision models | “Which visual task or service?” |
| Natural language | Sentiment, key phrases, language detection, entity recognition, translation, Q&A, summarization | “Which text feature?” |
| Speech | Speech-to-text, text-to-speech, speech translation, speaker-related capabilities | “Which speech service task?” |
| Generative AI | Prompts, tokens, completions, embeddings, grounding, RAG, content safety | “How do you reduce hallucination or ground an answer?” |
| Search and knowledge mining | Indexing, enrichment, semantic/vector search, retrieval | “How do users search across documents?” |
Azure AI Service Selection Rules
Use this table to answer service-selection questions quickly.
| Scenario | Best fit to consider | Why |
|---|---|---|
| Build, train, evaluate, deploy ML models | Azure Machine Learning | End-to-end ML lifecycle platform |
| Non-experts want to try models from data | Automated ML in Azure Machine Learning | Automates model and algorithm selection |
| Drag-and-drop ML workflow | Azure Machine Learning designer | Visual pipeline construction |
| Predict a numeric value | Regression model | Output is continuous numeric |
| Predict a category | Classification model | Output is a class or label |
| Group similar records without labels | Clustering model | No known target label |
| Analyze images for tags/captions/objects | Azure AI Vision | General computer vision tasks |
| Extract printed/handwritten text from images | OCR / Read capability in Azure AI Vision | Converts image text to machine-readable text |
| Extract fields from invoices, receipts, forms | Azure AI Document Intelligence | Structured document extraction |
| Analyze sentiment or key phrases in text | Azure AI Language | Text analytics capabilities |
| Translate text between languages | Azure AI Translator | Text translation |
| Convert audio speech to text | Azure AI Speech | Speech recognition |
| Convert text to spoken audio | Azure AI Speech | Speech synthesis |
| Create a chatbot or conversational app | Azure Bot Service, Copilot-related tools, or language services | Conversation orchestration and language understanding |
| Add enterprise document search | Azure AI Search | Indexing and retrieval over content |
| Use large language models for chat or generation | Azure OpenAI Service / Azure AI model catalog capabilities | Generative AI model access |
| Ground generated answers in your documents | Azure AI Search plus generative AI | Retrieval-augmented generation pattern |
Service names and portal experiences can evolve. For exam readiness, focus on the capability being tested: vision, language, speech, search, machine learning, or generative AI.
Fast Decision Path
flowchart TD
A[What is the scenario asking?] --> B{Train custom predictive model?}
B -->|Yes| C[Azure Machine Learning]
B -->|No| D{Image, video, or document?}
D -->|Image analysis / objects / OCR| E[Azure AI Vision]
D -->|Forms, invoices, receipts| F[Azure AI Document Intelligence]
D -->|No| G{Text or language?}
G -->|Sentiment, entities, key phrases| H[Azure AI Language]
G -->|Translation| I[Azure AI Translator]
G -->|No| J{Speech audio?}
J -->|Speech-to-text or text-to-speech| K[Azure AI Speech]
J -->|No| L{Search across content?}
L -->|Yes| M[Azure AI Search]
L -->|No| N{Generate or summarize with an LLM?}
N -->|Yes| O[Generative AI / Azure OpenAI Service]
N -->|No| P[Re-read scenario for workload clues]
Responsible AI: Principles and Traps
Microsoft commonly describes responsible AI using principles such as fairness, reliability and safety, privacy and security, inclusiveness, transparency, and accountability. AI-901 questions often describe a business risk and ask which principle applies.
| Principle | What it means | Exam clue |
|---|---|---|
| Fairness | AI should avoid unfair bias or disparate impact | “Model performs worse for one demographic group” |
| Reliability and safety | AI should behave consistently and safely under expected conditions | “System must work reliably in critical scenarios” |
| Privacy and security | Data and models should be protected; personal data handled appropriately | “Sensitive data, access control, encryption, privacy” |
| Inclusiveness | AI should work for people with diverse needs and abilities | “Accessible to users with disabilities or different languages” |
| Transparency | People should understand system behavior, limitations, and use of AI | “Explain how the model makes decisions” |
| Accountability | Humans and organizations remain responsible for AI outcomes | “Who is responsible for monitoring and governance?” |
Common Responsible AI Mistakes
Fairness vs inclusiveness: Fairness is about avoiding unjust bias in outcomes. Inclusiveness is about designing so more people can use the system.
Transparency vs accountability: Transparency explains what the system does and its limitations. Accountability assigns responsibility for decisions, oversight, and remediation.
Privacy vs security: Privacy concerns appropriate use and protection of personal information. Security concerns protecting systems, data, credentials, and access.
Safety is not just cybersecurity: Reliability and safety also include predictable behavior, testing, monitoring, and harm reduction.
Machine Learning Fundamentals
Core Vocabulary
| Term | Meaning |
|---|---|
| Dataset | Collection of records used for training, validation, testing, or inference |
| Feature | Input variable used by the model |
| Label | Known target value the model learns to predict |
| Training | Process of fitting a model to data |
| Validation | Used during model development and tuning |
| Test data | Used to estimate final model performance on unseen data |
| Inference | Using a trained model to make predictions |
| Model | Learned function or algorithm that maps inputs to outputs |
| Experiment | A tracked ML run or set of runs |
| Endpoint | Deployed access point for model inferencing |
Supervised vs Unsupervised Learning
| Type | Uses labels? | Output | Example |
|---|---|---|---|
| Supervised learning | Yes | Prediction based on known examples | Predict loan approval from past decisions |
| Unsupervised learning | No | Patterns, clusters, associations | Group customers by behavior |
| Reinforcement learning | Uses rewards | Action policy | Optimize decisions through trial and feedback |
AI-901 usually emphasizes supervised and unsupervised learning more than advanced reinforcement learning.
Regression, Classification, and Clustering
| Task | Predicts | Example | Trap |
|---|---|---|---|
| Regression | Numeric value | Predict sales revenue, temperature, delivery time | Not every “prediction” is regression |
| Binary classification | One of two classes | Fraud/not fraud, approve/deny | Output is category, not number |
| Multiclass classification | One of many classes | Identify product category | Still classification |
| Multi-label classification | Multiple possible labels | Tag image with several attributes | More than one label can apply |
| Clustering | Groups similar items | Segment customers | No predefined label |
Evaluation Metrics to Recognize
| Metric | Used for | Meaning |
|---|---|---|
| Accuracy | Classification | Overall proportion of correct predictions |
| Precision | Classification | Of predicted positives, how many were truly positive |
| Recall | Classification | Of actual positives, how many were found |
| F1 score | Classification | Balance between precision and recall |
| AUC / ROC | Classification | Ability to separate classes across thresholds |
| MAE | Regression | Average absolute prediction error |
| RMSE | Regression | Penalizes larger regression errors more strongly |
| R-squared | Regression | Proportion of variance explained by the model |
Precision vs Recall Decision Rule
- Choose precision when false positives are costly. Example: incorrectly flagging legitimate transactions as fraud.
- Choose recall when false negatives are costly. Example: failing to detect a serious disease or safety issue.
- Use F1 score when you need a balance between precision and recall.
Overfitting and Underfitting
| Problem | What happens | Typical cause | Possible response |
|---|---|---|---|
| Overfitting | Performs well on training data but poorly on new data | Model learned noise or is too complex | More data, regularization, simpler model, better validation |
| Underfitting | Performs poorly even on training data | Model too simple or features insufficient | Better features, more suitable algorithm, more training |
Data Leakage
Data leakage occurs when training data includes information that would not be available at prediction time. It can make model performance look unrealistically good.
Common leakage examples:
- Using a future outcome as a feature.
- Training and testing on duplicate or near-duplicate records.
- Preprocessing the full dataset before splitting into train/test sets.
- Including fields that directly reveal the label.
Azure Machine Learning Review
Azure Machine Learning is the main Azure platform for building, training, managing, and deploying machine learning models.
| Capability | What it does |
|---|---|
| Workspace | Top-level resource for Azure ML assets and collaboration |
| Compute | Training or inference compute resources |
| Data assets | Registered references to datasets |
| Jobs / experiments | Run and track training work |
| Automated ML | Tries algorithms and settings to find a strong model |
| Designer | Visual drag-and-drop ML workflow tool |
| Pipelines | Repeatable workflows for data prep, training, and deployment |
| Model registry | Manage trained model versions |
| Managed endpoints | Deploy models for online or batch inference |
| Responsible AI tools | Help assess model behavior, explainability, and fairness |
Azure ML Scenario Traps
| If the scenario says… | Think… |
|---|---|
| “No-code or low-code model creation” | Automated ML or designer |
| “Data scientist needs full control with code” | Notebooks, SDK/CLI, custom training jobs |
| “Repeatable ML workflow” | Pipeline |
| “Track model runs and metrics” | Experiments/jobs in Azure ML |
| “Deploy model for real-time predictions” | Online endpoint |
| “Run predictions over many records asynchronously” | Batch endpoint |
| “Choose the best algorithm automatically” | Automated ML |
Computer Vision Review
Computer vision workloads process images, video, or visual documents.
| Workload | Description | Example |
|---|---|---|
| Image classification | Assigns a label to an entire image | “This image is a dog” |
| Object detection | Finds and locates objects in an image | Draw boxes around cars |
| OCR | Reads text from images | Extract text from a scanned page |
| Image analysis | Tags, captions, categories, objects, metadata | Describe image content |
| Face detection | Detects presence/location of human faces | Count faces in an image |
| Document extraction | Extracts structured fields from forms | Invoice total, vendor, date |
Vision Service Matching
| Scenario | Likely service/capability |
|---|---|
| Read text from a sign or scanned image | Azure AI Vision OCR / Read |
| Identify objects in photos | Azure AI Vision |
| Create a custom model to classify specific product images | Custom vision capability / custom model approach |
| Extract fields from receipts, invoices, tax forms, IDs, or business forms | Azure AI Document Intelligence |
| Analyze images with tags and captions | Azure AI Vision image analysis |
| Detect unsafe or restricted visual content | Content safety or moderation-related capability |
Common Vision Mistakes
Image classification vs object detection: Classification labels the whole image. Object detection identifies and locates objects within the image.
OCR vs document intelligence: OCR extracts text. Document intelligence extracts structured fields and layout from forms and documents.
Face detection vs face recognition: Detection finds faces. Recognition or identification compares identity. Be careful with wording and responsible AI implications.
Computer vision vs machine learning platform: Use Azure AI Vision for prebuilt vision tasks. Use Azure Machine Learning when the question is about building and managing custom ML models more generally.
Natural Language Processing Review
Natural language processing, or NLP, helps applications understand, analyze, translate, and generate human language.
| NLP task | What it does | Example |
|---|---|---|
| Language detection | Identifies language of text | Detect English, French, Spanish |
| Sentiment analysis | Determines positive, negative, neutral sentiment | Analyze customer reviews |
| Opinion mining | Finds sentiment about specific aspects | “Food was great, service was slow” |
| Key phrase extraction | Finds important phrases | Extract topics from feedback |
| Named entity recognition | Identifies people, places, organizations, dates, etc. | Extract company names |
| PII detection | Finds personally identifiable information | Detect names, emails, phone numbers |
| Summarization | Condenses long text | Summarize support tickets |
| Question answering | Answers user questions from content | FAQ bot |
| Conversational language understanding | Interprets intents and entities | “Book a flight tomorrow” |
NLP Service Matching
| Scenario | Likely service |
|---|---|
| Detect sentiment in customer comments | Azure AI Language |
| Extract key phrases from survey responses | Azure AI Language |
| Detect names, locations, organizations, or PII | Azure AI Language |
| Build an FAQ-style question answering experience | Azure AI Language question answering capability |
| Translate documents or messages | Azure AI Translator |
| Interpret user intent in a conversational app | Azure AI Language conversational understanding |
| Generate free-form text or chat responses | Generative AI model, such as through Azure OpenAI Service |
NLP Traps
- Translation is not the same as language detection. Translation converts text; detection identifies the language.
- Entity recognition is not sentiment analysis. Entity recognition extracts things; sentiment measures attitude.
- Question answering is not general search. Q&A returns targeted answers from a knowledge source; search retrieves ranked documents or passages.
- NLP and generative AI overlap but are not identical. Traditional NLP can classify or extract. Generative AI can create new text, summarize, or answer conversationally.
Speech Review
Speech workloads process spoken audio or generate speech.
| Capability | Direction | Example |
|---|---|---|
| Speech recognition | Speech to text | Transcribe a meeting |
| Speech synthesis | Text to speech | Read a response aloud |
| Speech translation | Speech in one language to text/audio in another | Live multilingual captions |
| Speaker-related features | Analyze speaker characteristics depending on capability | Distinguish speakers or voices in supported scenarios |
Common Speech Mistakes
- Speech-to-text transcribes audio into text.
- Text-to-speech creates spoken audio from text.
- Translation changes language; transcription alone does not.
- NLP begins after text exists. If the source is audio, speech recognition may be needed before text analytics.
Generative AI Review
Generative AI creates new content such as text, code, images, summaries, or answers. Large language models, or LLMs, are a major generative AI category.
Key Concepts
| Concept | Meaning |
|---|---|
| Prompt | Input instruction or question given to the model |
| Completion / response | Model-generated output |
| Token | Unit of text processed by the model |
| Context window | Amount of text the model can consider at once |
| System message | High-priority instruction that shapes assistant behavior |
| Temperature | Controls randomness or creativity of responses |
| Grounding | Connecting responses to trusted data sources |
| RAG | Retrieval-augmented generation: retrieve relevant data, then generate answer |
| Embedding | Numeric representation of text or data for similarity search |
| Vector search | Finds semantically similar content using embeddings |
| Fine-tuning | Further training a model for a specific style or task |
| Content filtering | Detects or blocks harmful or policy-violating content |
Prompting Patterns
| Pattern | Use when… | Example |
|---|---|---|
| Direct instruction | The task is simple | “Summarize this paragraph in three bullets.” |
| Role prompting | You want style or perspective | “Act as a support agent.” |
| Few-shot prompting | You want a format or pattern | Provide examples, then ask for another |
| Grounded prompting | Accuracy depends on source material | “Answer only from the provided policy.” |
| Structured output | App needs parseable output | “Return JSON with these fields.” |
RAG vs Fine-Tuning vs Prompt Engineering
| Approach | Best for | Not best for |
|---|---|---|
| Prompt engineering | Better instructions, formatting, constraints | Adding large amounts of private knowledge |
| RAG | Grounding answers in current enterprise content | Teaching a model a new writing style by examples alone |
| Fine-tuning | Consistent style, domain-specific patterns, specialized behavior | Frequently changing facts or documents |
| Content filtering | Reducing harmful outputs | Guaranteeing factual correctness by itself |
Generative AI Traps
- Hallucination means the model may produce plausible but incorrect output.
- Grounding reduces hallucination risk by providing trusted context, but does not eliminate all risk.
- Embeddings do not generate answers. They represent content for similarity comparison.
- Vector search is not the final answer generator. It retrieves relevant content; an LLM may then generate a response from that content.
- Fine-tuning is not the default answer for adding current company documents. RAG is often the better pattern when facts change.
- Temperature affects variation, not truthfulness by itself.
- Content filters support safety, but human review and governance may still be required.
Azure AI Search and Knowledge Retrieval
Azure AI Search helps create searchable indexes over content. It can support traditional keyword search, semantic ranking, and vector-based retrieval patterns used in generative AI solutions.
| Concept | Meaning |
|---|---|
| Index | Searchable structure containing fields from source content |
| Indexer | Pulls data from a source into an index |
| Skillset | Enrichment steps such as OCR, entity extraction, or language processing |
| Analyzer | Processes text for search, such as tokenization |
| Semantic search | Improves relevance using language understanding |
| Vector search | Searches by similarity using embeddings |
| RAG | Uses retrieval results as grounding context for generation |
Search Traps
- Search is retrieval, not training. Indexing documents does not train a model in the same way ML training does.
- Keyword search and vector search solve different retrieval problems. Keyword search matches terms; vector search matches semantic similarity.
- OCR may be needed before search if source documents are scanned images.
- Azure AI Search can be part of a generative AI architecture but is not itself the LLM.
Conversational AI Review
Conversational AI allows users to interact through natural language, often in chat or voice interfaces.
| Component | Role |
|---|---|
| Bot channel | Where users interact, such as web chat or messaging |
| Conversation logic | Controls flow, prompts, and responses |
| Language understanding | Detects user intent and entities |
| Knowledge source | Provides FAQ or document-based answers |
| Backend integration | Connects to business systems |
| Generative model | May generate natural responses or summaries |
Conversational AI Traps
- Intent is what the user wants to do. Example: “BookFlight.”
- Entity is a detail needed to complete the task. Example: destination, date, passenger count.
- A bot is not automatically intelligent. It may use rules, language understanding, retrieval, generative AI, or a combination.
- Generative chat still needs grounding and safety controls for enterprise use.
Security, Privacy, and Governance Basics
AI-901 is not a deep security exam, but you should recognize basic cloud and AI governance concerns.
| Concern | What to look for |
|---|---|
| Authentication | Who or what can access the service |
| Authorization | What actions are allowed |
| Data privacy | Handling personal or sensitive information properly |
| Encryption | Protecting data at rest and in transit |
| Monitoring | Tracking behavior, performance, and errors |
| Human oversight | Reviewing high-impact or uncertain decisions |
| Model drift | Model performance changes as real-world data changes |
| Auditability | Ability to review decisions, data, and system activity |
Common Governance Mistakes
- Treating a model score as a final decision without human oversight in sensitive contexts.
- Ignoring monitoring after deployment.
- Assuming training performance will remain stable forever.
- Sending sensitive data to a model without considering privacy, access, retention, and compliance requirements.
- Failing to document known limitations.
Exam-Style Scenario Clues
Use these clues to move quickly in practice questions.
| Wording in question | Likely answer direction |
|---|---|
| “Predict next month’s revenue” | Regression |
| “Approve or reject an application” | Binary classification |
| “Classify support tickets by category” | Multiclass classification |
| “Find unusual transactions” | Anomaly detection |
| “Group customers by similar behavior” | Clustering |
| “Extract text from scanned images” | OCR |
| “Extract invoice total and vendor name” | Document intelligence |
| “Identify sentiment in customer reviews” | Language sentiment analysis |
| “Translate a web page” | Translator |
| “Transcribe phone calls” | Speech-to-text |
| “Read a response aloud” | Text-to-speech |
| “Search across product manuals” | Azure AI Search |
| “Answer questions using company documents” | RAG with search plus generative AI |
| “Generate draft email replies” | Generative AI |
| “Explain model decisions” | Transparency / explainability |
| “Ensure system works for users with disabilities” | Inclusiveness |
| “Reduce biased outcomes between groups” | Fairness |
| “Protect personal data” | Privacy and security |
Common Candidate Mistakes
Mistake 1: Choosing Azure Machine Learning for Every AI Scenario
Azure Machine Learning is for building and managing ML models. Many AI-901 scenarios are better answered by prebuilt Azure AI services.
| If the task is… | Prefer… |
|---|---|
| Sentiment analysis | Azure AI Language |
| Speech transcription | Azure AI Speech |
| Image tagging | Azure AI Vision |
| Invoice extraction | Azure AI Document Intelligence |
| Text translation | Azure AI Translator |
| Custom predictive modeling | Azure Machine Learning |
Mistake 2: Confusing Model Training with Inferencing
| Activity | Meaning |
|---|---|
| Training | Model learns from historical data |
| Evaluation | Model performance is measured |
| Deployment | Model is made available for use |
| Inferencing | Model makes predictions on new data |
| Monitoring | Performance and behavior are tracked after deployment |
Mistake 3: Ignoring Labels
If the scenario has known historical outcomes, it is likely supervised learning. If the scenario asks to discover natural groupings without known target values, it is clustering.
Mistake 4: Treating Accuracy as Always Best
Accuracy can be misleading with imbalanced classes. If only 1% of transactions are fraudulent, a model that always predicts “not fraud” can appear highly accurate but be useless for fraud detection.
Mistake 5: Using Generative AI Without Grounding
For enterprise answers based on policy, manuals, or internal documents, think about grounding and retrieval. A model’s general knowledge may be incomplete or outdated.
Mini Review Tables
Workload to Data Type
| Data type | Likely workload |
|---|---|
| Tabular rows and columns | Machine learning |
| Images or video frames | Computer vision |
| Scanned forms or invoices | Document intelligence |
| Written text | NLP or generative AI |
| Spoken audio | Speech |
| Large document collections | Search / knowledge mining |
| Prompts and generated responses | Generative AI |
Output Type to Model Type
| Output | Model type |
|---|---|
| Number | Regression |
| Yes/no | Binary classification |
| One category from many | Multiclass classification |
| Multiple labels | Multi-label classification |
| Group assignment without known labels | Clustering |
| Unusual/not unusual | Anomaly detection |
Risk to Responsible AI Principle
| Risk | Principle |
|---|---|
| Discriminatory outcomes | Fairness |
| Unpredictable behavior in production | Reliability and safety |
| Exposure of sensitive data | Privacy and security |
| Inaccessible design | Inclusiveness |
| Users do not know AI is involved | Transparency |
| No owner for system decisions | Accountability |
How to Use This Quick Review with Practice
A good AI-901 study loop is:
- Review the decision tables above.
- Complete topic drills by domain: ML, vision, language, speech, search, responsible AI, and generative AI.
- For every missed question, write down the clue you overlooked.
- Re-answer similar original practice questions until the service-selection pattern is automatic.
- Use mock exams to check timing and mixed-topic recognition.
- Read detailed explanations even for correct answers when you guessed.
When reviewing practice questions, focus less on memorizing wording and more on identifying the workload, data type, output type, and responsible AI concern.
Final Rapid Checklist
Before sitting for Microsoft Azure AI Fundamentals (AI-901), make sure you can answer these without hesitation:
- Can I distinguish regression, classification, clustering, and anomaly detection?
- Can I explain features, labels, training, testing, deployment, and inferencing?
- Can I match vision, language, speech, search, and generative AI scenarios to Azure services?
- Can I identify responsible AI principles from scenario wording?
- Can I explain why RAG helps ground generative AI responses?
- Can I distinguish OCR from document intelligence?
- Can I distinguish sentiment analysis, entity recognition, key phrase extraction, and translation?
- Can I choose between real-time and batch inference?
- Can I spot overfitting, underfitting, data leakage, and metric traps?
- Can I read a scenario and identify the main data type and desired output?
Practical Next Step
Use this Quick Review as your final concept pass, then move into IT Mastery practice: start with targeted topic drills, answer original practice questions by workload area, and use detailed explanations to close any gaps before attempting full mock exams.
Continue in IT Mastery
Use this Quick Review as a final concept map, then move into IT Mastery for focused topic drills, mixed practice sets, timed mock exams, and detailed explanations. The practice questions are original IT Mastery practice items; they are not official Microsoft questions, copied live-exam content, or exam dumps.