Master the AI-900

An interactive study guide built on 7 memory techniques to help you pass the Microsoft Azure AI Fundamentals exam.

AI Workloads 15-20%Machine Learning 20-25%Computer Vision 15-20%NLP 15-20%Generative AI 15-20%

What it covers

AI concepts, responsible AI principles, machine learning (regression, classification, clustering), Azure AI Vision, Azure AI Language, Azure OpenAI Service, and Microsoft Copilot.

Ideal for

IT professionals, business decision makers, students, and anyone exploring artificial intelligence and machine learning on Azure.

Aspire to this if

You're curious about AI, want to validate your understanding of Azure AI services, or plan to pursue associate-level AI or data science certifications.

Section 1 / Spatial Memory

The Map

Tap any component to see what you need to know for the exam.

🤖 AI Workloads

Prediction · Anomaly · Vision · NLP

⚖️ Responsible AI

Fairness · Reliability · Privacy

🧠 ML Fundamentals

Train · Validate · Evaluate

📈 Regression

Predict Numbers

🏷️ Classification

Predict Categories

🔵 Clustering

Find Groups

👁️ Computer Vision

Azure AI Vision · Custom Vision · Face

💬 Natural Language Processing

Azure AI Language · Speech · Translator

✨ Generative AI

Azure OpenAI · GPT · DALL-E · Copilot

☁️ Azure AI Services

APIs · Studio · Deployment

🔬 Azure Machine Learning

AutoML · Designer · Notebooks

🛡️ AI Safety & Governance

Content Filters · Grounding · Red Teaming

Section 2 / Narrative Memory

The Story

Follow the narrative to build a mental model of how everything connects.

🤖

The Spark of Intelligence

Artificial intelligence begins with a simple idea — machines that mimic human cognition. AI workloads span prediction (what will happen?), anomaly detection (what’s unusual?), computer vision (what do I see?), NLP (what does this mean?), and generative AI (create something new). Each workload maps to specific Azure services.

Exam IntelAI workload types: prediction, anomaly detection, computer vision, NLP, conversational AI, generative AI. Know which Azure service handles each. AI services are consumed as APIs, not built from scratch.

⚖️

The Ethical Compass

Microsoft’s six Responsible AI principles guide every AI deployment. Fairness ensures no group is disadvantaged. Reliability demands consistent, safe performance. Privacy protects personal data. Inclusiveness designs for everyone. Transparency explains decisions. Accountability keeps humans in control.

Exam IntelMemorise all 6 principles: Fairness, Reliability & Safety, Privacy & Security, Inclusiveness, Transparency, Accountability. Expect questions that describe a scenario and ask which principle applies. Accountability = humans are responsible, not the AI.

🧠

The Learning Machine

Machine learning is the engine powering AI. Models learn from data through training, validation, and testing. Supervised learning uses labelled data (regression, classification). Unsupervised learning discovers hidden patterns (clustering). The features are what goes in; the label is what comes out.

Exam IntelSupervised = labelled data (regression + classification). Unsupervised = no labels (clustering). Features = input columns. Label = predicted value. Training/validation/test split prevents overfitting. Know the difference between the three ML types.

📈

The Numbers Game

Regression predicts continuous values — house prices, temperatures, revenue. Classification sorts into categories — spam vs. legitimate, approved vs. denied. Clustering groups unlabelled data into natural segments. Each approach uses different algorithms and different evaluation metrics.

Exam IntelRegression metrics: MAE, RMSE, R². Classification metrics: accuracy, precision, recall, F1, AUC. Clustering: K-Means assigns to K groups. Confusion matrix: TP, FP, TN, FN. Binary classification = 2 classes, multi-class = 3+.

👁️

The Eyes of AI

Computer vision gives machines the power of sight. Azure AI Vision analyses images, extracts text via OCR, and generates captions. Custom Vision lets you train classifiers and object detectors with just a handful of images. Face API detects, identifies, and verifies faces — with strict ethical guidelines.

Exam IntelImage classification = label the whole image. Object detection = bounding boxes + labels. OCR = extract printed/handwritten text. Custom Vision: upload images → tag → train → predict. Face API: detection (where), identification (who), verification (confirm).

💬

The Language Bridge

Natural language processing helps AI understand and generate human language. Azure AI Language extracts sentiment, key phrases, and named entities. Azure AI Speech converts spoken words to text and back again. Conversational Language Understanding (CLU) powers chatbots by recognising intents and entities.

Exam IntelNLP capabilities: sentiment analysis, key phrases, NER, PII detection, language detection, question answering. CLU: intents (what the user wants) + entities (key details). Speech: STT, TTS, translation. Translator: 100+ languages, real-time.

✨

The Generative Revolution

Generative AI creates new content — text, images, code. Azure OpenAI Service provides enterprise access to GPT models, DALL-E for images, and Whisper for speech. Prompt engineering shapes the output. RAG (Retrieval-Augmented Generation) grounds responses in your own data to reduce hallucinations.

Exam IntelLLMs predict the next token. Temperature controls randomness (low = deterministic, high = creative). Prompt engineering: system message, few-shot examples, clear instructions. RAG = search your data + feed to LLM. DALL-E generates images from text prompts.

☁️

The Azure AI Toolbox

Azure AI services are the building blocks. Pre-built APIs handle common tasks (vision, language, speech). Azure AI Studio unifies the experience for building custom AI. Azure Machine Learning provides the full pipeline — AutoML for quick starts, Designer for no-code, and notebooks for data scientists.

Exam IntelAzure AI services = pre-built APIs (key + endpoint). Azure Machine Learning = full ML platform. AutoML = automated algorithm selection. Designer = drag-and-drop. AI Studio = unified portal for generative AI. All secured with keys or Microsoft Entra ID.

🤝

The Copilot Era

Microsoft Copilot brings AI to every product. Built on top of large language models, grounded in your organisational data via Microsoft Graph. Copilot in Microsoft 365 helps write documents, summarise emails, and create presentations. It’s the consumer-facing expression of the AI platform underneath.

Exam IntelCopilot = LLM + Microsoft Graph + your data. Microsoft 365 Copilot: Word, Excel, PowerPoint, Teams, Outlook. Bing Chat (Copilot): web search + LLM. Copilot Studio: build custom copilots. Grounding in organisational data is the key differentiator.

🛡️

The Safety Net

AI must be safe before it can be trusted. Azure AI Content Safety detects harmful text and images across four categories. Content filters in Azure OpenAI apply configurable severity thresholds. Red teaming stress-tests models. The Responsible AI dashboard provides fairness analysis, error analysis, and model interpretability.

Exam IntelContent Safety: 4 categories — violence, hate, sexual, self-harm. Each rated severity 0–6. Azure OpenAI content filters are on by default. Grounding reduces hallucinations. Red teaming = adversarial testing. Responsible AI dashboard = fairness + error analysis + interpretability.

Section 3 / Acronym Memory

Mnemonic Wall

Memorable acronyms and phrases to anchor key exam concepts in your memory.

⚖️

FRIPTA

Microsoft’s 6 Responsible AI principles. Expect scenario questions — match the described issue to the correct principle.

🧠

SUS

Three ML paradigms. Supervised = labelled data. Unsupervised = no labels. Reinforcement = reward-based. AI-900 focuses on supervised and unsupervised.

📈

RCC

Three core ML task types. Regression = numbers. Classification = categories. Clustering = groups. Know which is supervised vs unsupervised.

📊

MAR

Regression metrics: MAE (average error), RMSE (penalises large errors), R² (0–1, closer to 1 = better fit).

🏷️

PRFA

Classification metrics. Precision = correct positives / all predicted positives. Recall = correct positives / all actual positives. F1 = harmonic mean. AUC = area under ROC curve.

👁️

COF

Three Azure AI Vision capabilities. Classification = label whole image. Object detection = bounding boxes. Face = detect/identify/verify.

💬

SKEP

Core Azure AI Language capabilities. Also includes language detection and question answering. NER recognises people, places, organisations.

🎤

STT-TTS

Azure AI Speech bidirectional capabilities. STT = transcription. TTS = synthesis. Also supports real-time speech translation.

🤖

CLU

Build chatbots by defining intents (what user wants) and entities (key details). Replaces LUIS. Trained with example utterances.

✨

GDW

Azure OpenAI models. GPT = text generation. DALL-E = image generation. Whisper = speech-to-text. All accessed via Azure OpenAI Service.

📝

RAG

Ground LLM responses in your own data. Search relevant docs → inject into prompt → generate grounded answer. Reduces hallucinations.

🎯

TEMP

Temperature parameter: low (0) = deterministic, focused. High (1+) = creative, random. Controls randomness of token selection.

🛡️

VHSS

Four Azure AI Content Safety categories. Each rated by severity (0–6). Content filters on Azure OpenAI are enabled by default.

🔬

ADN

Three ways to build ML in Azure Machine Learning. AutoML = automated. Designer = drag-and-drop. Notebooks = code-first (Python/R).

🤝

LMG

How Copilot works: LLM for reasoning, Microsoft Graph for organisational data, grounding for accuracy. This triad powers all Copilot experiences.

Section 4 / Contrast Memory

Versus Arena

Side-by-side comparisons to sharpen your understanding of similar concepts.

RegressionvsClassification

Regression vs Classification

Aspect	Regression	Classification
Predicts	Continuous number	Category/class label
Output	Price, temperature, score	Spam/not spam, yes/no
Learning	Supervised	Supervised
Metrics	MAE, RMSE, R²	Accuracy, Precision, Recall, F1, AUC
Algorithms	Linear regression, decision forest	Logistic regression, decision tree
Example	Predict house price	Predict loan approval (yes/no)

SupervisedvsUnsupervised

Supervised vs Unsupervised

Aspect	Supervised	Unsupervised
Labels	Requires labelled data	No labels needed
Goal	Predict outcomes	Discover patterns/groups
Types	Regression, classification	Clustering
Example	Predict churn (yes/no)	Segment customers into groups
Training data	Features + known label	Features only
Evaluation	Compare prediction vs actual	Cluster quality metrics

Image ClassificationvsObject Detection

Image Classification vs Object Detection

Aspect	Image Classification	Object Detection
Output	Single label for whole image	Bounding boxes + labels
Answers	“What is this image?”	“What objects are where?”
Count	One label per image	Multiple objects detected
Use case	Photo categorisation	Autonomous driving, retail
Azure service	Custom Vision (classifier)	Custom Vision (detector)
Complexity	Simpler	More complex

Azure AI ServicesvsAzure ML

Azure AI Services vs Azure ML

Aspect	Azure AI Services	Azure ML
Approach	Pre-built APIs	Build custom models
Skill level	Developer (no ML expertise)	Data scientist / ML engineer
Customisation	Limited (some fine-tuning)	Full control
Setup time	Minutes (API call)	Hours to weeks (training)
Examples	Vision, Language, Speech APIs	AutoML, Designer, Notebooks
Best for	Common AI tasks	Unique/complex ML problems

GPTvsDALL-EvsWhisper

GPT vs DALL-E vs Whisper

Aspect	GPT	DALL-E
Modality	Text generation	Image generation \| Speech-to-text
Input	Text prompt	Text prompt \| Audio file
Output	Text/code	Images \| Transcribed text
Use case	Chat, summarise, code	Create visuals \| Transcribe meetings
Temperature	Controls randomness	N/A \| N/A
Azure service	Azure OpenAI	Azure OpenAI \| Azure OpenAI

PrecisionvsRecall

Precision vs Recall

Aspect	Precision	Recall
Question	How many positives are correct?	How many actual positives found?
Formula	TP / (TP + FP)	TP / (TP + FN)
High when	Few false positives	Few false negatives
Prioritise when	False alarm is costly (spam filter)	Missing cases is costly (disease)
Trade-off	Higher precision → lower recall	Higher recall → lower precision
Balance	F1 score = harmonic mean of both	F1 score = harmonic mean of both

CLUvsQuestion Answering

CLU vs Question Answering

Aspect	CLU	Question Answering
Purpose	Understand user intent	Answer from knowledge base
Input	User utterance	User question
Output	Intent + entities	Best matching answer
Training	Example utterances	FAQ pairs / documents
Use case	Chatbot command routing	FAQ bot, help desk
Azure feature	Conversational Language Understanding	Custom Question Answering

AutoMLvsDesignervsNotebooks

AutoML vs Designer vs Notebooks

Aspect	AutoML	Designer
Approach	Automated \| No-code	Code-first
Audience	Business users \| Citizen devs	Data scientists
Customisation	Algorithm auto-selected \| Drag-and-drop	Full control
Speed	Fastest \| Moderate	Slowest (but most flexible)
Output	Best model selected \| Pipeline	Custom pipeline
Language	None \| None	Python / R

Content SafetyvsResponsible AI

Content Safety vs Responsible AI Principles

Aspect	Content Safety	Responsible AI
What	Technical service / API	Guiding principles
Scope	Detect harmful content	Govern all AI systems
Categories	Violence, hate, sexual, self-harm	Fairness, reliability, privacy, inclusiveness, transparency, accountability
Applied	Runtime content filtering	Design-time decisions
Automation	Automated severity scoring	Human-led assessment
Azure tool	Azure AI Content Safety	Responsible AI dashboard

Section 5 / Grouping Memory

Cheat Sheet

Organised reference grouped by exam domain — everything you need on one page.

AI Workloads & Responsible AI

15-20%

AI Workload Types

Prediction: forecast future values (regression, classification)
Anomaly detection: identify unusual patterns
Computer vision: analyse images and video
NLP: understand and generate human language
Conversational AI: chatbots and virtual assistants
Generative AI: create new content (text, images, code)

Responsible AI Principles

Fairness: treat all people equitably, avoid bias
Reliability & Safety: perform consistently and safely
Privacy & Security: protect personal data
Inclusiveness: empower everyone, design for all
Transparency: explain how AI makes decisions
Accountability: humans responsible for AI outcomes

Key Concepts

AI = machines mimicking human cognitive functions
ML = subset of AI, learning from data
Deep learning = subset of ML, neural networks
Azure AI services = pre-built APIs for common tasks

Machine Learning Fundamentals

20-25%

ML Types

Supervised: labelled data (regression + classification)
Unsupervised: no labels (clustering)
Features = input variables, Label = target to predict
Training/validation/test split prevents overfitting

Regression

Predicts continuous numeric values
Algorithms: linear regression, decision forest
Metrics: MAE, RMSE, R² (closer to 1 = better)

Classification

Predicts categories (binary or multi-class)
Algorithms: logistic regression, decision tree, neural network
Metrics: accuracy, precision, recall, F1, AUC-ROC
Confusion matrix: TP, FP, TN, FN

Clustering

Groups similar data without labels
K-Means: assign data to K clusters
Use cases: customer segmentation, anomaly detection

Azure Machine Learning

AutoML: automatic algorithm and hyperparameter selection
Designer: drag-and-drop pipeline builder
Notebooks: code-first with Python/R
MLflow for experiment tracking and model registry

Computer Vision

15-20%

Azure AI Vision

Image analysis: tags, captions, categories, objects
OCR: extract printed and handwritten text
Spatial analysis: people counting, movement tracking
Smart cropping: intelligent image thumbnails

Custom Vision

Image classification: label the whole image
Object detection: bounding boxes + labels
Train with small datasets (as few as 5 images per tag)
Export models to edge devices (ONNX, TensorFlow)

Face API

Detection: locate faces, return attributes
Identification: match face to known person
Verification: confirm two faces are same person
Strict ethical guidelines and limited access policy

Natural Language Processing

15-20%

Azure AI Language

Sentiment analysis: positive, negative, neutral, mixed
Key phrase extraction: main talking points
Named Entity Recognition (NER): people, places, organisations
PII detection: identify and redact personal info
Language detection: identify the language of text

Azure AI Speech

Speech-to-text (STT): transcription
Text-to-speech (TTS): synthesis with neural voices
Speech translation: real-time spoken language translation
Speaker recognition: identify who is speaking

Conversational AI

CLU: define intents (goals) and entities (details)
Question Answering: knowledge base from FAQs/docs
Azure Bot Service: build and deploy chatbots
Replaces LUIS and QnA Maker (now in Azure AI Language)

Generative AI

15-20%

Azure OpenAI Service

GPT models: text and code generation
DALL-E: image generation from text prompts
Whisper: speech-to-text transcription
Enterprise security, compliance, and regional deployment

Prompt Engineering

System message: sets behaviour and constraints
Few-shot examples: show desired input/output pairs
Temperature: 0 = deterministic, 1+ = creative
Top-p: nucleus sampling, alternative to temperature

RAG & Grounding

RAG: retrieve relevant docs → inject into prompt → generate
Azure AI Search: vector + semantic search for retrieval
Grounding reduces hallucinations by anchoring to real data
On Your Data feature in Azure OpenAI for easy RAG setup

Microsoft Copilot

LLM + Microsoft Graph + grounding = Copilot
Microsoft 365 Copilot: Word, Excel, PowerPoint, Teams
Copilot Studio: build custom copilots (low-code)
Bing Chat (Copilot): web search + LLM answers

Section 6 / Method of Loci

The Memory Palace

Walk through themed rooms — each object anchors a concept in spatial memory.

The Principles Lobby

AI Workloads & Responsible AI — Where it all begins

🤖

AI Workloads

Prediction, anomaly detection, vision, NLP, conversational AI, generative AI

⚖️

Fairness

Treat all people equitably. Avoid bias in data and outcomes

🛡️

Reliability & Safety

AI must perform consistently and handle errors gracefully

🔒

Privacy & Security

Protect personal data. Secure models and endpoints

👁️

Transparency

Users should understand how AI makes decisions. Explain outputs

🤝

Accountability

Humans are responsible for AI outcomes — not the machine

The Training Lab

Machine Learning — Where models learn

📈

Regression

Predict numbers. Linear regression. Metrics: MAE, RMSE, R²

🏷️

Classification

Predict categories. Logistic regression. Metrics: precision, recall, F1, AUC

🔵

Clustering

Find groups. K-Means. Unsupervised — no labels needed

⚙️

AutoML

Automated algorithm selection and hyperparameter tuning

🔧

Designer

Drag-and-drop pipeline builder. No code required

💻

Notebooks

Jupyter notebooks for code-first ML. Python and R supported

The Vision Gallery

Computer Vision — Where machines see

🖼️

Image Analysis

Tags, captions, categories, objects. Azure AI Vision service

📄

OCR

Extract printed and handwritten text from images and PDFs

🎯

Custom Vision

Train classifiers and detectors with small datasets. Export to edge

👤

Face API

Detection (where), identification (who), verification (confirm match)

🔲

Object Detection

Bounding boxes with labels and confidence scores. Multiple objects per image

📷

Spatial Analysis

People counting, movement tracking, zone monitoring from video

The Language Library

NLP & Speech — Where words have meaning

💬

Sentiment Analysis

Positive, negative, neutral, mixed. Scored per sentence and document

🔑

Entity Recognition

Identify people, places, organisations, dates, quantities in text

🎤

Speech Services

STT (transcribe), TTS (synthesise), translation, speaker recognition

🤖

CLU

Intents = what user wants. Entities = key details. Train with utterances

❓

Question Answering

Build knowledge bases from FAQs and documents. Returns best match

🌐

Translator

Real-time text translation across 100+ languages. Custom models available

The Generation Chamber

Generative AI & Copilot — Where AI creates

✨

GPT Models

Text generation by predicting next token. Temperature controls randomness

🎨

DALL-E

Generate images from text prompts. Variations and edits supported

📝

Prompt Engineering

System message + few-shot examples + clear instructions = better output

🔍

RAG Pattern

Retrieve docs → inject into prompt → generate grounded answer. Reduces hallucinations

🤝

Microsoft Copilot

LLM + Microsoft Graph + grounding. Built into M365, Bing, Azure

🛡️

Content Safety

Violence, hate, sexual, self-harm filtering. Severity 0–6. On by default in Azure OpenAI

Section 7 / Pattern Recognition

Pattern Spotter

Decision trees and trigger-answer pairs — see the pattern, know the answer.

Which ML Task Type?

Which ML Task Type?
  ├── Predict a number (price, temperature) → Regression (supervised)
  ├── Predict a category (yes/no, spam/legit) → Classification (supervised)
  ├── Group similar items (no labels) → Clustering (unsupervised)
  ├── Detect unusual patterns in data → Anomaly detection
  ├── Generate new text, images, or code → Generative AI
  ├── Understand or process human language → NLP workload
  └── Analyse or interpret images/video → Computer vision workload

Which Azure AI Service?

Which Azure AI Service?
  ├── Analyse images, generate captions, OCR → Azure AI Vision
  ├── Train custom image classifier/detector → Custom Vision
  ├── Detect/identify faces → Azure AI Face
  ├── Sentiment, entities, key phrases → Azure AI Language
  ├── Speech-to-text or text-to-speech → Azure AI Speech
  ├── Translate text across languages → Azure AI Translator
  ├── Generate text, code, or chat → Azure OpenAI Service (GPT)
  └── Generate images from prompts → Azure OpenAI Service (DALL-E)

Which Responsible AI Principle?

Which Responsible AI Principle?
  ├── Model disadvantages a specific group → Fairness
  ├── System behaves unpredictably or unsafely → Reliability & Safety
  ├── Personal data exposed or misused → Privacy & Security
  ├── System excludes certain users → Inclusiveness
  ├── Users can’t understand how AI decided → Transparency
  └── No one owns the AI’s outcomes → Accountability

How to Build an ML Model in Azure?

How to Build an ML Model in Azure?
  ├── Fastest path, no ML expertise → AutoML (automated algorithm selection)
  ├── Visual pipeline, no code → Azure ML Designer (drag-and-drop)
  ├── Full control, custom code → Azure ML Notebooks (Python/R)
  ├── Common AI task (no custom model needed) → Azure AI Services (pre-built APIs)
  └── Generative AI / LLM application → Azure OpenAI Service + AI Studio

Decision Cards

“predict a numeric value” or “continuous output”→Regression

“predict a category” or “binary/multi-class”→Classification

“group similar items” or “no labelled data”→Clustering (K-Means)

“confusion matrix” or “TP/FP/TN/FN”→Classification evaluation

“R-squared” or “RMSE” or “MAE”→Regression evaluation metrics

“bounding boxes” on images→Object detection (Custom Vision)

“OCR” or “extract text from images”→Azure AI Vision (Read API)

“sentiment analysis” or “key phrases”→Azure AI Language

“intents and entities” for a chatbot→Conversational Language Understanding (CLU)

“speech-to-text” or “text-to-speech”→Azure AI Speech

“temperature” parameter for text generation→Controls LLM randomness (low = focused, high = creative)

“RAG” or “retrieval-augmented generation”→Search your data + inject into prompt + generate grounded answer

“prompt engineering” or “system message”→Optimise LLM output via instructions, examples, constraints

“Copilot” or “Microsoft 365 AI assistant”→LLM + Microsoft Graph + grounding

“content safety” or “harmful content detection”→Azure AI Content Safety (violence, hate, sexual, self-harm)

“treats all people equitably” or “bias in AI”→Fairness (Responsible AI principle)

“AutoML” or “automatic algorithm selection”→Azure Machine Learning AutoML

“drag-and-drop ML pipeline”→Azure Machine Learning Designer

Ready to certify?

Train with practitioners, not presenters

Lucid Labs delivers Microsoft certification training led by Microsoft Certified Trainers (MCTs) and grounded in real-world project experience. We adapt every session to your team's environment, data stack, and business objectives — because the best exam prep comes from engineers who build these solutions every day.

🎯

Tailored Content

Training built around your actual AI use cases and business goals — not generic slides.

🛠️

Hands-On Labs

Work through real Azure AI scenarios with expert guidance — Vision, Language, OpenAI, and more.

📈

Exam + Capability

Pass the exam and build lasting AI literacy your team can apply from day one.

Talk to us about Azure AI Fundamentals training

Custom training for teams & individuals — remote or on-site across Australia

Keith Oak

Chief Technology Officer — Lucid Labs

Microsoft Solutions Partner architect specialising in Fabric, Azure Data & AI, and GitHub Enterprise. 18+ years delivering data platforms for Australian businesses — building the systems these exams test every day.

LinkedIn ↗lucidlabs.com.au ↗Published 29-03-2026