AI-900 exam concepts in 5 categories — Search or filter by category
Quick stats
Percentage of correct predictions out of all predictions.
Different types of AI tasks: Computer Vision, NLP, ML, Conversational AI, Document Intelligence, Knowledge Mining, Generative AI, Anomaly Detection.
Set of rules or instructions that ML uses to learn patterns from data.
AI workload that identifies data points or events that deviate significantly from normal patterns.
Azure ML feature that automatically tries multiple algorithms and picks the best model.
Detect harmful text and images: hate, violence, sexual content, self-harm. Severity 0 (safe) to 6 (severe).
Train custom image classification and object detection models with your own labeled images.
OCR + AI to extract structured data from documents; prebuilt models for invoices, receipts, IDs, W-2s, business cards.
Detect faces, analyze attributes (age, emotion, glasses), optional identification (restricted).
Build, evaluate, deploy AI apps; model catalog (100+ models), prompt flow, evaluation tools.
Unified NLP service: sentiment analysis, NER, key phrase extraction, language detection, question answering, conversational understanding.
Language = text analysis (sentiment, entities). Translator = language translation only. Speech = audio (speech-to-text, text-to-speech).
Monitor metrics and detect anomalies automatically; root cause analysis, alerts.
Full-text, vector, and hybrid search; AI enrichment (OCR, entities) during indexing.
Convert speech to text, text to speech, translate spoken language, assess pronunciation.
Real-time text translation, document translation, custom glossaries.
Extract insights from video: faces, objects, text, speech, scenes, emotions, topics.
Analyze images: tags, descriptions, objects, OCR, face detection, content moderation.
Vision = prebuilt general models. Custom Vision = train your own models on custom categories.
Build, test, deploy bots; multi-channel (Teams, Slack, web); integrates with Language (QnA, CLU).
Top-level resource organizing all ML assets: compute, data, models, experiments, pipelines, endpoints.
Top-level resource container for all Azure ML assets: datasets, models, experiments, endpoints, compute.
AutoML = fully automated algorithm testing. Designer = visual drag-and-drop pipeline builder with more control.
Visual interface for building ML pipelines without writing code.
Built-in content filters for prompts and outputs: hate, violence, sexual, self-harm. Severity: safe, low, medium, high.
Access OpenAI models (GPT for text, DALL-E for images, Whisper for speech) via Azure with enterprise security.
Azure AI services aren't available in all regions; must deploy in supported region.
Conversational AI applications that interact with users via text or voice to answer questions, provide support, or complete tasks.
ML task type that predicts discrete classes or categories.
ML task type that predicts discrete categories or classes.
ML task that groups similar data points without pre-labeled categories.
Unsupervised ML that groups similar data points without predefined labels.
Top confusions: categories vs numbers, extract text vs understand meaning, learning vs predicting.
Classification = one label for whole image. Detection = multiple objects + locations (bounding boxes).
Compute Instance = single VM for dev; Compute Cluster = scalable cluster for training.
AI workload that analyzes visual content (images, videos) to detect objects, recognize faces, read text, classify images.
ML models return confidence (probability) for each prediction; set thresholds to control precision vs recall.
Table showing true positives, true negatives, false positives, false negatives for classification models.
Azure OpenAI has built-in content filters (hate, violence, sexual, self-harm); configurable severity levels.
AI workload that enables natural, human-like conversations through chatbots and virtual assistants.
Process of manually tagging training data (images, text) with correct answers before training custom models.
Subset of ML using artificial neural networks with multiple layers to learn complex patterns from large amounts of data.
AI workload that extracts structured data (text, tables, key-value pairs) from documents, invoices, receipts, forms, IDs.
40-60 questions, 180 minutes (3 hours), passing score 700/1000, no prerequisites.
Use smaller models, limit max_tokens, cache prompts, batch processing.
API key vs Managed Identity; answer is always Managed Identity for production.
Scenario describes AI problem; identify which of 6 principles is violated.
Most common exam pattern: match task/scenario to correct Azure service.
Scenario describes a prediction task; identify the correct ML type.
Match business requirement to Azure service. Know the service-to-task mapping.
Show which features influenced a prediction; required for transparency and regulations.
Harmonic mean of precision and recall; balanced metric when you need both to be good.
Detection: locate faces in images. Recognition: identify who the person is (restricted use).
Features are input data (columns); labels are what you're trying to predict (target).
Using AI to identify fraudulent transactions, activities, or patterns.
Free tier: limited transactions/month for testing. Paid tier: pay-per-use or commitment for production.
Unique risks of generative AI: creating biased/harmful content, false information, privacy violations, impersonation.
AI workload that creates new content (text, images, code, audio) based on patterns learned from training data.
LLMs sometimes generate plausible-sounding but false information; mitigation: RAG, low temperature, citations.
Critical decisions require human review, especially in healthcare, finance, legal.
Settings you configure before training (learning rate, number of layers, tree depth); model doesn't learn these.
Assigns a single label/category to an entire image.
Extracting vendor name, total, date, line items from invoices using AI.
Identifies main topics or important concepts in text.
AI workload that discovers hidden patterns and insights by indexing and searching large amounts of unstructured data.
Identifies the language of text input and returns language code with confidence score.
Massive neural networks trained on billions of text documents to understand and generate human language.
Track all AI decisions, inputs, outputs, and confidence scores for accountability and compliance.
Subset of AI where systems learn from data to make predictions without being explicitly programmed.
The trained artifact that makes predictions; result of applying algorithm to training data.
Real-time = instant predictions (milliseconds). Batch = process many predictions at once (minutes/hours).
Make trained model available for inference; real-time (REST API), batch (scheduled), edge (IoT devices).
Monitor models for data drift (input changes) and concept drift (feature-target relationship changes).
Identifies and categorizes entities in text: people, places, organizations, dates, quantities, etc.
AI workload that enables computers to understand, interpret, and generate human language (text and speech).
Computing systems inspired by biological neurons, organized in layers that process information and learn patterns.
Identifies multiple objects in an image AND their locations (bounding boxes).
Extracts text from images and documents (printed or handwritten).
Overfitting = model memorizes training data; Underfitting = model too simple.
Precision = of predicted positives, how many are correct? Recall = of actual positives, how many did we find?
Precision: Of predicted positives, how many were correct? Recall: Of actual positives, how many did we find?
Using AI to predict when equipment will fail so maintenance can be performed proactively.
Art of crafting effective instructions (prompts) to get desired outputs from AI models.
Extracting merchant, date, total, items from receipts; extracting any form fields from custom layouts.
Testing AI systems by trying to make them fail, generate harmful content, or bypass safety measures.
ML task type that predicts numbers (price, temperature, sales).
ML task type that predicts continuous numeric values.
ML where an agent learns by trial and error, receiving rewards for good actions and penalties for bad ones.
Microsoft's ethical AI framework; high exam weight. Know which principle applies to each scenario.
Make video lectures accessible; real-time captions, translation to multiple languages.
Extract vendor, total, date from 10,000 paper invoices.
Training data had historical bias; solution: audit data, balance dataset, fairness constraints, monitor.
Analyze social media images to count appearances of company logo and brand mentions.
Transcribe calls, analyze sentiment, extract key phrases, identify trends for quality improvement.
EU customers can request data deletion; AI systems must support this.
Bank must explain why AI denied a loan; use XAI (Explainable AI) and Responsible AI dashboard.
Classify chest X-rays as normal/abnormal with 2,000 labeled images.
Chatbot supporting English, Spanish, French for global customers.
Flag suspicious credit card transactions immediately.
Camera monitors shelves; count products, detect empty spaces, identify out-of-stock items.
Employees search 100,000 documents by meaning ('find docs about customer complaints').
Search by meaning and intent, not just exact keyword matching.
Determines emotional tone of text: positive, negative, neutral, or mixed.
Converts spoken audio into written text (transcription).
ML where you train with labeled data (input-output pairs); model learns to map inputs to correct outputs.
Converts written text into natural-sounding spoken audio.
Predict future values from historical time-ordered data (sales, demand, energy, stock prices).
ML datasets are split into training (70-80%), validation (10-15%), and test (10-15%) sets.
Using a pre-trained model as starting point instead of training from scratch; saves time and data.
Use a model pre-trained on large data; fine-tune for your specific task with less data and time.
Converts text from one language to another (100+ languages supported).
Users should know when AI is being used, understand its limitations, and see how decisions are made.
ML with unlabeled data; algorithm finds hidden patterns or groupings without being told what to look for.
AI-powered assistants that respond to voice commands and perform tasks (set reminders, answer questions, control devices).
Capability of a computer system to mimic human-like cognitive functions such as learning and problem-solving.
Prebuilt = ready to use, general purpose. Custom = train on your data for domain-specific needs.