Microsoft Azure AI Fundamentals
Last reviewed: May 2026
A scannable reference of architectural patterns the AI-900 exam tests. Read top-to-bottom, or jump to a section.
An AI model performs poorly for a specific demographic group due to underrepresentation in training data.
Use the Azure Machine Learning Responsible AI dashboard to identify and measure fairness disparities (e.g., demographic parity). Mitigate bias through data augmentation, resampling, or using fairness-aware algorithms.
Why: Fairness requires equitable treatment across all groups. Simply removing protected attributes is insufficient as proxy variables can still encode bias.
Regulators require an explanation for every AI-driven loan denial decision.
Use model interpretability and explainability techniques like SHAP (SHapley Additive exPlanations) via the Responsible AI dashboard to generate feature importance scores for individual predictions.
Why: Transparency is about making AI decisions understandable. Aggregate accuracy metrics are insufficient for explaining individual outcomes.
A high-stakes AI system (e.g., medical diagnosis) needs human oversight before its recommendations are acted upon.
Implement a human-in-the-loop workflow where a qualified human reviews, approves, or overrides AI-generated recommendations.
Why: Accountability ensures humans are ultimately responsible for AI system outcomes, especially in critical applications.
An AI-powered chatbot must be prevented from providing harmful or incorrect advice in sensitive domains like healthcare.
Implement rigorous testing, performance monitoring, and content filtering. Establish operational boundaries to prevent use in unintended or harmful contexts.
Why: Reliability and Safety requires that AI systems operate consistently and do no harm.
Extract searchable insights, entities, and relationships from a large, heterogeneous corpus of unstructured documents (PDFs, DOCX, images).
Use Azure AI Search. Create an enrichment pipeline that uses AI skills (e.g., OCR, entity recognition, language detection) to index content for semantic search.
Why: Knowledge mining is for creating a structured, searchable index from unstructured data, not just processing individual documents.
Monitor real-time sensor data from factory equipment to detect unusual patterns that indicate an impending failure.
Use the Azure Anomaly Detector service. Stream time-series data to the service API to identify outliers in real-time.
Why: Anomaly Detector is a pre-built service optimized for identifying unusual patterns in time-series data without requiring custom model training.
An AI-powered quality inspection system on a factory floor requires sub-50ms inference times and must operate despite intermittent internet connectivity.
Deploy an optimized AI model (e.g., ONNX format) to an edge device (like Azure IoT Edge) for local inference.
Why: Edge AI meets low-latency and offline requirements by processing data locally, avoiding network round-trips to the cloud.
Predict a continuous numerical value, such as a house price or future sales revenue.
Use a regression algorithm (e.g., linear regression, boosted trees). Evaluate with metrics like Root Mean Squared Error (RMSE) or Mean Absolute Error (MAE).
Why: Regression is for "how much?" or "how many?" questions. Use classification for "which category?" questions.
Predict a discrete category, such as whether a transaction is "fraudulent" or "legitimate".
Use a classification algorithm (e.g., logistic regression, random forest). For two categories, use binary classification. For 3+, use multi-class classification. Evaluate with a confusion matrix, precision, and recall.
Why: Classification predicts a label from a predefined set of categories.
Discover natural groupings in a customer dataset to create targeted marketing segments, without predefined categories.
Use an unsupervised clustering algorithm like K-Means. The data is unlabeled.
Why: Clustering is an unsupervised technique used to find structure in unlabeled data. Use it when you don't have a predefined target to predict.
A team with limited ML expertise needs to quickly build a model by automatically testing multiple algorithms and hyperparameters.
Use Automated ML (AutoML) in Azure Machine Learning. Specify the task (classification, regression) and let AutoML iterate through models to find the best performer.
Why: AutoML automates the time-consuming tasks of algorithm selection and hyperparameter tuning, accelerating model development.
A citizen data scientist needs to build an ML pipeline using a visual, no-code, drag-and-drop interface.
Use Azure Machine Learning designer. Connect pre-built modules for data ingestion, transformation, training, and scoring on a visual canvas.
Why: Designer is the no-code/low-code solution for building ML pipelines, distinct from the code-first SDK approach.
A data scientist needs an interactive development environment (Jupyter) and a scalable, multi-node resource for batch training.
Use an Azure ML compute instance for interactive development. Use an Azure ML compute cluster for scalable, distributed training.
Why: Compute instances are single-node workstations for development. Compute clusters are multi-node, auto-scaling resources for production training and batch inference.
A model must provide instant predictions for a web app, while another needs to score millions of records overnight.
Deploy the web app model to a real-time managed online endpoint. Deploy the batch scoring model to a batch endpoint.
Why: Online endpoints are for low-latency, synchronous inference. Batch endpoints are for high-throughput, asynchronous scoring of large datasets.
A fraud detection model trained on data with 99% non-fraudulent transactions performs poorly, missing most actual fraud cases.
Address the severe class imbalance. Techniques include oversampling the minority class (e.g., SMOTE), undersampling the majority class, or using class-weighted loss functions.
Why: With imbalanced data, a model can achieve high accuracy by simply predicting the majority class. Focus on metrics like recall and precision for the minority class.
Identify multiple items on a warehouse conveyor belt and provide the location of each item with a bounding box.
Use an object detection model. This task returns both a class label and coordinates for a bounding box for each object found.
Why: Object detection answers "what is in this image and where is it?" Image classification only answers "what is in this image?".
Digitize handwritten city council minutes from scanned images to make them searchable.
Use the Azure AI Vision service with the Read API (OCR). It supports both printed and handwritten text.
Why: OCR (Optical Character Recognition) is the specific task for extracting text from images. The Read API is Azure's most advanced OCR engine.
Identify the exact boundaries of a tumor in an MRI scan, classifying each pixel as "tumor" or "healthy tissue".
Use image segmentation. This technique classifies every pixel in an image, providing precise outlines of objects.
Why: Segmentation provides pixel-level detail, which is more precise than the bounding boxes from object detection.
Build a model to classify 50 specific types of cuisine dishes using only 75 labeled images per category.
Use the Azure AI Custom Vision service. It uses transfer learning to build effective custom classifiers with small datasets.
Why: Custom Vision is ideal for domain-specific classification tasks when you have limited training data, avoiding the need for a large, general-purpose model.
A security system must detect a person, determine their approximate age, and verify they are a live person, not a photo.
Use the Azure AI Face service. Use face detection for attributes (age, glasses) and liveness detection to prevent spoofing.
Why: The Face service provides specialized capabilities for face detection, attribute analysis, verification, identification, and anti-spoofing that general vision services lack.
Automatically extract structured data (vendor name, invoice number, line items, total) from thousands of invoices in various formats.
Use Azure AI Document Intelligence (formerly Form Recognizer) with its prebuilt invoice model.
Why: Document Intelligence is purpose-built to understand document structure and extract key-value pairs and tables, far surpassing what simple OCR can do.
Analyze a video library to automatically create a searchable index of spoken words, identified speakers, visual text, and topics.
Use Azure AI Video Indexer. It combines multiple AI models (speech-to-text, OCR, face detection, topic modeling) into a single service for deep video analysis.
Why: Video Indexer is the comprehensive service for extracting multi-modal insights from video content, not just analyzing individual frames.
A retail store wants to analyze foot traffic by counting how many people enter a specific promotional zone in real-time video.
Use the Spatial analysis capability of Azure AI Vision. Configure zones and line-crossing events to monitor people's movement.
Why: Spatial analysis is designed for understanding the movement of people in physical spaces using video feeds, a specialized task beyond general object detection.
An application needs to automatically identify and tag mentions of specific companies, people, and locations within news articles.
Use the Named Entity Recognition (NER) feature of the Azure AI Language service.
Why: NER specifically identifies and categorizes known entities into predefined types. Key Phrase Extraction finds important topics but doesn't categorize them.
Analyze a product review to determine not just the overall sentiment, but also that the "battery life" was viewed positively and the "screen" negatively.
Use Sentiment Analysis with the opinion mining feature in the Azure AI Language service.
Why: Opinion mining (aspect-based sentiment) provides granular insights by linking sentiment to specific attributes (targets) within the text.
Create a chatbot that answers customer questions using the company's existing FAQ documents and product manuals as a knowledge base.
Use the Custom Question Answering feature of Azure AI Language. Ingest the documents to create a knowledge base that the service can query.
Why: Question Answering is designed to find the best answer from a given text corpus, which is ideal for FAQ and knowledge base scenarios.
Build a virtual assistant that understands user intents like "reset password" or "book a flight" and extracts relevant entities (e.g., destination, date).
Use Conversational Language Understanding (CLU) in Azure AI Language to build a custom model that maps utterances to intents and entities.
Why: CLU is purpose-built for understanding the conversational intent in short utterances, which is different from analyzing long-form documents.
Automatically categorize incoming support emails into predefined categories like "Billing," "Technical," or "Account Inquiry".
Use the Custom Text Classification feature of Azure AI Language. Train a model with labeled examples for each category.
Why: This is a classic document classification task. Custom classification allows you to define your own business-specific categories.
A voice-enabled application must convert spoken commands to text and then generate a natural-sounding spoken response.
Use the Azure AI Speech service. Use the speech-to-text API for transcription and the text-to-speech API for synthesis.
Why: The Speech service is the centralized hub for all speech-related AI tasks, including transcription, synthesis, translation, and speaker recognition.
A voice-based security system needs to verify that a person speaking is who they claim to be.
Use the Speaker Recognition feature of Azure AI Speech to perform speaker verification against an enrolled voiceprint.
Why: Speaker recognition identifies or verifies a person by their unique voice characteristics, distinct from just transcribing what they say.
An e-commerce platform needs to translate product descriptions into 30 different languages while preserving HTML formatting.
Use the Azure AI Translator service. It supports text and document translation across 100+ languages and can preserve formatting.
Why: Translator is the dedicated, scalable service for multilingual translation, more optimized for this task than a general-purpose language model.
Deploy a chatbot to multiple channels, including a company website, Microsoft Teams, and Slack, from a single codebase.
Use the Azure Bot Service. It provides a framework and channel connectors to manage communication across various platforms.
Why: Bot Service handles the complexities of adapting a conversational AI core (like CLU or QnA) to the specific protocols of different chat platforms.
A chatbot must answer questions exclusively based on a company's internal, frequently updated knowledge base, not its general pre-trained knowledge.
Implement the Retrieval-Augmented Generation (RAG) pattern. Use Azure AI Search to retrieve relevant documents and pass them as context to an Azure OpenAI model to generate a grounded answer.
Why: RAG grounds the model in specific, current data without expensive retraining, reducing hallucinations and ensuring factual accuracy from a trusted source.
A developer needs a GPT model to respond in a specific format (e.g., JSON).
Use few-shot prompting. Provide 2-3 examples of the desired input-output format directly in the prompt before the actual request.
Why: Few-shot prompting guides the model's behavior and output structure through in-context examples, which is faster and cheaper than fine-tuning.
Improve a model's accuracy on a multi-step reasoning problem (e.g., a math word problem).
Use Chain-of-Thought (CoT) prompting by adding a phrase like "Think step by step" to the prompt.
Why: CoT encourages the model to break down the problem and show its reasoning, which significantly improves performance on complex logical tasks.
Control the creativity versus predictability of a generative model's text output.
Adjust the `temperature` parameter. A low value (~0.1) makes output more deterministic and focused. A high value (~0.9) makes it more creative and random.
Why: Temperature directly controls the randomness of token selection, allowing you to tune the output style for the specific use case (e.g., factual summary vs. creative writing).
An enterprise needs to use OpenAI's GPT-4 and DALL-E models within their secure Azure environment, with private networking and integrated identity management.
Use the Azure OpenAI Service. It provides OpenAI models with Azure's enterprise-grade security, compliance, regional availability, and content filtering.
Why: Azure OpenAI provides a secure, enterprise-ready wrapper around OpenAI models, integrating them into the Azure ecosystem.
Build a search system that finds documents based on semantic meaning, not just keyword matches (e.g., "car maintenance" finds "vehicle service intervals").
Use an Azure OpenAI embeddings model (e.g., `text-embedding-ada-002`) to convert documents and queries into numerical vectors. Use a vector database (like Azure AI Search) to find the closest vectors by cosine similarity.
Why: Embeddings capture the semantic meaning of text, enabling searches based on conceptual similarity rather than lexical overlap.
An application using Azure OpenAI must automatically prevent the generation of content related to violence, hate speech, sexual content, or self-harm.
Rely on the built-in content filtering, powered by Azure AI Content Safety. Configure the severity levels (low, medium, high) for each harm category.
Why: Azure OpenAI includes a mandatory, multi-layered safety system that filters both prompts and completions to align with responsible AI principles.
A marketing team needs to generate custom product images for advertising campaigns from text descriptions.
Use the DALL-E model available through Azure OpenAI Service. Craft a detailed prompt describing the desired image.
Why: DALL-E is a text-to-image generation model, specifically designed for creating novel images from natural language prompts.
A generative AI assistant needs to access real-time data (e.g., current stock prices) or execute actions (e.g., book a meeting) by calling external APIs.
Use the function calling capability of Azure OpenAI models. Define available functions in the API request; the model will generate a structured JSON object specifying which function to call with which arguments.
Why: Function calling allows LLMs to interact with external tools and APIs, overcoming the limitation of their static training data and enabling them to take action.
A team needs to build, evaluate, and deploy a complex generative AI application by orchestrating LLM calls, Python scripts, and prompt templates in a visual workflow.
Use Azure AI Foundry (formerly AI Studio) and its Prompt flow feature. Build the application as a visual graph of connected tools.
Why: Prompt flow is the orchestration tool for building and testing complex LLM-based applications, chaining together multiple components into a reproducible workflow.
An IT team needs to build a custom copilot for internal use that can answer employee questions and integrate with enterprise systems (e.g., ServiceNow, SAP) using a low-code platform.
Use Microsoft Copilot Studio. It provides a low-code graphical interface for building custom copilots with pre-built connectors and generative AI capabilities.
Why: Copilot Studio abstracts the complexity of building enterprise-grade AI assistants, enabling rapid development without extensive coding.
