AI-901 Study Guide & Cheat Sheet (Azure AI Fundamentals)

Microsoft Azure AI Fundamentals (AI-901) Exam Guide

Exam domains

Who it’s for: You are at the beginning of a career in AI solution development. Microsoft expects conceptual knowledge of AI solutions in Azure plus the foundational technical skills to build them: Python coding syntax, Azure resources, and familiarity with REST APIs, SDKs, and CLIs. If you studied for AI-900, note that regression, clustering, and Azure Machine Learning designer are no longer the focus. Microsoft Foundry is.

Study & test-day tips

• AI-901 replaced AI-900 on June 30, 2026. If your study material talks about regression, clustering, and Azure Machine Learning designer as major topics, it was written for the retired exam.
• Over half the exam is Microsoft Foundry implementation. Open the Foundry portal, deploy a model, and chat with it before exam day. Reading about it is not enough.
• You do not have to write Python from scratch, but you must read it. Expect snippets using the Foundry SDK where you pick the correct line, argument, or client call.
• Know the Foundry vocabulary cold: Foundry resource, project, model catalog, deployment, endpoint, playground, and Foundry Tools.
• Agents come up repeatedly. Understand what an agent adds over a bare model: instructions, tools, and the ability to take actions in a loop.
• Azure Content Understanding is its own sub-domain. It extracts structured information from documents, forms, images, audio, and video. Do not confuse it with the older Document Intelligence branding.
• Multimodal means one model handling text plus images or audio. Expect questions where the right answer is 'use a multimodal model' instead of chaining separate vision and language services.
• The six responsible AI principles are still worth easy points. Memorize them and one concrete example each.
• Temperature, top_p, and max tokens are configuration parameters you will be asked to reason about. Higher temperature means more random output, lower means more deterministic.
• Read the scenario for the deciding constraint: cost, latency, privacy, or accuracy. Microsoft writes distractors that are technically real services but wrong for the stated constraint.

Cheat sheet

The two domains (and their real weights)

• Identify AI concepts and capabilities: 40-45%
• Implement AI solutions by using Microsoft Foundry: 55-60%
• If your prep is mostly concepts, you are studying under half the exam.

Microsoft Foundry building blocks

• Portal: deploy models, use the playground, build and test agents
• Model catalog: browse and compare models before deploying
• Deployment: a hosted instance of a model with its own endpoint
• Foundry SDK: the Python client you use to call the deployment from an app
• Foundry Tools: includes Azure Speech and Azure Content Understanding

Model configuration parameters

• temperature: randomness. Low for factual and deterministic, high for creative
• top_p: nucleus sampling. Tune this or temperature, not both at once
• max tokens: caps the response length and therefore cost
• context window: how much input plus output the model can hold at once
• system prompt: sets persistent behavior. User prompt: the immediate request

Which service for which job

• Chat, summarize, generate: a deployed generative model via Foundry
• Take actions with tools in a loop: an agent
• Speech to text or text to speech: Azure Speech in Foundry Tools
• Pull fields from invoices, forms, images, audio, video: Content Understanding
• Describe or classify an image in a prompt: a multimodal model

Responsible AI, six principles

• Fairness: avoid bias across groups
• Reliability and safety: perform consistently, fail safely
• Privacy and security: protect data at rest, in transit, and in prompts
• Inclusiveness: work for people of all abilities
• Transparency: users understand what the system does and its limits
• Accountability: humans remain responsible for outcomes

Retired with AI-900 (deprioritize)

• Regression, classification, and clustering as major topics
• Azure Machine Learning designer and automated ML workflows
• Individual Cognitive Services branding as the primary framing
• These may still appear as background concepts, but they are no longer the focus

Glossary

Agent

A model plus instructions, tools, and the ability to act in a loop. Use an agent when the task needs steps or tool calls; use a bare model when a single response is enough.

Artificial intelligence (AI)

Software that performs tasks normally requiring human intelligence, such as recognizing images, understanding language, or making predictions.

Azure AI Face

Service that detects, analyzes, and recognizes human faces and facial attributes in images.

Azure AI Foundry

Former name of Microsoft Foundry. Study material and exam items may use either name; they refer to the same platform.

Azure Language in Foundry Tools

Formerly Azure AI Language. Service (formerly Text Analytics/LUIS) for sentiment analysis, entity recognition, key phrase extraction, and language understanding.

Foundry Tools

The name the AI-901 outline uses for prebuilt AI capabilities you call from Microsoft Foundry, such as Azure Speech in Foundry Tools and Azure Content Understanding in Foundry Tools. Older material calls these Azure AI services, and before that Cognitive Services.

Azure Speech in Foundry Tools

Formerly Azure AI Speech. Service for speech-to-text, text-to-speech, speech translation, and speaker recognition.

Azure Translator in Foundry Tools

Formerly Azure AI Translator. Service that performs real-time text translation across many languages.

Azure Vision in Foundry Tools

Formerly Azure AI Vision. Service (formerly Computer Vision) for image classification, object detection, captioning, and OCR.

Azure Content Understanding

The Foundry Tools service that extracts structured information from documents, forms, images, audio, and video. Its own sub-domain on AI-901.

Azure Machine Learning

Azure's cloud platform for building, training, deploying, and managing ML models at scale.

Azure OpenAI Service

Azure service providing access to OpenAI models including GPT (text/chat), embeddings, and DALL-E (image generation).

Classification

A supervised ML task that predicts which category or class an item belongs to.

Context window

The total tokens a model can consider at once, covering the prompt plus the response. Long documents may need chunking to fit.

Deep learning

Machine learning that uses multi-layered neural networks to model complex patterns, powering vision and language tasks.

Deployment

A hosted instance of a chosen model with its own endpoint and capacity. Your app talks to the deployment, not to the catalog entry.

Embedding

A numeric vector representation of text that captures meaning, used for search, similarity, and grounding.

Endpoint

A deployed model exposed as a callable service so applications can request predictions in real time or batch.

Foundry portal

The web interface for Microsoft Foundry where you deploy models, use the playground, and create and test agents without writing code.

Foundry project

The workspace that holds your deployments, agents, connected data, and tools. You create a project before deploying anything.

Foundry SDK

The client library (commonly Python) used to call a deployed model or agent from an application. AI-901 shows you snippets and asks which call or argument is correct.

Generative AI

AI that creates new content - text, images, code, or audio - in response to a prompt.

Grounding

Supplying a model with authoritative source data at request time so its answers reflect your content instead of only its training data.

Large language model (LLM)

A generative model trained on vast text data to understand and produce human-like language.

Machine learning (ML)

A subset of AI in which models learn patterns from data rather than being explicitly programmed with rules.

Max tokens

The cap on how many tokens a response may contain, which bounds both length and cost.

Microsoft Foundry

The unified platform (formerly Azure AI Foundry / Azure AI Studio) for building AI solutions: browse the model catalog, deploy models, test in the playground, build agents, and call everything from the Foundry SDK. AI-901 is built around it.

Model catalog

The browsable library of models in Foundry from Microsoft, OpenAI, and open-source providers, used to compare and pick a model before deploying it.

Multimodal model

A single model that accepts more than one input type, such as text plus images or audio, removing the need to chain separate vision and language services.

Named entity recognition

An NLP task that identifies entities such as people, places, organizations, and dates within text.

Natural language processing (NLP)

AI that interprets, analyzes, and generates human language in text or speech.

Object detection

A vision task that identifies objects in an image and locates each with a bounding box.

Optical character recognition (OCR)

Technology that reads printed or handwritten text from images and documents into machine-readable text.

Playground

The interactive chat surface in the Foundry portal for testing prompts and configuration against a deployment before writing code.

Prompt and completion

The prompt is the input you give a generative model; the completion is the output it generates in response.

Responsible AI

Microsoft's framework of six principles - fairness, reliability and safety, privacy and security, inclusiveness, transparency, and accountability - for building trustworthy AI.

Retrieval Augmented Generation (RAG)

A pattern that retrieves relevant documents and includes them in the prompt so the model answers from your data with citations.

Sentiment analysis

An NLP task that classifies text as positive, negative, or neutral in tone.

System prompt

Instructions that set a model's persistent behavior and role for the whole conversation, as opposed to the user prompt, which is the immediate request.

Temperature

A configuration parameter controlling randomness. Lower values give focused, deterministic output; higher values give more varied, creative output.

Token

A chunk of text (word or sub-word) that a language model processes; usage and limits are measured in tokens.

top_p

Nucleus sampling: the model considers only the most probable tokens whose cumulative probability reaches p. Tune temperature or top_p, not both at once.

Transformer

A neural network architecture using attention mechanisms; the foundation of modern large language models.