Complete guide
AI-901 Azure AI Fundamentals
Tap any service to see more details and official docs.
Exam breakdown
40–60
Questions
~50
Min read
700
Passing score
Identify AI concepts and capabilities
40–45% of the examWhat is Artificial Intelligence?
AI is the simulation of human intelligence processes by computer systems. In Azure, AI is delivered through models and services that let applications see, hear, speak, understand, search and generate new content.
Vision
Interpret images and video, and generate new images
Speech
Recognize and synthesize human speech
Language
Understand, extract information from, and generate natural text
Search
Index and search by meaning (retrieval)
Generative AI
Create new content: text, code, images
Agentic AI
Systems that plan and execute tasks autonomously
Responsible AI — Microsoft's 6 principles
⚖️ Fairness
AI systems should treat everyone fairly, without bias based on race, gender, or other characteristics.
🔒 Reliability and safety
Systems should operate reliably and safely, even in unexpected situations.
🛡️ Privacy and security
Systems should protect data and respect people's privacy.
🌍 Inclusiveness
Systems should empower everyone, including people with disabilities.
🔍 Transparency
Systems should be understandable. People should know how they work.
✅ Accountability
People should be accountable for the AI systems they design and deploy.
Responsible AI's 6 principles show up often on the exam: Fairness, Reliability and safety, Privacy and security, Inclusiveness, Transparency, Accountability.
How generative AI models work
Generative AI models (LLMs) are trained on massive amounts of text and learn to predict the most likely next word (token) given the prior context. Scaled up, that ability produces coherent text, code, and even step-by-step reasoning.
| Concept | Definition |
|---|---|
| LLM (Large Language Model) | A model trained on trillions of tokens to understand and generate text |
| Token | The unit of text a model processes (~4 characters or ~¾ of a word) |
| Multimodal model | Accepts and/or generates more than one content type: text, image, audio |
| Prompt | The instruction or input text given to a model |
| Temperature | Controls response creativity: 0 = deterministic, 1 = very creative |
| Grounding | Anchoring a model's responses in real data to reduce hallucinations |
Choosing the right model
The Microsoft Foundry model catalog offers models from multiple providers (OpenAI, Meta, Mistral, DeepSeek and more). Choosing the right model depends on the task, not just "which one is newest."
Text vs. multimodal
A text-only model can't interpret an image or audio in the prompt.
Size and cost
Larger models reason better but cost more and respond more slowly.
Reasoning
Some models are optimized for step-by-step problems, not just conversation.
Open vs. proprietary
Open-weight models allow more control over deployment and fine-tuning.
Deployment options and configuration parameters
| Option | What it controls |
|---|---|
| Serverless / pay-as-you-go | Billed by tokens consumed, no reserved capacity |
| Provisioned throughput (PTU) | Reserved capacity for predictable load and low latency |
| Region / data residency | Where the deployment's data is processed and stored |
| Temperature / top-p | How deterministic or creative the generated response is |
| Max tokens | Length limit on the generated response |
Access to certain models in Foundry requires prior approval — not every model is automatically available in every project or region.
AI workloads — identifying the right scenario
Click any service to see more details
Text analysis
| Capability | What it detects/generates | Example |
|---|---|---|
| Sentiment Analysis | Positive / Negative / Neutral + score | "I loved the product" → Positive 0.97 |
| Key Phrase Extraction | Main concepts in the text | "Microsoft Foundry" in a technical article |
| Entity Detection (NER) | People, places, dates, organizations | Microsoft (org), Redmond (place), 2026 (date) |
| Summarization | Automatic summary of a long text | Summarizing a 10-page report into 3 lines |
Speech recognition and synthesis
🎤 Speech to Text (STT)
Transcribes audio in real time or in batch. Supports 100+ languages.
🔊 Text to Speech (TTS)
Converts text to natural speech. High-quality neural voices.
🌐 Speech Translation
Translates speech in real time between languages.
🎙️ Speaker Recognition
Identifies or verifies a person's identity from their voice.
Computer vision and image generation
| Capability | Description | Use case |
|---|---|---|
| Image Analysis | Describes the image, detects objects and tags | Cataloging a photo inventory |
| OCR / Read API | Extracts text from images and documents | Digitizing invoices and forms |
| Content Understanding | Classifies and extracts data from images with your own schema | Manufacturing QA, defect classification |
| Image generation | Creates new images from a description | Marketing illustrations from a prompt |
Multimodal information extraction
Extracting structured information from unstructured content (free text, images, audio, video) is one of the exam's key workloads — and it\'s the gateway to domain 2, where it\'s implemented with Content Understanding.
Documents
Key fields, tables and text from forms and contracts
Images
Objects, embedded text and visual metadata
Audio
Transcription, topics and mentioned entities
Video
Key moments, transcription and detected objects
Implement AI solutions by using Microsoft Foundry
55–60% of the examMicrosoft Foundry — the implementation platform
Click any service to see more details
Generative apps and agents with Foundry
Designing good system and user prompts is the foundation of any generative solution: the system prompt sets the model's role and rules, the user prompt is the one-off request.
Deploy in the portal
Pick a model from the catalog, deploy it and test it right in the Foundry portal, no code required.
Chat client with the SDK
Build a lightweight chat client using the Foundry SDK, pointed at the project endpoint.
Portal Agents
Build a single-agent solution directly in the portal — no infrastructure to maintain.
Client for the agent
Build a lightweight app that calls that agent and displays its responses.
Portal Agents = no code, no infrastructure of your own, defined in the portal or via SDK/REST. Hosted Agents = your own code (Agent Framework, LangGraph, OpenAI Agents SDK) packaged and running on an endpoint managed by Foundry. The exam distinguishes both concepts.
Text and speech with Foundry
Click any service to see more details
App with text analysis
Build a lightweight app that uses Foundry Tools (Language) for sentiment, entities or summaries.
Spoken prompts to a multimodal model
Respond to audio directly with a deployed model — no separate speech service in between.
App with Azure Speech in Foundry Tools
Use Azure Speech (inside Foundry Tools) to transcribe or synthesize speech in your app.
Vision and image generation with Foundry
Interpreting images in the prompt
A multimodal model deployed in Foundry can accept an image alongside text and answer about its content.
Generating new images
Use an image-generation model to create visual content from a description.
App with vision capabilities
Build a lightweight app that combines both capabilities: interpreting and generating images.
Information extraction with Content Understanding
Click any service to see more details
| Source | What Content Understanding extracts |
|---|---|
| Documents and forms | Key fields, tables, text — with prebuilt analyzers for invoices, contracts, receipts |
| Images | Objects, embedded text and visual metadata |
| Audio | Transcription, topics and mentioned entities |
| Video | Key moments, transcription and detected objects |
Content Understanding is first-class in the Microsoft Foundry portal — when the exam asks about "extracting information" inside a Foundry solution, this service is almost always the answer.
Interactive diagram — follow a question's flow from the user to the final answer. Hover over each node to see its role.
Grounding with RAG
RAG (Retrieval-Augmented Generation) is the most common pattern for grounding a model's responses in your own data, within a solution built on Foundry.
Indexing
Documents are split into chunks and turned into embeddings → Azure AI Search
Retrieval
The user's question is turned into an embedding and searched against AI Search
Generation
Relevant docs + the question are sent as a prompt to the model deployed in Foundry
Response
The model generates a response grounded in the retrieved documents
Hallucinations happen when a model generates plausible but incorrect information. RAG with grounding reduces this by anchoring responses in real documents.
Tips for exam day
The exam has only 2 domains: "Identify AI concepts and capabilities" (40–45%) and "Implement AI solutions by using Microsoft Foundry" (55–60%). Almost 60% is about Foundry.
Responsible AI has 6 principles: Fairness, Reliability and safety, Privacy and security, Inclusiveness, Transparency, Accountability. Memorize them.
Portal Agents = no-code agent, no infrastructure of your own. Hosted Agents = your own code running on an endpoint managed by Foundry. The exam distinguishes both.
A multimodal model can accept an image or audio directly in the prompt — you no longer need a separate vision or speech service just to "interpret" the content.
Content Understanding extracts information from documents, images, audio and video — it's the answer when the exam asks about "extracting information" inside Foundry.
Azure Speech lives inside Foundry Tools (formerly Azure AI Services). It's the right service for "text and speech by using Foundry".
Access to certain models in Foundry requires approval — it's not automatically available to every project.
Content filtering is enabled by default on Foundry models. It can be configured but not fully disabled.
RAG (Retrieval-Augmented Generation) = Azure AI Search (retrieval) + a model deployed in Foundry (generation). Find relevant docs → pass them to the model as context.
Foundry SDK = a thin client with a single project endpoint. Higher-level frameworks (like Agent Framework) are built on top of it, not a replacement for it.
Want to go deeper?
This guide is a quick summary. Visit the Deep Dive section for an exhaustive walkthrough with real cases, cost analysis and architecture decisions.
Practice AI-901 with interactive simulations
Click-through labs: navigate a step-by-step portal simulation, with clear objectives and verification on every action — no real costs.
Ready to practice?
Put what you learned to the test with original AI-901 exam questions.
Start AI-901 mock examUnlimited mock exams with the Pro plan
Was this guide helpful?
Not ready to start yet?
Leave your email and create your free account whenever you're ready — start with 15 AI-901 questions, no credit card required.
Already have an account? Go straight to the mock exam →