Domain 1: Identify AI concepts and capabilities
- The six Microsoft Responsible AI principles are fairness, reliability and safety, privacy and security, inclusiveness, transparency, and accountability. Expect to match a scenario to a principle rather than to recite the list.
- Fairness is about a model treating groups comparably; inclusiveness is about the solution being usable by people of differing abilities and circumstances. They are easy to confuse and the exam separates them.
- Transparency means a person can understand how and why a system reached an output, and accountability means named people remain answerable for what it does. Transparency is a property of the explanation; accountability is a property of the organisation.
- Reliability and safety cover consistent behaviour and safe failure, including in conditions the system was not trained for. Privacy and security cover what data the system holds and who can reach it.
- A generative model predicts the next token given the tokens before it. Text is broken into tokens, and embeddings represent meaning as numeric vectors, which is what makes similarity search and retrieval possible.
- A prompt is the input and a completion is the response. The context window bounds how much text the model can consider at once, which is why long documents are retrieved in pieces rather than pasted in whole.
- Model configuration parameters change the output without retraining anything: temperature and top-p control how varied the sampling is, and a maximum token setting bounds the length of the response.
- Choosing a model is a capability question before it is a cost question: whether it needs to handle images or audio as well as text, how large a context it must hold, what languages it covers, and how fast and expensive it is per token.
- Deployment options differ in who provides the capacity. A serverless endpoint is billed per token with no infrastructure to manage; managed compute is dedicated capacity you provision and pay for whether or not it is busy.
- The AI workloads to recognise are generative and agentic AI, text analysis, speech, computer vision, and information extraction. Being able to place a described scenario into one of those is a large part of the first skill area.
- Common text analysis techniques are keyword and key phrase extraction, entity detection, sentiment analysis and summarisation. Entity detection finds the things named in the text; sentiment analysis judges the attitude expressed.
- Speech divides into recognition and synthesis: speech to text transcribes what was said, and text to speech generates spoken audio. Speech translation combines the two across languages.
- Computer vision covers interpreting an existing image - captions, tags, objects, text through optical character recognition - while image generation models create new images from a prompt. The exam distinguishes reading an image from making one.
- Information extraction pulls structured fields out of unstructured sources: text, images, audio and video. It is the workload behind turning an invoice, a recording or a scanned form into data an application can use.
Domain 2: Implement AI solutions by using Microsoft Foundry
- Microsoft Foundry is the platform the second skill area is built on: projects and connected resources, a model catalogue to browse and compare foundation models, and the portal where a model is deployed and tried.
- A system prompt sets the role, constraints and tone that persist across the conversation; a user prompt is the individual request. Putting the standing instructions in the user prompt is the common mistake, because they are then easy for later turns to override.
- Prompt technique matters more than prompt length: zero-shot asks directly, few-shot supplies worked examples of the format wanted, and asking a model to work through its reasoning helps on multi-step problems.
- Deploying a model in the Foundry portal produces an endpoint and a key, and the portal playground lets you interact with it before any code exists. That is the fastest path from choosing a model to seeing its behaviour.
- A lightweight chat client is built with the Foundry SDK against that endpoint: authenticate, send the message history, read the completion. The SDK is why Microsoft expects familiarity with Python and with calling REST APIs.
- An agent differs from a single prompt by having tools and steps. It can call functions, consult knowledge sources, and take several actions towards a goal rather than returning one response.
- A single-agent solution is created and tested in the Foundry portal by giving the agent instructions, attaching its tools and knowledge, and running it against sample inputs before any client application is written.
- Tool or function calling is what lets an agent act: the model requests a named function with arguments, your code runs it, and the result is returned to the model. The model never executes anything itself.
- Grounding and retrieval-augmented generation supply the model with your own trusted content at query time so answers reflect current, specific data. It is the standard answer to a model that is fluent but wrong.
- A text analysis application built on Foundry sends the text and reads back the extracted sentiment, entities, key phrases or summary, so the analysis techniques from the first skill area become an API call in the second.
- A multimodal model can accept speech or an image in the same prompt as text, which is how a spoken question is answered directly. Azure Speech in Foundry Tools is the alternative when transcription or synthesis is wanted as its own step.
- Vision capabilities work the same way in both directions: a multimodal model interprets an image supplied in the prompt, and an image-generation model produces new visual output from a text description.
- Azure Content Understanding is the information-extraction service in Foundry Tools. It pulls fields from documents and forms, from images, and from audio and video, using a schema that names the fields you want back.
- Operating a deployed solution responsibly means content filtering against harmful output, defences against prompt injection, evaluation on groundedness and relevance, and watching quota and cost - a token-billed endpoint has no natural ceiling.
AI-901 exam tips
- Two skill areas, not five, and the Foundry implementation half is the larger at 55-60%. Spend your time there rather than on concept definitions.
- Learn the six Responsible AI principles by name and practise matching a scenario to one. Fairness against inclusiveness, and transparency against accountability, are the pairs that get confused.
- Do not expect classical machine-learning questions. The current skills list covers generative and agentic AI, text analysis, speech, computer vision and information extraction; regression, clustering and train/test splits are not in it.
- Know Azure Content Understanding by name. Information extraction from documents, images, audio and video is a published sub-topic and it is the one candidates most often have never touched.
- Coding is expected, contrary to what a fundamentals label suggests. Microsoft asks for Python syntax and programming techniques and familiarity with REST APIs, SDKs and CLIs, so read the Foundry SDK quickstarts rather than only the concept pages.
Study guide FAQ
Is AI-900 still available?
No. Microsoft retired AI-900 and replaced it with AI-901. Register for AI-901; both grant the same Azure AI Fundamentals credential.
Do I need coding experience for AI-901?
Yes, some. Microsoft’s audience profile states that you need knowledge of Python coding syntax and programming techniques, and the exam notes add that you should be familiar with REST APIs, SDKs and CLIs. Several published objectives are about building a lightweight client application against a deployed model or agent, so treat it as a fundamentals exam with hands-on expectations rather than a purely conceptual one.
What changed from AI-900 to AI-901?
The exam reorganised into two skill areas - identifying AI concepts and capabilities, and implementing AI solutions by using Microsoft Foundry. The emphasis moved to generative AI, agents, multimodal models and Content Understanding, and away from the classical machine-learning framing that AI-900 used.
How much of the exam is Microsoft Foundry?
The Foundry implementation skill area is 55-60% of the exam, so the majority. It covers generative apps and agents, text and speech, computer vision and image generation, and information extraction with Content Understanding.
Official exam sources
The domain names and weightings on this page follow the published exam blueprint. Each source below records what it confirmed and when it was read, so the split can be checked rather than taken on trust.
- Microsoft Learn - Exam AI-901 study guideBlueprint revised April 15, 2026 · link and content verified 8 September 2026Two skills areas: 40-45% and 55-60%.