The short version
AI-901 has two domains: concepts and capabilities (40-45%), and implementing with Microsoft Foundry (55-60%). Under them sit seven skill areas, each with a few bullets in Microsoft’s study guide. This page walks through all seven using the “Skills measured as of April 15, 2026”. Microsoft notes that its bullets illustrate how a skill is assessed, so related topics may appear too.
Domain 1: Identify AI concepts and capabilities (40-45%)
Describe principles of responsible AI
What it covers. Six principles: fairness, reliability and safety, privacy and security, inclusiveness, transparency, and accountability. The study guide asks for “considerations” for each, which means you apply them to scenarios.
Typical traps.
- Mixing up transparency and accountability. Transparency is about telling people how the system works and what it can and cannot do. Accountability is about who is answerable and what governance exists.
- Reading “a model that works poorly for people with accents” as fairness when the stem is really about inclusiveness. Look for the group that is being left out versus a group that is treated unequally.
Practice. Write one scenario per principle in your own words and say which principle it is and why not the neighbor.
Identify AI model components and configurations
What it covers. How generative models work (tokens, embeddings, predicting the next token), picking a model by capability, and choosing a deployment option and configuration parameters.
Typical traps.
- Believing a bigger model is always the right answer. Small models can win on cost, speed, and on-device use.
- Confusing parameters. Temperature changes how varied the output is, and a token limit caps how long the answer can be. A question that says “more predictable” points at temperature, not length.
- Treating fine-tuning as the way to add fresh knowledge. Grounding the model with your own data at query time is the usual answer for content that changes.
Practice. In the portal, deploy a model and change one parameter at a time. Note what changes.
Identify AI workloads
What it covers. Matching scenarios to workloads: generative and agentic AI, text analysis, speech, computer vision and image generation, and information extraction. It includes text techniques such as keyword extraction, entity detection, sentiment analysis, and summarization.
Typical traps.
- Key phrases versus entities. Key phrases are the main talking points with no type. Entities are typed, such as person, place, or organization.
- Speech recognition versus speech synthesis. One turns audio into text, the other turns text into audio.
- Extractive versus abstractive summaries. Extractive reuses existing sentences, abstractive writes new ones.
Practice. Make a two-column list: a plain-language need on the left and the workload on the right. Quiz yourself from it.
Domain 2: Implement AI solutions by using Microsoft Foundry (55-60%)
Implement generative AI apps and agents by using Foundry
What it covers. Writing effective system and user prompts, deploying a model and using it in the Foundry portal, building a lightweight chat client with the Foundry SDK, and creating and testing a single agent, plus a client app for it.
Typical traps.
- Putting persona and rules in the user message when they belong in the system prompt.
- Forgetting that an agent combines a model, instructions, and tools. A question may ask which of the three to change.
Practice. Build a small chat client, then turn it into a single agent with one tool. Read the code until each line makes sense.
Implement AI solutions for text and speech by using Foundry
What it covers. A lightweight text-analysis app, responding to spoken prompts with a multimodal model, and a lightweight app using Azure Speech in Foundry Tools.
Typical traps.
- Choosing live transcription when the audio is a stored file, or the reverse. Think about whether the audio is streaming or already recorded, and whether you need the result immediately.
- Skipping the setup details: an endpoint, a key or credential, and the right resource.
Practice. Transcribe a short recording and run text analysis on a few sentences. Note the inputs the code needs.
Implement AI solutions with computer vision and image-generation capabilities by using Foundry
What it covers. Sending images in prompts to a multimodal model, creating images with generative models, and a lightweight app with vision features.
Typical traps.
- Mixing up analyzing an image with creating one.
- Forgetting how an image reaches the model: as a link or encoded data, not a file path on your own computer.
Practice. Ask a multimodal model about a photo, then generate a new image from a prompt and compare the two workflows.
Implement AI solutions for information extraction by using Foundry
What it covers. Using Azure Content Understanding in Foundry Tools to pull information from documents and forms, images, and audio and video, and building a lightweight app around it.
Typical traps.
- Confusing values that are copied from the source with values the model infers, such as a summary.
- Overlooking that the same service handles several kinds of input.
Practice. Run one document, one image, and one audio file through an analyzer and compare the outputs.
More guides for AI-901
- AI-901 Exam Format, Questions and Scoring
- AI-901 Sample Questions With Answers
- Is AI-901 Hard? An Honest Difficulty Check
- AI-901 Prerequisites: What to Know First
- AI-901 Renewal and Retakes Explained
- Is AI-901 Worth It? An Honest Look
- AI-901 Cheat Sheet: Key Facts to Review
- AZ-900 vs AI-901: Which First?
- AB-900 vs AI-901: Which Exam First?
- AWS AIF-C01 vs AI-901: Which First?
Next step
Pick your two weakest areas and spend your next study session on them. The study plan shows where each one fits. Our AI-901 practice tests report results by topic, which makes weak areas easy to see.