Expect Python SDK questions.
Several 2026 pass reports say the hardest questions were Python snippets on client setup and response handling, not the AI concepts.

Validates foundational knowledge of AI concepts, responsible AI and building generative AI, vision and speech solutions.
AI-901 checks that you can explain core AI ideas and pick the right model or workload, and that you can read and complete small Python apps that call Microsoft Foundry. Work through the two official learning paths, build the exercises in your own Azure subscription, then practise the SDK and portal questions.
Example scenario: A support team wants to tag incoming emails as positive, neutral or negative. Which text analysis technique is that, and what would you check for fairness?
Example scenario: A short Python script creates a Foundry client but every request fails with an authentication error. What two settings would you check first?
AI-901 study guide · Skills measured as of April 15, 2026
Choose the best answer from the options, or the stated number of answers when a question asks for more than one.
A company wants a model to summarize long contracts. Which type of AI workload is this?
A. Writing a new, shorter text from a long one is a generative AI task.
Drag items from a list into the matching places; an item can be used once, more than once or not at all.
Drag each responsible AI principle to the example it matches. The examples are explaining why a loan was refused, treating all applicants equally, and protecting applicant data.
Explaining why a loan was refused → Transparency, treating all applicants equally → Fairness, protecting applicant data → Privacy and security.
Pick one value from each drop-down list in a block of text, code or a diagram to complete it.
Select the correct value in each list. Raising the temperature makes model output [more varied / more predictable], and a system prompt sets the model's [role and rules / training data].
More varied, and role and rules. A higher temperature adds randomness, and the system prompt sets how the model should behave.
Sign in to Microsoft Learn with a personal Microsoft account, complete your certification profile, then schedule AI-901 with Pearson VUE. Students and educators can schedule through Certiport.
Read the study guide, then work through the two official learning paths and build each exercise in Foundry.
Most candidates spend 10–30 hours. Plan for the low end if you already know basic Python and some Azure; plan for the high end if Python, cloud or AI concepts are new to you.
The two official learning paths map to the two skill areas. A short video course or John Savill's cram helps you review.
Run the code yourself. The SDK questions are easier once you have created a client, sent a prompt and handled a response in Python. Copy the learning path samples and change one setting at a time.
Start with Microsoft's free practice assessment on AI Skills Navigator, which needs a sign-in. Then use a question bank and rerun any code example you got wrong.
Choose the checklist for your test-centre or online appointment.
Several 2026 pass reports say the hardest questions were Python snippets on client setup and response handling, not the AI concepts.
Passers say some questions expect you to know where things are in the Foundry portal, so they recommend clicking through it with a free Azure trial before the exam.
Some passers used only the Microsoft Learn paths and a practice test, sometimes after a few days of study, and finished John Savill's cram as a last review.