Launch pricing — Azure certification prep for €19,95 €29,95

Microsoft Azure · AI Fundamentals

AI-901 Practice Questions and Answers

AI-901 · 15 free questions · Skills outline of 15 April 2026

These are 15free AI-901 practice questions, each with the answer and a short explanation. They follow Microsoft's current AI-901 skills outline (15 April 2026), so they cover Microsoft Foundry, agents and multimodal models rather than the older AI-900 topics.

AI-901 replaced AI-900 for the Azure AI Fundamentals certification after AI-900 retired on 30 June 2026. If you are not sure what changed, read AI-900 vs AI-901: what changed first. Try to answer each question before you open the answer.

Identify AI concepts and capabilities (40-45%)

Responsible AI principles, how generative models work, choosing a model, and recognising AI workloads such as text analysis, speech, vision and information extraction.

Question 1

A bank's loan approval model approves applicants from one region far less often than equally qualified applicants from other regions. Which responsible AI principle is the main concern?

  • A. Transparency
  • B. Fairness
  • C. Reliability and safety
  • D. Inclusiveness
Show answer

B. Fairness is about an AI system treating similar people in the same way. Transparency would be the concern if applicants could not find out how the decision was made.

Question 2

A company tells customers when they are talking to an AI agent and explains what the agent can and cannot do. Which responsible AI principle does this support?

  • A. Accountability
  • B. Privacy and security
  • C. Transparency
  • D. Fairness
Show answer

C. Transparency means people understand that they are dealing with AI, how it works and where its limits are.

Question 3

An app adds live captions and a read-aloud option so that people with hearing or vision impairments can use it. Which responsible AI principle does this support?

  • A. Inclusiveness
  • B. Reliability and safety
  • C. Accountability
  • D. Transparency
Show answer

A. Inclusiveness means designing AI so that it works for everyone, including people with disabilities.

Question 4

How does a generative language model produce its answer?

  • A. It looks up a stored answer in a database of questions
  • B. It predicts the next token, one at a time, based on the prompt and the tokens it has produced so far
  • C. It follows a fixed set of if-then rules written by developers
  • D. It copies the best matching web page
Show answer

B. A language model generates text token by token, each time predicting a likely next token. That is also why it can produce fluent text that is wrong.

Question 5

You deploy a model for a support bot and want its answers to be more consistent and less creative. Which setting do you lower?

  • A. Maximum tokens
  • B. Temperature
  • C. The number of past messages included
  • D. The number of model deployments
Show answer

B. A lower temperature makes the output more predictable. Maximum tokens only limits how long a response can be.

Question 6

Users of your app will upload a photo of a damaged product and ask questions about it in text. Which type of model do you need?

  • A. A speech synthesis model
  • B. A text-only language model
  • C. A multimodal model that accepts images and text
  • D. An image-generation model
Show answer

C. A multimodal model can take an image and text in the same prompt and answer about both. An image-generation model creates new images; it does not analyse the one you give it.

Question 7

A retailer wants to know whether thousands of product reviews are positive, negative or mixed. Which text analysis technique fits?

  • A. Keyword extraction
  • B. Entity detection
  • C. Summarization
  • D. Sentiment analysis
Show answer

D. Sentiment analysis scores the opinion in a text. Entity detection finds names of people, places and organizations; keyword extraction pulls out the main terms.

Implement AI solutions by using Microsoft Foundry (55-60%)

Prompts, deploying models, agents, and building small text, speech, vision and information extraction apps with Foundry.

Question 8

Your chat app must always answer in plain English and only about your company's products, whatever the user types. Where do you put that instruction?

  • A. In the system prompt
  • B. In each user prompt
  • C. In the model's name
  • D. In the maximum tokens setting
Show answer

A. The system prompt sets the model's role, tone and rules for the whole conversation. User prompts are the questions people type.

Question 9

What makes an AI agent different from a basic chat app built on the same model?

  • A. An agent always uses a larger model
  • B. An agent can use tools and take steps to complete a task, not only reply with text
  • C. An agent does not need instructions
  • D. An agent only works with speech input
Show answer

B. An agent combines a model with instructions and tools (such as search, code or business systems) so it can act towards a goal. The model itself can be the same.

Question 10

You want to try a model's answers before writing any code. Where in Foundry can you do this after deploying it?

  • A. Only through the Foundry SDK
  • B. In the Foundry portal, by chatting with the deployed model
  • C. In Azure Content Understanding
  • D. Nowhere: a model must be called from an app first
Show answer

B. The Foundry portal lets you deploy a model and interact with it directly. The SDK is for building your own client app afterwards.

Question 11

A call centre app must read answers aloud to callers in a natural-sounding voice. Which capability is this?

  • A. Speech recognition
  • B. Speech synthesis
  • C. Sentiment analysis
  • D. Entity detection
Show answer

B. Speech synthesis (text to speech) turns text into spoken audio. Speech recognition is the other direction: spoken audio to text.

Question 12

A finance team wants to pull the invoice number, date and total from thousands of scanned invoices. Which Foundry tool is built for this?

  • A. Azure Speech in Foundry Tools
  • B. Azure Content Understanding in Foundry Tools
  • C. An image-generation model
  • D. Sentiment analysis
Show answer

B. Content Understanding extracts information from documents and forms. It also handles images, audio and video.

Question 13

A marketing team wants to create a new product image from a written description. Which kind of model do they need?

  • A. An image-generation model
  • B. Image analysis (computer vision)
  • C. Optical character recognition
  • D. Speech synthesis
Show answer

A. Image-generation models create new visual output from a prompt. Image analysis describes or tags an image that already exists.

Question 14

A company records its training videos and wants a summary and the key moments of each one extracted automatically. Which Foundry tool fits?

  • A. Azure Speech in Foundry Tools only
  • B. A text-only language model
  • C. Azure Content Understanding in Foundry Tools
  • D. An image-generation model
Show answer

C. Content Understanding extracts information from audio and video as well as documents and images. Speech alone would give you a transcript, not structured key moments.

Question 15

For each statement about prompts, answer Yes or No.

  • 1. Stating a role and the expected answer format in the prompt usually improves the response.
  • 2. Giving one or two examples of the answer you want can help the model.
  • 3. Once a conversation has started, the model ignores the system prompt.
Show answer

Yes, Yes, No. Clear instructions, a defined format and examples all steer the model; the system prompt applies to every turn of the conversation.

How to use these questions

If you got most of the first section right, the concepts are in place. The second section carries more weight on the exam and is where AI-901 differs most from AI-900: Microsoft recommends hands-on practice, so deploy a model in the Foundry portal, build a small agent and call it from a few lines of Python before you book.

We don't offer a full AI-901 question bank yet. Taking another Microsoft fundamentals exam as well? We have full banks for AZ-900, SC-900 and DP-900, each with 15 free questions to start.

Frequently asked questions

▶

Are these real AI-901 exam questions?

No. They are original practice questions written from Microsoft's AI-901 study guide (skills measured as of 15 April 2026). Real exam questions are confidential, and using leaked ones breaks Microsoft's exam rules.

▶

What is the passing score for AI-901?

700 on a scale of 1 to 1,000. It is a scaled score, so it does not mean 70% of the questions.

▶

How many questions are on the AI-901 exam?

Microsoft does not publish a fixed number. It says most of its exams have between 40 and 60 questions, and the count can change.

▶

Do I need to know Python for AI-901?

Some. Microsoft says candidates need knowledge of Python coding syntax and programming techniques and should be familiar with Azure resources, because over half of the exam is about building small solutions in Microsoft Foundry.

▶

Is there an official AI-901 practice assessment?

Yes. Microsoft's free practice assessment for AI-901 runs on AI Skills Navigator, and you need to sign in to start it. The link is in the AI-901 study guide on Microsoft Learn.

Sources: the AI-901 study guide, the Azure AI Fundamentals certification page and exam duration and question types, Microsoft Learn.