AI-901 (AI901) Practice Questions and Exam Facts
10 original AI-901 (AI901) practice questions with answers, the official exam facts and skills measured, and what replaced AI-900 in 2026.
In this article
AI-900 is gone: Microsoft retired it on June 30, 2026 and replaced it with AI-901: Microsoft Azure AI Fundamentals, which is now the only exam for the Microsoft Certified: Azure AI Fundamentals certification. Below are 10 original AI-901 practice questions with answers and short explanations, split across the two official skill areas in roughly the same proportion as the real exam.
Try each question before you read the answer. If you were studying with an AI-900 practice test, the first four questions will feel familiar. The last six are the new part.
Key takeaways
- AI-901 replaced AI-900 on June 30, 2026. The certification name stayed the same: Microsoft Certified: Azure AI Fundamentals.
- 55–60% of AI-901 is about implementing AI solutions with Microsoft Foundry, so 6 of the 10 questions below come from that skill area.
- Microsoft expects AI-901 candidates to know Python coding syntax. AI-900 did not.
- You get 45 minutes and need a scaled score of 700 out of 1,000. Microsoft doesn't publish the number of questions.
What happened to AI-900?
Microsoft retired the AI-900 exam on June 30, 2026. The AI-900 exam page now says the exam "has been replaced by AI-901", and the Azure AI Fundamentals certification page lists AI-901 as the only required exam.
The change is bigger than a new code. AI-900 measured five conceptual areas: AI workloads, machine learning on Azure, computer vision, natural language processing and generative AI. AI-901 has two skill areas, and the larger one is hands-on:
| AI-900 (retired) | AI-901 (current) | |
|---|---|---|
| Skill areas | 5 conceptual domains | 2: AI concepts (40–45%) and implementing AI with Microsoft Foundry (55–60%) |
| Coding | Not required | Knowledge of Python coding syntax expected |
| Focus | Describe AI workloads and Azure services | Describe AI concepts, then build small chat apps, agents, speech, vision and extraction solutions in Foundry |
If your AI-900 practice test covers only the left column, it still helps with the concept questions, but it misses more than half of AI-901.
AI-901 exam facts at a glance
These facts come from the official AI-901 study guide (skills measured as of April 15, 2026), the AI-901 exam page, Microsoft's page on exam duration and its page on exam scoring, all checked on 2026-10-26. For the full fact sheet and a study plan, see our AI-901 exam prep page.
| Fact | Official value |
|---|---|
| Exam | AI-901: Microsoft Azure AI Fundamentals |
| Certification | Microsoft Certified: Azure AI Fundamentals |
| Time | 45 minutes (65 minutes seat time) |
| Passing score | 700 on a scale of 1–1,000 (scaled score) |
| Questions | Not published (Microsoft says most exams have 40–60) |
| Price | Depends on the country or region where you take the exam |
| Penalty for guessing | None |
| Skill area | Weight | Questions below |
|---|---|---|
| 1. Identify AI concepts and capabilities | 40–45% | 1–4 |
| 2. Implement AI solutions by using Microsoft Foundry | 55–60% | 5–10 |
The study guide also says most questions cover features that are generally available, that the exam may include commonly used preview features, and that you should be familiar with REST APIs, SDKs and CLIs.
Skill area 1: Identify AI concepts and capabilities
Question 1
A company releases a customer service chatbot. Every conversation starts with a message saying the user is talking to an AI system, what it can help with and where it may be wrong. Which Microsoft responsible AI principle does this mainly support?
- A. Fairness
- B. Inclusiveness
- C. Transparency
- D. Reliability and safety
Answer: C. Transparency means people understand that they are using an AI system, what it is for and what its limits are. Telling users up front that they are talking to AI and where it can be wrong is a direct example.
- A is wrong because fairness is about the system treating all groups of people equitably.
- B is wrong because inclusiveness is about designing for people with different abilities and backgrounds, for example accessible interfaces.
- D is wrong because reliability and safety is about the system working as intended and failing safely; a disclosure message doesn't change how the model behaves.
The six principles in the study guide are fairness, reliability and safety, privacy and security, inclusiveness, transparency and accountability. Expect scenario questions that ask you to match a practice to one of them.
Question 2
Which statement best describes how a large language model produces a response to a prompt?
- A. It searches its training data for the stored document that best matches the prompt and returns it.
- B. It repeatedly predicts the most likely next token, based on the prompt and the tokens it has already generated.
- C. It runs a fixed set of rules that developers wrote for each type of question.
- D. It looks the answer up in a database on the internet each time.
Answer: B. A language model generates text one token at a time. A token is a chunk of text, often part of a word. At each step the model predicts a likely next token from everything that came before, adds it, and repeats until the answer is complete.
- A is wrong because the model doesn't store and return whole documents; it generates new text.
- C is wrong because the behavior is learned from data during training, not written as rules.
- D is wrong because a model on its own doesn't search the internet. An app or agent can add search as a tool, but that is a separate step.
Question 3
A retailer wants to process thousands of product reviews and pull out the names of products, brands and cities that customers mention. Which text analysis technique fits?
- A. Sentiment analysis
- B. Entity detection (named entity recognition)
- C. Summarization
- D. Language detection
Answer: B. Entity detection finds and categorizes named things in text, such as products, organizations, people and locations.
- A is wrong because sentiment analysis tells you whether a review is positive, negative or neutral, not which things it names.
- C is wrong because summarization shortens text instead of listing specific items.
- D is wrong because language detection only tells you which language the review is written in.
Question 4
A navigation app must read driving directions aloud to the driver. Which speech capability does it need?
- A. Speech recognition (speech to text)
- B. Speech synthesis (text to speech)
- C. Speaker recognition
- D. Keyword extraction
Answer: B. Speech synthesis turns text into spoken audio, which is exactly what reading directions aloud requires.
- A is wrong because speech recognition does the opposite: it turns spoken audio into text, for example to understand the driver's voice commands.
- C is wrong because speaker recognition identifies who is speaking.
- D is wrong because keyword extraction is a text analysis technique and produces no audio.
Skill area 2: Implement AI solutions by using Microsoft Foundry
Question 5
A developer builds a chat app in Microsoft Foundry. Every answer must be polite, in plain English and limited to questions about the company's products, whatever the user types. Where should these rules go?
- A. In the system prompt
- B. In each user prompt, typed by the user
- C. In the model's name
- D. In the app's error messages
Answer: A. The system prompt (system message) sets the model's role, tone, scope and rules for the whole conversation. The user prompt carries the user's actual question, which changes on every turn and can't be relied on to repeat the rules.
- B is wrong because users won't add the rules themselves, and the app shouldn't depend on them to.
- C is wrong because the model or deployment name only identifies which model is called.
- D is wrong because error messages don't change how the model answers.
Our guide on how to write a system prompt goes deeper on what belongs in it.
Question 6
A developer picks a model in the Foundry model catalog and wants to call it from a Python app. What must the developer do before the app can send prompts to it?
- A. Fine-tune the model on company data
- B. Deploy the model in the Foundry resource and use that deployment's endpoint and name in the app
- C. Download the model weights to the laptop
- D. Create a Content Understanding analyzer
Answer: B. A model in the catalog has to be deployed before you can use it. The app then sends requests to the deployment, using the project endpoint, the deployment name and credentials such as a key or Microsoft Entra ID. The study guide lists both "deploy a model and interact with it in the Foundry portal" and "create a lightweight chat client application by using the Foundry SDK".
- A is wrong because fine-tuning is optional and not needed to call a model.
- C is wrong because Foundry hosts the deployed model; you call it over the network.
- D is wrong because an analyzer is for information extraction with Content Understanding, not for chatting with a model.
Question 7
A developer writes this simple chat loop in Python. Users complain that the assistant forgets what they said one message earlier. What is the fix?
system = {"role": "system", "content": "You are a helpful travel assistant."}
while True:
user_input = input("You: ")
messages = [system, {"role": "user", "content": user_input}]
reply = send_to_model(messages) # calls the deployed chat model
print("Assistant:", reply)
- A. Raise the temperature so the model remembers more
- B. Keep one
messageslist for the whole conversation and append each user message and each assistant reply to it before the next call - C. Move the system message to the end of the list
- D. Deploy a larger model
Answer: B. A chat model doesn't remember earlier requests. Each call only sees the messages you send. The loop rebuilds messages from scratch every turn, so earlier turns are lost. Creating the list once and appending the user's input and the assistant's reply each turn gives the model the conversation history.
- A is wrong because temperature changes how random the output is, not what the model can see.
- C is wrong because the order of the system message doesn't bring back missing turns.
- D is wrong because a larger model still only sees the messages in the request.
Question 8
A team creates a single agent in the Foundry portal to answer employees' questions about the company's HR policies, which are stored as PDF files. Which change lets the agent ground its answers in those documents?
- A. Add the file search tool to the agent and upload the policy documents
- B. Increase the maximum response length
- C. Add an image generation model to the agent
- D. Remove the agent's instructions so it answers freely
Answer: A. Tools extend what an agent can do. The file search tool indexes uploaded files in a vector store so the agent can retrieve relevant passages and answer from them instead of from the model's general knowledge.
- B is wrong because a longer response limit doesn't give the agent new knowledge.
- C is wrong because image generation creates pictures, not answers from documents.
- D is wrong because the instructions define the agent's role; removing them makes it less reliable, not better informed.
The pattern behind this is retrieval-augmented generation, and our post on tool calling and AI agents explains how agents decide when to use a tool.
Question 9
A field technician takes a photo of a damaged machine part and asks the company's app, "What is this part and what could be wrong with it?" Which approach fits best in Foundry?
- A. Send the photo and the question together in a prompt to a deployed multimodal model that accepts image input
- B. Send the question to an image generation model
- C. Convert the photo to audio with speech synthesis
- D. Run sentiment analysis on the photo's file name
Answer: A. A multimodal model can take an image and text in the same prompt and reason about both. The study guide calls this interpreting visual input in prompts by using a deployed multimodal model.
- B is wrong because an image generation model creates new images; it doesn't explain an existing photo.
- C is wrong because speech synthesis turns text into audio and has nothing to do with images.
- D is wrong because sentiment analysis works on text opinions, and a file name says nothing about the part.
Question 10
A call center wants to process its recorded customer calls and get structured fields for each call: a short summary, the product discussed and whether the customer was satisfied. Which Foundry tool is built for this?
- A. Azure Content Understanding in Foundry Tools, with an analyzer that defines the fields
- B. A text embedding model
- C. An image generation model
- D. The file search tool on an agent
Answer: A. Content Understanding processes documents, images, audio and video into a structured output. You define the fields you want in an analyzer schema, and it can extract, classify or generate values, such as a call summary or a satisfaction label. Post-call analytics is one of the use cases Microsoft lists.
- B is wrong because embeddings turn text into vectors for search; they don't produce labeled fields from audio.
- C is wrong because image generation creates pictures.
- D is wrong because file search helps an agent answer questions from documents; it doesn't turn each recording into structured fields.
How to score yourself
Count your correct answers out of 10, then look at where you missed. Ten questions are too few to predict an exam result, but they show which skill area needs work.
- 9–10 correct: practice with longer, timed sets and keep reviewing the questions you miss.
- 6–8 correct: go back to the bullets of the skill area you missed in the study guide.
- 5 or fewer: start with the concepts. Our prompt engineering learning path covers the generative AI ideas behind questions 2 and 5–8.
If you missed several of questions 5–10, the Foundry part is your gap. Those skills are easiest to learn by doing: deploy a model, write a system prompt, and build a small Python chat client yourself.
Microsoft also offers a free practice assessment for this certification on AI Skills Navigator (you need to sign in). It is linked from the AI-901 exam page.
How these questions were written
TokIQ Editorial wrote these questions from the skills listed in the AI-901 study guide (skills measured as of April 15, 2026). They are not real exam questions and are not taken from any question dump. Exam facts were checked against the Microsoft Learn pages linked above on 2026-10-26. TokIQ is an independent practice app and is not affiliated with, endorsed by or sponsored by Microsoft. Microsoft, Azure and Microsoft Foundry are trademarks of the Microsoft group of companies.
Studying for AWS too? Try our AWS AI Practitioner (AIF-C01) sample questions.
Practice more AI-901 questions
If you want a short set like this every day, TokIQ (on Google Play, coming soon to iPhone) has an AI-901 prep mode: a daily session that brings back the questions you missed with spaced repetition, leans toward your weakest skill area, and includes timed mock exams. Each question cites the official Microsoft Learn source it is based on, and there is an explanation after every answer, in your own language.
Not sure AI-901 is the right first exam? Compare it with other options in which AI certification to take first, or try Google Generative AI Leader practice questions.
Frequently asked questions
Was the AI-900 exam retired?
Yes. Microsoft retired AI-900 on June 30, 2026 and replaced it with AI-901: Microsoft Azure AI Fundamentals. To earn the Microsoft Certified: Azure AI Fundamentals certification you now pass AI-901, which is the only required exam (Microsoft Learn, checked 2026-10-26).
What is the difference between AI-900 and AI-901?
AI-900 had five conceptual domains (AI workloads, machine learning on Azure, computer vision, NLP and generative AI). AI-901 has two skill areas: Identify AI concepts and capabilities (40–45%) and Implement AI solutions by using Microsoft Foundry (55–60%). Microsoft also expects AI-901 candidates to know Python coding syntax, so more than half of the exam is about building small apps and agents with Foundry.
Can I still use AI-900 practice tests for AI-901?
Only partly. The concept questions on responsible AI, text analysis, speech and computer vision still help with the first skill area. AI-900 sets don't cover deploying models, prompts, agents and Content Understanding in Microsoft Foundry, which is 55–60% of AI-901.
How long is the AI-901 exam and what score do you need to pass?
Microsoft Fundamentals exams last 45 minutes, with 65 minutes of seat time. Scores are reported on a scale of 1 to 1,000 and you need 700 or more to pass. It is a scaled score, so 700 is not the same as 70% of the questions (Microsoft Learn, checked 2026-10-26).
How many questions are on AI-901?
Microsoft does not publish the number of questions for AI-901. It says most of its certification exams have 40 to 60 questions and that the number can change. There is no penalty for guessing, so answer every question.
Are these real AI-901 exam questions?
No. TokIQ Editorial wrote all 10 questions from the skills listed in the official AI-901 study guide. Microsoft keeps its exam content confidential, so sites selling real questions are a risk, not a shortcut. Microsoft also offers its own free practice assessment for this certification on AI Skills Navigator (sign-in required).
- #AI-901
- #AI-900
- #Azure AI Fundamentals
- #practice questions
- #certification prep