Explore / Microsoft Azure AI Fundamentals (AI-901) Exam Prep / Practice questions
Microsoft Azure AI Fundamentals (AI-901) practice questions
10 questions from the bank behind this course's timed mock. Free to read and to answer, no account needed, and each has its answer and an explanation. They belong to Microsoft Azure AI Fundamentals (AI-901) Exam Prep.
Answer them one at a time →Read lesson 1 free
Question 1
A learner takes the AI-901 practice run and wants to know what the real exam requires. Their colleague says AI-901 is scored as a percentage and that 70% of questions must be correct. According to the exam facts covered in this course, which statement is accurate?
- Scores run from 0 to 100, and a score of 70 or greater is required to pass
- Scores run from 1 to 1000, and a score of 700 or greater is required to pass
- Scores run from 1 to 1000, and a score of 800 or greater is required to pass
- The exam is pass/fail with no numeric score, based on 70% of questions answered correctly
Show the answer
Answer: B. Scores run from 1 to 1000, and a score of 700 or greater is required to pass
AI-901 is scored from 1 to 1000 and the pass mark is 700 or greater. The 0–100 option is the nearest distractor, but it confuses a scaled score with a simple percentage of correct answers.
Question 2
A product manager is about to roll out an Azure-based speech transcription service. Before launch, she wants to read Microsoft's own write-up of how the service works, what it is intended for, and where it is known to fall short. Which Microsoft resource is designed for this, and which principle does it support?
- An internal review board report, which supports accountability
- An accessibility conformance statement, which supports inclusiveness
- A data encryption policy, which supports privacy and security
- A Transparency Note, which supports transparency
Show the answer
Answer: D. A Transparency Note, which supports transparency
Microsoft publishes Transparency Notes so customers can see how its AI services work and what their known limitations are. Publishing limitations is a transparency practice. A review board is an accountability mechanism that oversees development. It is not the document that explains a service's limits to the people using it.
Question 3
A market research analyst must classify the sentiment of 500,000 archived support tickets with a Foundry model. Nobody needs the results until next week, and the main goal is to keep the cost as low as possible. Which deployment type should the analyst choose?
- Global Provisioned
- Global Standard
- Developer
- Global Batch
Show the answer
Answer: D. Global Batch
Batch deployments are built for large asynchronous jobs. They cost 50% less than standard pricing and aim to finish within 24 hours. Global Standard would also work, but it charges full pay-per-token prices for speed this job doesn't need.
Question 4
A contact-centre manager has an archive of 40,000 call recordings from last year. She wants all of them turned into searchable text, and nobody needs results while a call is still happening. Which Azure Speech in Foundry Tools feature fits best?
- Text to speech with a custom neural voice
- Read text (OCR) in Azure Vision in Foundry Tools
- Batch transcription in speech to text
- Real-time transcription in speech to text
Show the answer
Answer: C. Batch transcription in speech to text
Batch transcription is built for large volumes of prerecorded audio. Real-time transcription is also speech to text, but it streams results while someone is still talking, and that isn't needed here.
Question 5
A legal operations analyst must convert 20,000 scanned contracts into searchable text that keeps the tables and section structure. No named fields are needed, and the budget is tight. Which Azure Content Understanding analyzer is the best fit?
- A custom analyzer based on prebuilt-document with a field schema
- prebuilt-invoice
- prebuilt-layout
- prebuilt-audioSearch
Show the answer
Answer: C. prebuilt-layout
prebuilt-layout returns OCR text plus tables, figures and sections without calling a language model, so it is the cheaper choice when only text and structure are needed. A custom analyzer with a field schema would also work, but it calls generative model deployments and costs tokens for fields nobody asked for.
Question 6
An online retailer runs a product-description generator. It sends a steady, high volume of requests all day, and the operations lead needs guaranteed capacity with more consistent latency than pay-per-token offers. Which deployment type should the AI engineer choose in the Foundry portal?
- Data Zone Standard
- Provisioned
- Global Standard
- Standard
Show the answer
Answer: B. Provisioned
Provisioned deployments reserve throughput, which gives guaranteed capacity and more consistent latency for steady high volume. Global Standard is the pay-per-token default and has the broadest model availability, but it doesn't reserve capacity for you.
Question 7
A travel company is building an agent in the Foundry portal. Customers will ask it about current flight disruptions and today's airport news, which change hour by hour. The answers must include citations. Which addition best meets this need?
- File search with an uploaded PDF of airport information
- Code interpreter
- Web search
- Azure AI Search connected to last month's index of travel articles
Show the answer
Answer: C. Web search
Web search is a knowledge source that grounds answers in the current public web and gives citations, which suits information that changes all the time. File search and an Azure AI Search index only hold the content you loaded into them, so they can't keep up with news that changes by the hour.
Question 8
A support engineer is building a Python console client for an existing Foundry agent. The call `openai.responses.create(conversation=conv.id, input=user_text)` works, but the script prints a large object instead of the agent's answer. The engineer wants to print only the reply text. What should the script print?
- response.instructions
- conversation.id
- response.input
- response.output_text
Show the answer
Answer: D. response.output_text
The agent's reply text is in `response.output_text`. `conversation.id` only identifies the server-side conversation the turn belongs to. It holds no reply text.
Question 9
A developer is building a helpdesk voice bot for callers in the UK. They set speech_config.speech_synthesis_voice_name = "en-GB-SoniaNeural" and expect British callers to be transcribed correctly. Transcripts of British speakers are poor, but the code raises no errors. What is the best fix?
- Replace recognize_once_async() with continuous recognition
- Change the voice to speech_synthesis_voice_name = "en-GB-RyanNeural"
- Set speech_config.speech_recognition_language = "en-GB" on the SpeechConfig used by the SpeechRecognizer
- Send the audio as an input_audio content part with format set to "wav"
Show the answer
Answer: C. Set speech_config.speech_recognition_language = "en-GB" on the SpeechConfig used by the SpeechRecognizer
The voice name only affects synthesis. The language the recogniser expects comes from speech_recognition_language, and a mismatch lowers accuracy without raising an error. Picking a different en-GB voice changes the speaker output and does nothing for recognition.
Question 10
A records team at a law firm is digitizing 40,000 scanned archive pages. They need only the OCR text and the table structure from each page, and their compliance lead has said that no language model deployment may be involved. Which Content Understanding analyzer best fits?
- prebuilt-documentSearch
- prebuilt-layout
- prebuilt-document
- prebuilt-contract
Show the answer
Answer: B. prebuilt-layout
prebuilt-layout is a content extraction analyzer that returns OCR text, tables and layout without needing a language model. prebuilt-documentSearch also returns document content, but it produces markdown and summaries for search, and that kind of output relies on a model.
The full timed mock
50 questions in 45 minutes, with a pass mark of 70%. Each paper is drawn from a bank of questions written for the mock alone, and these 10 are left out of it. Full access opens it, with every level of the course.
LevelUpWith is independent of the organisation that sets this exam. These questions were written for this course; they are not taken from the exam.