CertifyAIxAce every AI certification.

AI-901 sample questions with answers

10 free practice questions for the Microsoft Azure AI Fundamentals (AI-901) exam. Try each one, then open the answer to see why the right option wins and every other option loses.

Question 1Identify AI concepts and capabilities

An engineer is investigating a harmful completion that reached a user. The call returned HTTP 200, the finish reason was length, and the content filter results object on the response carries an error entry instead of category results. What is the most likely explanation?

  1. A.

    The filtering system could not complete in time, so the request finished without content filtering applied.

  2. B.

    The user's account was blocked for repeated misuse and the error entry records that action.

  3. C.

    The prompt itself was blocked, which is what the error entry on the response records.

  4. D.

    The deployment has no content filter attached, so completions are never annotated at all.

Show answer

Answer: A

An error object inside the content filter results means the filtering system did not run, so the request completed unfiltered with a normal status code.

  • A. A filtering outage lets the request complete unfiltered and is reported as an error object inside the content filter results.
  • B. Blocking an abusive user is an account action taken by the application or platform, not a per-response filter annotation.
  • C. A blocked prompt returns HTTP 400 with a content filter error code; this call returned 200 with a completion.
  • D. A deployment without a filter produces no annotations at all rather than an error entry on this particular response.
Question 2Identify AI concepts and capabilities

During a procurement review, a supplier states that nothing can be said about how its model reaches an answer, so it will provide no documentation of the system's behaviour. Which response is best supported by Microsoft's guidance on transparency?

  1. A.

    Transparency asks for a useful account of behaviour, limitations and appropriate uses, not for the model's internals.

  2. B.

    The buyer should rely on the content filter, which makes documentation of behaviour unnecessary.

  3. C.

    The buyer should instead require the training data and the model weights, which transparency defines as the deliverable.

  4. D.

    The position is acceptable, because the behaviour of a deep learning model genuinely cannot be described.

Show answer

Answer: A

Transparency is satisfied by documenting behaviour, limitations and appropriate uses, so a claim that the model is unexplainable does not excuse providing nothing.

  • A. Documented transparency practice is an account of behaviour, limitations and appropriate uses rather than model internals.
  • B. A content filter is a runtime safety control and provides no information about capabilities or limits.
  • C. Transparency does not define weights or training data as the deliverable, and neither would answer the buyer's question.
  • D. Behaviour, limitations and intended uses can be documented even where internal mechanics are hard to interpret.
Question 3Identify AI concepts and capabilities

A clinic is building a Microsoft Foundry assistant that drafts referral letters from clinician intake notes. The notes contain patient names, dates of birth and record numbers. The compliance lead asks for the measure that most directly upholds the privacy and security principle. What should you recommend?

  1. A.

    Use a provisioned throughput deployment so that capacity is reserved for the clinic

  2. B.

    Set the deployment temperature to 0 so that the assistant produces the same letter each time

  3. C.

    Publish a transparency note describing the assistant's intended uses and known limitations

  4. D.

    Increase the maximum output tokens so that a complete letter is always returned

  5. E.

    Remove or mask personal identifiers before the notes are sent to the model, and restrict access to the deployment by using Microsoft Entra ID role assignments

Show answer

Answer: E

Minimising the personal data sent to the model and gating the deployment with Entra ID directly serves privacy and security.

  • A. Provisioned throughput reserves capacity for predictable latency and cost, not for data protection.
  • B. Temperature controls how varied the wording is; it has no effect on who sees personal data.
  • C. A transparency note documents intended use and limitations, which evidences transparency rather than privacy.
  • D. Maximum output tokens only bounds reply length and does nothing to protect identifiers in the prompt.
  • E. Data minimisation plus Entra ID role assignments reduces personal data exposure and limits who can call the deployment.
Question 4Identify AI concepts and capabilities

An insurer is building an application in which a claims adjuster uploads a photograph of vehicle damage and then asks questions about it in plain language, receiving written answers. You must pick a model from the Microsoft Foundry model catalog. Which type of model should you deploy?

  1. A.

    A text-to-speech neural voice that reads text aloud

  2. B.

    A multimodal chat model that accepts image and text input and returns text

  3. C.

    A speech recognition model that transcribes recorded audio

  4. D.

    An image generation model that creates images from a text description

  5. E.

    A text embedding model that converts text into numeric vectors

Show answer

Answer: B

The workload sends an image plus a question and expects text back, which only a multimodal chat model does.

  • A. Text to speech only renders audio from text and cannot inspect an uploaded photograph.
  • B. Image plus text in and text out is the signature of a multimodal chat completion model.
  • C. Speech recognition transcribes audio; there is no audio in this workload and no image understanding.
  • D. Image generation runs the wrong direction: it creates a picture from text instead of interpreting one.
  • E. Embedding models return vectors for similarity search and cannot produce a natural-language answer.
Question 5Identify AI concepts and capabilities

A haulage company records driver debrief calls. Safety officers need each recording turned into searchable text, with each speaker's turns identified separately, so that they can find the point in a call where an incident was discussed. Which capability meets the requirement?

  1. A.

    Sentiment analysis applied to the recorded audio

  2. B.

    Speech recognition, with speaker separation applied to the transcript

  3. C.

    Text translation of the call summary into the officers' language

  4. D.

    Speech synthesis using a selected neural voice

Show answer

Answer: B

Converting recorded audio to searchable text with speaker turns is speech recognition with speaker separation.

  • A. Sentiment analysis needs text as input and reports tone, not the searchable content of the call.
  • B. Speech to text creates the searchable transcript and speaker separation attributes each turn to a speaker.
  • C. Translation converts text between languages and cannot produce a transcript from audio.
  • D. Speech synthesis generates audio from text, which is the opposite of transcribing an existing recording.
Question 6Identify AI concepts and capabilities

A picture agency holds 200,000 untagged photographs. Editors must find pictures by typing a phrase such as "cyclists in heavy rain", and nobody will be adding tags or metadata to the library. Which capability supports this?

  1. A.

    Multimodal embeddings, which place images and text queries in the same vector space so a phrase can be matched to pictures by meaning.

  2. B.

    A custom image-classification model trained on the agency's own labelled examples of each category editors search for.

  3. C.

    Image tagging, which applies labels drawn from a large set of recognisable objects, living things, scenery and actions.

  4. D.

    Optical character recognition, which extracts any text that happens to appear inside the pictures themselves.

Show answer

Answer: A

Multimodal embeddings vectorise images and text into one space, so a text phrase retrieves pictures without any tags being stored.

  • A. Images and text share one vector space, so a typed phrase retrieves semantically close pictures with no tags stored.
  • B. A custom classifier needs labelled examples for every category and retraining whenever editors want a new one.
  • C. Tagging produces labels from a fixed vocabulary that someone must store and maintain, which the agency has ruled out.
  • D. Text recognition only finds characters printed inside a picture, which is absent from most photographs.
Question 7Identify AI concepts and capabilities

A water company photographs domestic meters and needs the reading shown on each meter recorded against the customer account. The team trained an image classification model on thousands of labelled meter photographs. In testing the model returns one of ten category labels for every photograph and never returns the digits. What is the underlying mistake?

  1. A.

    The photographs were not cropped to the dial before classification, so the model saw too much background around the digits that it should have read.

  2. B.

    Image classification cannot run on photographs taken outdoors, where lighting varies between readings.

  3. C.

    An extraction problem was attacked with classification: the reading has to be read out of the image, which is what optical character recognition does.

  4. D.

    The classification model was trained on too few photographs, so its confidence on each category is low.

Show answer

Answer: C

Reading characters out of a picture is an extraction task; classification can only choose a label, so the workload family was chosen wrongly.

  • A. Cropping helps a recogniser find the digits, but the classifier would still return a category label rather than the digits themselves.
  • B. Classification runs on outdoor photographs perfectly well; lighting affects accuracy, not the kind of output produced.
  • C. Classification returns a label from a fixed set; the reading has to be extracted from the pixels, which is optical character recognition.
  • D. More training data improves accuracy within the label set; it cannot make a classifier emit a value that is not one of its labels.
Question 8Identify AI concepts and capabilities

A team is reviewing a new speech-driven assistant. Which change most directly serves the inclusiveness principle?

  1. A.

    Encrypt the audio recordings the assistant stores and restrict which administrators are able to retrieve them.

  2. B.

    Publish the assistant's measured word error rate for each language it claims to support, alongside the conditions it was tested in.

  3. C.

    Assign a named owner who is answerable for the assistant's behaviour and who signs off before each new release goes live.

  4. D.

    Ensure the assistant also accepts typed input and presents its responses as text for people who cannot use speech.

Show answer

Answer: D

Inclusiveness asks whether people of all abilities can use the system, which means offering an alternative to speech.

  • A. Encrypting stored audio protects data, which is the privacy and security principle.
  • B. Publishing accuracy figures serves transparency, even though the figures might reveal an inclusiveness gap.
  • C. Naming an owner establishes who is answerable, which is accountability.
  • D. Providing a non-speech path removes a barrier for people who cannot use a speech-only interface.
Question 9Identify AI concepts and capabilities

A lender receives one PDF per application containing several different document types bound together, such as payslips, bank statements and an identity document. Each type must reach the analyzer that knows its fields. Which two capabilities does the pipeline need? (Choose TWO.)

Choose 2.

  1. A.

    Smart cropping, to select the region of interest on each page

  2. B.

    Abstractive summarization, to recap each section in fresh wording

  3. C.

    Diarization, to attribute the content to a speaker

  4. D.

    Segmentation, to divide the file into logical sections

  5. E.

    Classification, to categorise each section by document type and route it

Show answer

Answer: D, E

A bound file must first be split into logical sections and then each section categorised so it can be routed to the right analyzer.

  • A. Smart cropping chooses a region of interest in an image for reframing; it neither splits nor routes documents.
  • B. A recap of each section does not tell the pipeline what type the section is or where its boundaries fall.
  • C. Diarization is an audio feature that attributes transcript turns to speakers and has no role in a document pipeline.
  • D. Segmentation divides the bound file into logical sections, which is the prerequisite for treating each document separately.
  • E. Classification categorises each section by type so it can be routed to the analyzer that knows that type's fields.
Question 10Identify AI concepts and capabilities

Coho Vineyard is setting up a programme to reduce bias in a model that scores job applications. Which TWO actions are documented Microsoft practices for fairness? Each correct answer presents part of the solution.

Choose 2.

  1. A.

    Set the deployment's temperature to zero so that the model returns the same answer every time.

  2. B.

    Test whether the training data represents the groups the system will be used on.

  3. C.

    Build a design team whose members bring different backgrounds, experiences and perspectives to the work.

  4. D.

    Store the applications in the Azure region closest to the applicants.

  5. E.

    Turn on the protected material filter so that the model does not reproduce copyrighted text.

Show answer

Answer: B, C

Attracting a diverse pool of talent to the design team and testing the data for representation are both named in Microsoft's fairness practices.

  • A. A temperature of zero makes output repeatable; a biased model repeats its bias just as reliably.
  • B. Testing that the data represents the groups served is the documented way to identify bias in datasets.
  • C. Attracting a diverse pool of talent so the design team reflects its users is a named fairness practice.
  • D. Region choice affects latency and data residency, not whether applicants are scored equitably.
  • E. The protected material filter addresses copyrighted output, which is unrelated to how groups are treated.

Keep going with 580 more AI-901 questions

Free papers every day, in the real exam formats, with progress by exam domain. Unlock every paper and timed mock exam when you are ready.

Practise freeSee all AI-901 papers
AI-901 sample questions with answers (10 free) · CertifyAIx