Microsoft Azure & FabricLive · 102 questions · 11 notes

AI-901: Microsoft Azure AI Fundamentals

Free practice exam, study notes with architecture diagrams and the official blueprint for Microsoft Azure AI Fundamentals.

Passing score
700 / 1000
Domains
2
Exam code
AI-901

Official AI-901 exam blueprint

Domains and weights from Microsoft Azure & Fabric's exam guide.

DomainWeight
Identify AI concepts and capabilities40–45%
Implement AI solutions by using Microsoft Foundry55–60%

Free AI-901 sample questions

The first questions of the free diagnostic. Pick your answer, then check it. Why every option is right or wrong, and the What-If twist, open in the workspace.

  1. Question 1 · Identify AI workloads

    A retailer's support team must route the angriest product reviews to a senior agent within an hour of posting. A developer wires the review feed to Azure Language in Foundry Tools, and the router receives this result for each new review:

    JSON
    {
      "id": "1",
      "sentiment": "negative",
      "confidenceScores": {"positive": 0.02, "neutral": 0.05, "negative": 0.93},
      "sentences": [
        {"text": "The battery dies by noon.", "sentiment": "negative"}
      ]
    }

    What does this output measure?

    1. AThe key phrases that summarize the review
    2. BThe named entities, such as products and places, in the review
    3. CThe opinion the review expresses, as sentiment labels with confidence scores
    4. DThe language the review is written in
    Check the answer

    Answer: C. Sentiment analysis labels text positive, neutral, negative or mixed with confidence scores at document and sentence level.

    Official docs: Microsoft Learn: Sentiment analysis and opinion mining in Azure Language
    Unlock Full 4-Option Autopsy & What-If Twist in Workspace
  2. Question 2 · Implement generative AI apps and agents by using Foundry

    A router maker's support app must refuse questions about rival brands and only discuss its own products. Before the release ships, a reviewer who has never built a chat app asks what the first message in this request is for:

    JSON
    {
      "messages": [
        {"role": "system", "content": "You are Contoso's router support assistant. Answer only questions about Contoso routers, in under 100 words."},
        {"role": "user", "content": "How do I reset my router?"}
      ],
      "temperature": 0.2
    }

    What does the system message do in this request?

    1. AIt sets the assistant's role, scope and response limits for the whole conversation
    2. BIt stores the user's earlier questions so the model can remember them
    3. CIt retrains the model's weights on Contoso's router manuals
    4. DIt controls how random the model's wording is
    Check the answer

    Answer: A. System messages define the assistant's role, scope, format and rules; user messages carry the request; temperature sets randomness.

    Official docs: Microsoft Learn: System message design
    Unlock Full 4-Option Autopsy & What-If Twist in Workspace
  3. Question 3 · Implement generative AI apps and agents by using Foundry

    A support app built with the Foundry SDK must start calling a newer model deployment named gpt-5-mini. It keeps the same Foundry project and the same Microsoft Entra sign-in, and only one line may change in this release:

    Python
    from azure.ai.projects import AIProjectClient
    from azure.identity import DefaultAzureCredential
    
    project = AIProjectClient(
        endpoint="https://contoso.services.ai.azure.com/api/projects/support",
        credential=DefaultAzureCredential(),
    )
    openai = project.get_openai_client()
    response = openai.responses.create(model="gpt-4.1-mini", input="Summarize ticket 4411")
    print(response.output_text)

    Which value must the team change?

    1. AThe model argument, from gpt-4.1-mini to gpt-5-mini
    2. BThe endpoint of the project
    3. CThe credential, from DefaultAzureCredential to an API key
    4. DThe input text of the request
    Check the answer

    Answer: A. In the Foundry SDK, the project endpoint selects the project, the credential signs in, and the model argument selects the deployment.

    Official docs: Microsoft Learn: Get started with Microsoft Foundry SDKs and endpoints
    Unlock Full 4-Option Autopsy & What-If Twist in Workspace
Answer all 10 free AI-901 questions

AI-901 study notes: key concepts

What each note covers. The full notes, with summaries, examples and memory hooks, open with a free account.

Architecture flow · Identify AI workloads

Where each AI workload runs in Microsoft Foundry

Business app or chat client

  • leads to Foundry Agent Service prompt agents · HTTPS 443 requests
  • leads to Azure Language for text analysis · HTTPS 443 requests
  • leads to gpt-image models for image generation

Microsoft Foundry resource

Foundry Agent Service prompt agents

  • leads to Multimodal chat and reasoning models
  • leads to Azure Content Understanding for extraction

Model deployments

Multimodal chat and reasoning models

gpt-image models for image generation

Foundry Tools

Azure Language for text analysis

Azure Speech for speech workloads

  • leads to Multimodal chat and reasoning models

Azure Content Understanding for extraction

Drawn fromMicrosoft Learn: What are Foundry Tools?

Identify AI workloads

  • Generative AI creates new text, code or images from a prompt, while agentic AI adds planning and tool calls so an agent can complete multi-step tasks.
  • Text analysis workloads extract sentiment, key phrases, entities and summaries from written language, and Azure Language in Foundry Tools offers these as features.
  • Speech workloads cover speech recognition, which turns spoken audio into text, and speech synthesis, which turns text into natural-sounding spoken audio.
  • Computer vision interprets existing images and video, while image generation models such as the gpt-image series create entirely new pictures from text prompts.
  • Information extraction turns unstructured documents, images, audio and video into structured fields that downstream systems such as finance or claims apps can use directly.

Describe principles of responsible AI

  • Fairness means people with similar circumstances receive similar outcomes, so unequal approval or shortlist rates between groups are a fairness problem to investigate.
  • Reliability and safety means testing systems on unusual and edge cases and designing them to fail safely, such as handing over to a person.
  • Privacy and security means protecting personal data and resisting attacks, using real controls such as least-privilege access and redaction rather than prompt instructions.
  • Inclusiveness means designing for everyone, including people with disabilities, and involving them in design and testing so the system works for them.
  • Transparency and accountability differ: transparency helps people understand a system, while accountability makes named people answerable with authority to review and override.

Identify AI model components and configurations

  • Large language models split text into tokens, words or parts of words, and both prompts and replies are counted and billed in tokens.
  • Generation is repeated next-token prediction from the prompt and the tokens already written, which explains varied wording and hallucinations without grounding.
  • Embeddings are vectors that represent meaning, so vector search can find passages with similar meaning even when they share no exact words.
  • Retrieval-augmented generation retrieves relevant passages and adds them to the prompt, grounding answers in trusted data while keeping each prompt small.
  • Chat models generate text, embedding models produce vectors, and the two often work together in one RAG app built on Azure AI Search.

Identify AI model components and configurations

  • Pick a model by modality and task: multimodal for images with text, reasoning for complex problems, small models for fast simple classification.
  • Global Standard is Microsoft's recommended starting deployment type, while Data Zone types keep processing inside the US, EU or APAC data zone.
  • Provisioned deployment types reserve throughput with lower latency variance, batch types discount asynchronous bulk jobs, and Developer suits fine-tuned model evaluation.
  • Temperature and top_p control randomness, max_tokens limits reply length, and reasoning models use max_completion_tokens without temperature or top_p.
  • Each deployment's tokens-per-minute rate limit comes from the subscription's regional quota, and requests beyond it are throttled with 429 errors.

Identify AI workloads

  • Sentiment analysis labels text positive, neutral, negative or mixed with confidence scores, and opinion mining links each opinion to the aspect it describes.
  • Key phrase extraction returns uncategorized talking points, while entity recognition returns items such as people, places and dates together with their categories.
  • Speech recognition turns spoken audio into text, speech synthesis turns text into speech, and speech translation transcribes and translates live speech in one step.
  • Image classification labels a whole image, object detection locates each object with a bounding box, and OCR reads text printed in an image.
  • Information extraction returns structured fields from documents, images, audio and video, so systems can use the values without anyone retyping them.

Implement AI solutions with computer vision and image-generation capabilities by using Foundry

  • Multimodal chat models take images as image_url content parts, as a link or inline base64 data, together with text, and answer in text.
  • The detail setting on image input can be low, high or auto, trading processing cost and speed against how finely the image is read.
  • gpt-image models generate images from prompts, with size for resolution, n for how many images and quality for detail versus speed.
  • Face identification and verification are Limited Access features that need an approved application, and liveness detection is the anti-spoofing check against printed photos.
  • The Image Analysis API retires, so OCR moves to Document Intelligence read, faces to the Face API and captions to GPT models.
Read the full AI-901 study notes

AI-901 FAQ

What is the passing score for AI-901?

The passing score for Microsoft Azure AI Fundamentals is 700 / 1000.

Which domains does AI-901 cover?

The official exam guide lists 2 domains: Identify AI concepts and capabilities (40–45%); Implement AI solutions by using Microsoft Foundry (55–60%).

Is there a free AI-901 practice test?

Yes. Gyan AI's free diagnostic is 10 AI-901 questions with no sign-up, each with an explanation for every option and a link to the official documentation. A free account unlocks the full question set and study notes.

Are these the real AI-901 exam questions?

No. Real exam content is confidential under vendor exam agreements, and using leaked "dumps" can get a certification revoked. Gyan AI questions are original and written to the published skills outline.