Free AI-901 practice exam questions

What the 20 questions cover

Every question here passed the same checks as the full bank: the quoted sentence was found on the live official page, and an independent blind reviewer matched the answer. How we check questions.

The questions

Answers are hidden. Open “Show answer and source” under a question after you have picked yours.

1. Question 1 of 20 · D1: Identify AI concepts and capabilities · ai9-d1-006

   Which statement reflects Microsoft's description of the reliability and safety principle for AI applications?

   • A. Option A: A model is reliable as soon as it scores above 95 percent accuracy on its own training data
   • B. Option B: AI models are probabilistic, so applications must plan for wrong outputs and mitigate the risk
   • C. Option C: Safety is owned by the model provider, so app developers don't need to add their own controls
   • D. Option D: Once a model is deployed, its outputs are deterministic and can be trusted without any review
   Show answer and source

   Answer: B. Reliability and safety start from accepting that models can be wrong and designing mitigations, from confidence thresholds to human review.

   • A (incorrect): Training-set accuracy says little about behavior on new or unexpected inputs, which is where reliability is tested.
   • B (correct): Correct. Microsoft stresses that AI is not infallible and that applications have to account for that.
   • C (incorrect): Microsoft places responsibility on the people building the application to mitigate risks in their context.
   • D (incorrect): Generative and predictive models produce probabilistic outputs; deployment doesn't change that.

   Official sourcelearn.microsoft.com

   AI is based on probabilistic models, it is not infallible.

   learn.microsoft.com/en-us/training/modules/get-started-ai-fundamentals/7-responsible-ai (official page)

2. Question 2 of 20 · D1: Identify AI concepts and capabilities · ai9-d1-016

   A nonprofit designs a benefits advice assistant for people with low literacy, limited mobile data and a range of disabilities, so that nobody is shut out of using it. Which principle is the team primarily applying?

   • A. Option A: Reliability and safety
   • B. Option B: Inclusiveness
   • C. Option C: Fairness
   • D. Option D: Privacy and security
   Show answer and source

   Answer: B. Fairness and inclusiveness are often confused. Fairness looks at outcomes for people who do use the system; inclusiveness looks at who is able to use it in the first place.

   • A (incorrect): Reliability concerns consistent, safe behavior. Designing for low literacy and limited data is about access.
   • B (correct): Correct. Designing so that the solution doesn't exclude some users is how Microsoft defines inclusiveness.
   • C (incorrect): Fairness is about avoiding discriminatory outputs for similar people. The team's focus is on who can access and use the assistant.
   • D (incorrect): Privacy covers protecting personal data. Nothing in the scenario is about how data is stored or shared.

   Official sourcelearn.microsoft.com

   AI developers should strive to ensure that their solutions don't exclude some users.

   learn.microsoft.com/en-us/training/modules/get-started-ai-fundamentals/7-responsible-ai (official page)

3. Question 3 of 20 · D1: Identify AI concepts and capabilities · ai9-d1-027

   An app stores embedding vectors for product terms and needs to measure how semantically similar two terms are. Which calculation does Microsoft describe for this?

   • A. Option A: A comparison of the two terms' positional encodings
   • B. Option B: The number of letters the two terms have in common
   • C. Option C: The cosine similarity of the two embedding vectors
   • D. Option D: The numeric difference between the two token IDs
   Show answer and source

   Answer: C. Embeddings place semantically related tokens in similar directions in vector space. Cosine similarity compares those directions.

   • A (incorrect): Positional encodings record where a token sits in a sequence, not what it means.
   • B (incorrect): Shared spelling isn't shared meaning. 'Dog' and 'puppy' share no letters but are close semantically.
   • C (correct): Correct. Tokens used in similar contexts point in similar directions, and cosine similarity measures that.
   • D (incorrect): Token IDs are just labels assigned in order of appearance. Nearby IDs say nothing about meaning.

   Official sourcelearn.microsoft.com

   We can measure how close tokens are to one another semantically by calculating the cosine similarity of their vectors.

   learn.microsoft.com/en-us/training/modules/fundamentals-generative-ai/3-language-models (official page)

4. Question 4 of 20 · D1: Identify AI concepts and capabilities · ai9-d1-037

   An insurer needs to redact personal details from millions of claim notes. It wants predictable output and doesn't want to design prompts or train a model. What does Microsoft suggest for a well-defined use case like this?

   • A. Option A: Use an image generation model to blur the personal details in scanned copies of the claim notes
   • B. Option B: Fine-tune a general-purpose chat model on examples of redacted claim notes before deploying it
   • C. Option C: Deploy the largest reasoning model available and ask it to redact each note with a detailed prompt
   • D. Option D: Use a Foundry tool built on prebuilt models instead of choosing a general model from the catalog
   Show answer and source

   Answer: D. When the task is well defined, such as PII detection, a prebuilt Foundry tool is often a better fit than a general model.

   • A (incorrect): The notes are text, and image generation isn't a redaction tool.
   • B (incorrect): Fine-tuning is custom modeling, the effort the insurer wants to avoid, and output can still vary.
   • C (incorrect): A general model can redact text, but results vary between calls and the insurer wants predictable output without prompt design.
   • D (correct): Correct. For well-defined tasks, Foundry tools give predictable performance and fast time to value without custom modeling.

   Official sourcelearn.microsoft.com

   Foundry tools are powered by prebuilt models that provide predictable performance

   learn.microsoft.com/en-us/training/modules/get-started-with-generative-ai-and-agents/2-generative-ai-models (official page)

5. Question 5 of 20 · D1: Identify AI concepts and capabilities · ai9-d1-048

   A product team uses a deployed model to write catalog descriptions, but the wording varies a lot every time they rerun the same prompt. They want more consistent output. Which parameter should they lower?

   • A. Option A: Model version
   • B. Option B: Tokens-per-minute limit
   • C. Option C: Temperature
   • D. Option D: Max output tokens
   Show answer and source

   Answer: C. In the playground and in code, temperature, max output tokens and system instructions are the main runtime levers.

   • A (incorrect): Changing the model version may change behavior, but it isn't the control for run-to-run randomness.
   • B (incorrect): The TPM limit controls throughput and rate limiting, not output variation.
   • C (correct): Correct. Temperature controls creativity versus determinism; lowering it makes outputs more consistent.
   • D (incorrect): Max output tokens caps response length. It doesn't make the wording more predictable.

   Official sourcelearn.microsoft.com

   Temperature : controls creativity vs. determinism.

   learn.microsoft.com/en-us/training/modules/get-started-with-generative-ai-and-agents/3-using-generative-ai-models (official page)

6. Question 6 of 20 · D1: Identify AI concepts and capabilities · ai9-d1-058

   A grocery chain mounts cameras on shelves to spot when products are running low and alert staff. Which AI workload does this use?

   • A. Option A: Speech recognition
   • B. Option B: Generative AI
   • C. Option C: Natural language processing
   • D. Option D: Computer vision
   Show answer and source

   Answer: D. Computer vision analyzes photographs, video and live camera feeds, and retail stock monitoring is a common use.

   • A (incorrect): There is no spoken input to transcribe.
   • B (incorrect): The system analyzes what it sees. It doesn't need to generate new content.
   • C (incorrect): NLP analyzes text. The input here is camera images.
   • D (correct): Correct. Monitoring stock levels from camera images is a computer vision scenario Microsoft lists explicitly.

   Official sourcelearn.microsoft.com

   Monitoring stock levels or identifying items for checkout in retail scenarios.

   learn.microsoft.com/en-us/training/modules/get-started-ai-fundamentals/3-computer-vision (official page)

7. Question 7 of 20 · D1: Identify AI concepts and capabilities · ai9-d1-069

   Users say an app's synthesized voice sounds robotic and flat, even though every word is pronounced correctly. Which part of speech synthesis is most likely at fault?

   • A. Option A: Prosody generation
   • B. Option B: Speech recognition
   • C. Option C: Language detection
   • D. Option D: Text normalization
   Show answer and source

   Answer: A. Prosody covers pitch contours, duration, intensity, pauses and stress patterns. Modern systems use transformers to predict it from sentence context.

   • A (correct): Correct. Robotic-sounding speech often comes from flat, monotone prosody rather than imperfect pronunciation.
   • B (incorrect): Speech recognition converts speech to text. It isn't part of the synthesis pipeline.
   • C (incorrect): Detecting the language of input text doesn't affect the rhythm and intonation of the output.
   • D (incorrect): Normalization errors would cause wrongly expanded numbers or abbreviations, not a flat delivery.

   Official sourcelearn.microsoft.com

   Robotic-sounding speech often results from flat, monotone prosody—not from imperfect phoneme pronunciation.

   learn.microsoft.com/en-us/training/modules/introduction-ai-speech/4-speech-synthesis (official page)

8. Question 8 of 20 · D1: Identify AI concepts and capabilities · ai9-d1-079

   Which TWO components make up a typical information extraction solution for scanned documents? Select TWO.

   • A. Option A: A model that maps the extracted values to fields
   • B. Option B: An image model that redraws the original scan
   • C. Option C: A vocoder that turns field values into audio
   • D. Option D: OCR that detects and extracts text from images
   • E. Option E: A speech model that reads the document aloud
   Show answer and source

   Answer: A, D. OCR says what text exists; field mapping says what it means, such as which value is the total.

   • A (correct): Correct. Machine learning or generative AI maps the recognized text to specific data fields.
   • B (incorrect): Regenerating the image doesn't extract any data.
   • C (incorrect): Vocoders generate audio waveforms in text-to-speech.
   • D (correct): Correct. Computer vision detects the text in image-based content.
   • E (incorrect): Speech synthesis has no role in extracting structured data from scans.

   Official sourcelearn.microsoft.com

   Value identification and mapping from the OCR results to data fields.

   learn.microsoft.com/en-us/training/modules/introduction-information-extraction/2-overview (official page)

9. Question 9 of 20 · D2: Implement AI solutions by using Microsoft Foundry · ai9-d2-005

   A developer builds a prompt that mixes instructions, three source documents and a question. Which TWO practices from Microsoft's prompt guidance help the model parse it correctly? Select TWO.

   • A. Option A: Use Markdown or XML when unsure which syntax to choose
   • B. Option B: Remove all headings so the prompt uses fewer tokens
   • C. Option C: Use separators such as --- between sources and steps
   • D. Option D: Put every element on one line to avoid line breaks
   • E. Option E: Randomize the order of the sources on each request
   Show answer and source

   Answer: A, C. Structure helps the model tell parts of the prompt apart and can also provide stopping conditions.

   • A (correct): Correct. Models have been trained on lots of Markdown and XML, which can give better results.
   • B (incorrect): Headings help structure the prompt. Saving a few tokens isn't worth the lost clarity.
   • C (correct): Correct. Clear syntax, such as separators between sections, communicates intent and makes output easier to parse.
   • D (incorrect): Cramming everything together makes the boundaries between parts harder to see.
   • E (incorrect): Shuffling sources adds noise. It doesn't help the model tell instructions from content.

   Official sourcelearn.microsoft.com

   separators ( --- in this case) have been added between different sources of information or steps.

   learn.microsoft.com/en-us/azure/foundry/openai/concepts/prompt-engineering (official page)

10. Question 10 of 20 · D2: Implement AI solutions by using Microsoft Foundry · ai9-d2-015

    Which TWO things happen when you deploy a model in Foundry? Select TWO.

    • A. Option A: The model's weights are copied to client devices
    • B. Option B: The model is retrained on your project's files
    • C. Option C: An API endpoint is created for invoking the model
    • D. Option D: Monitoring and logging of usage and latency begin
    • E. Option E: A new Foundry resource is created for the model
    Show answer and source

    Answer: C, D. Deployment also allocates compute and locks in configuration such as model version and safety settings.

    • A (incorrect): The model runs on Foundry-allocated compute, not on client devices.
    • B (incorrect): Deployment doesn't train the model. Fine-tuning is a separate process.
    • C (correct): Correct. Deployment creates an endpoint so applications can call the model securely.
    • D (correct): Correct. Usage metrics, performance, latency, errors and costs are tracked once the model is deployed.
    • E (incorrect): Deployments are created inside an existing resource and project.

    Official sourcelearn.microsoft.com

    Monitoring and logging become active: usage metrics, performance, latency, errors, and costs are tracked

    learn.microsoft.com/en-us/training/modules/get-started-with-generative-ai-and-agents/2-generative-ai-models (official page)

11. Question 11 of 20 · D2: Implement AI solutions by using Microsoft Foundry · ai9-d2-025

    A developer uses the OpenAI Python SDK against a Foundry resource and wants to authenticate with an API key instead of Microsoft Entra ID. What does Microsoft's SDK guidance say?

    • A. Option A: API keys work on the /openai/v1 endpoint; pass the key as api_key
    • B. Option B: API keys are no longer accepted by any Foundry model endpoints
    • C. Option C: API keys only work on the project endpoint used by the Foundry SDK
    • D. Option D: API keys must be placed inside the user prompt for each request
    Show answer and source

    Answer: A. Entra ID through DefaultAzureCredential is recommended, but key authentication remains available on the OpenAI-compatible route.

    • A (correct): Correct. The Foundry SDK samples use Entra ID, but API keys work on /openai/v1 when passed as api_key.
    • B (incorrect): Keys are still supported on /openai/v1, though Entra ID is the default in samples.
    • C (incorrect): The guidance names /openai/v1 as the route where keys work.
    • D (incorrect): Credentials never go in prompts. They're passed to the client.

    Official sourcelearn.microsoft.com

    API keys work on /openai/v1 . Pass the key as api_key instead of a token provider.

    learn.microsoft.com/en-us/azure/foundry/how-to/develop/sdk-overview (official page)

12. Question 12 of 20 · D2: Implement AI solutions by using Microsoft Foundry · ai9-d2-034

    A small team wants its first text-based agent with the least to manage: configure instructions, a model and tools, and let Foundry run it with no code or infrastructure. Which agent type fits?

    • A. Option A: A hosted agent
    • B. Option B: A voice agent
    • C. Option C: A batch agent
    • D. Option D: A prompt agent
    Show answer and source

    Answer: D. Microsoft suggests starting with a prompt agent, either in the Foundry portal or with code.

    • A (incorrect): Hosted agents bring your own code and framework as a container, which is more to manage.
    • B (incorrect): Voice-based prompt agents target real-time voice. The team wants a text agent.
    • C (incorrect): Batch is a deployment type for asynchronous requests, not an agent type.
    • D (correct): Correct. Prompt agents are configured with instructions, a model and tools, and Foundry runs them.

    Official sourcelearn.microsoft.com

    Want the least to manage for text? Start with a prompt agent .

    learn.microsoft.com/en-us/azure/foundry/agents/overview (official page)

13. Question 13 of 20 · D2: Implement AI solutions by using Microsoft Foundry · ai9-d2-044

    Which capability does calling an agent through the Foundry Project API give a client application?

    • A. Option A: Bypassing the agent's guardrails for trusted internal callers
    • B. Option B: Integrating the agent into web apps, bots or backend workflows
    • C. Option C: Running the agent offline without any Foundry deployments
    • D. Option D: Changing the base model's weights for every request it sends
    Show answer and source

    Answer: B. The Project API also supports orchestrating multi-step tasks and passing structured inputs or tool calls.

    • A (incorrect): Calling from code doesn't bypass guardrails.
    • B (correct): Correct. The Project API lets you integrate agents into apps and workflows and run them at scale.
    • C (incorrect): Agents run on your Foundry deployments.
    • D (incorrect): Calling an agent doesn't modify model weights.

    Official sourcelearn.microsoft.com

    Integrate agents into web apps, bots, or backend workflows

    learn.microsoft.com/en-us/training/modules/get-started-with-generative-ai-and-agents/4-creating-an-agent (official page)

14. Question 14 of 20 · D2: Implement AI solutions by using Microsoft Foundry · ai9-d2-054

    An agent already uses a capable language model. Why might a team still connect Azure Language to it through MCP?

    • A. Option A: The model by itself can't perform deterministic, structured text analysis
    • B. Option B: The model can't read text at all unless Azure Language converts it first
    • C. Option C: MCP replaces the model's instructions with the Language service's own rules
    • D. Option D: Azure Language makes the model respond faster by caching its earlier answers
    Show answer and source

    Answer: A. The agent combines the reasoning of the language model with the precision of Azure Language, for example to route tickets by detected language.

    • A (correct): Correct. Adding Azure Language gives the agent consistent and predictable text analysis.
    • B (incorrect): Language models read text directly. The value is deterministic analysis.
    • C (incorrect): The agent keeps its instructions and calls the Language tool when relevant.
    • D (incorrect): Azure Language is an analysis tool, not a response cache.

    Official sourcelearn.microsoft.com

    that model alone can't perform text analysis tasks that require deterministic, structured analysis.

    learn.microsoft.com/en-us/training/modules/get-started-text-analysis-azure/4-language-mcp (official page)

15. Question 15 of 20 · D2: Implement AI solutions by using Microsoft Foundry · ai9-d2-063

    A media company stores 20,000 recorded calls in Azure Blob Storage and needs transcripts. Results aren't needed in real time. Which Azure Speech approach fits?

    • A. Option A: Text to speech with SSML, which turns each of the recordings into text
    • B. Option B: Batch transcription, with the files referenced by a SAS URI or a container
    • C. Option C: Real-time speech to text from a microphone stream, one recording at a time
    • D. Option D: Language identification on its own, which produces the full transcripts
    Show answer and source

    Answer: B. Batch jobs are scheduled on a best-effort basis; at peak times they can take up to 30 minutes to start and up to 24 hours to complete.

    • A (incorrect): Text to speech goes the other way: from text to audio.
    • B (correct): Correct. Batch transcription handles large volumes of stored audio and returns results asynchronously.
    • C (incorrect): Real-time recognition is for live audio, and processing 20,000 files this way is impractical.
    • D (incorrect): Language identification detects which language is spoken. It doesn't produce transcripts.

    Official sourcelearn.microsoft.com

    You can upload your own data or use existing audio files via public URI or shared access signature (SAS) URI.

    learn.microsoft.com/en-us/azure/ai-services/speech-service/batch-transcription (official page)

16. Question 16 of 20 · D2: Implement AI solutions by using Microsoft Foundry · ai9-d2-073

    An insurer currently runs separate OCR, object detection and captioning services, then a language model, to process claim photos. What is the main advantage of switching to a multimodal model?

    • A. Option A: It removes the need for any responsible AI review, since the model is pretrained
    • B. Option B: It accepts text and images together, reducing the need for separate vision pipelines
    • C. Option C: It runs fully offline, so no Foundry deployment or endpoint is needed anymore
    • D. Option D: It guarantees exact, identical outputs every time, unlike the prebuilt vision services
    Show answer and source

    Answer: B. Vision-enabled models allow flexible, general-purpose visual reasoning without deep computer vision expertise.

    • A (incorrect): Responsible AI review still applies to multimodal solutions.
    • B (correct): Correct. Multimodal models reduce the need for separate vision pipelines and simplify end-to-end experiences.
    • C (incorrect): The model still runs on a Foundry deployment that the app calls.
    • D (incorrect): Generative models are probabilistic. They don't guarantee identical outputs.

    Official sourcelearn.microsoft.com

    they reduce the need for separate vision pipelines and make it easier to build end-to-end intelligent experiences.

    learn.microsoft.com/en-us/training/modules/get-started-vision-azure/2-vision-enabled-models (official page)

17. Question 17 of 20 · D2: Implement AI solutions by using Microsoft Foundry · ai9-d2-083

    A user asks an image generation app for violent graphic content, and no image is produced. What is the most likely reason?

    • A. Option A: The deployment ran out of tokens-per-minute quota partway through the request
    • B. Option B: The model only produces images when a URL output option has been requested
    • C. Option C: The Image APIs include a content moderation filter that blocks harmful prompts
    • D. Option D: The prompt was too short for the model to understand what it was asked for
    Show answer and source

    Answer: C. The error code returned for this case is content_policy_violation; modify the prompt to comply with the content policy.

    • A (incorrect): Quota problems produce rate-limit errors, not content policy blocks.
    • B (incorrect): Output format has nothing to do with whether harmful prompts are blocked.
    • C (correct): Correct. If the service recognizes the prompt as harmful, it doesn't generate an image.
    • D (incorrect): Short prompts still produce images. This is a content policy decision.

    Official sourcelearn.microsoft.com

    If the service recognizes your prompt as harmful content, it doesn't generate an image.

    learn.microsoft.com/en-us/azure/foundry/openai/how-to/dall-e (official page)

18. Question 18 of 20 · D2: Implement AI solutions by using Microsoft Foundry · ai9-d2-092

    Why does a team define its extraction logic in a Content Understanding analyzer rather than in ad hoc prompts for each document?

    • A. Option A: Analyzers remove the need for a schema, since they infer fields every time
    • B. Option B: Analyzers apply the same logic consistently and produce predictable JSON
    • C. Option C: Analyzers return free-form prose answers that change with each document's content
    • D. Option D: Analyzers don't need a Foundry resource, so they cost nothing to run
    Show answer and source

    Answer: B. Predictable results make downstream storage, search and automation easier.

    • A (incorrect): An analyzer includes a schema that defines the fields and structure.
    • B (correct): Correct. Once configured, an analyzer reuses the schema consistently for every request.
    • C (incorrect): Analyzers produce predictable structured JSON, not free-form prose.
    • D (incorrect): Content Understanding is part of a Foundry resource and has its own pricing.

    Official sourcelearn.microsoft.com

    Analyzers consistently apply the same extraction logic to all incoming content.

    learn.microsoft.com/en-us/training/modules/get-started-information-extraction/2-documents (official page)

19. Question 19 of 20 · D2: Implement AI solutions by using Microsoft Foundry · ai9-d2-102

    A retail chain wants to detect, count and extract details about products on store shelves from photos, returning the same fields every time for its inventory system. What fits best?

    • A. Option A: An embeddings model applied to each of the shelf photographs
    • B. Option B: A Content Understanding image analyzer for shelf analysis
    • C. Option C: A chat with a vision model, asking it new questions each time
    • D. Option D: A batch transcription job applied to the store's camera footage
    Show answer and source

    Answer: B. Use analyzers when you need consistent, schema-based output; use vision chat for open-ended questions.

    • A (incorrect): Embeddings support similarity search, not counting products.
    • B (correct): Correct. Shelf analysis and inventory management is a listed image analyzer scenario with structured results.
    • C (incorrect): A multimodal chat can describe shelves, but free-form answers vary and don't fit a fixed inventory schema.
    • D (incorrect): Transcription converts audio to text.

    Official sourcelearn.microsoft.com

    Detect, count, and extract specific details about retail products

    learn.microsoft.com/en-us/azure/ai-services/content-understanding/image/overview (official page)

20. Question 20 of 20 · D2: Implement AI solutions by using Microsoft Foundry · ai9-d2-112

    A developer inspects the JSON returned by the prebuilt-invoice analyzer. Which of the following does Microsoft say it shows?

    • A. Option A: The markdown, the fields with their data, and confidence scores
    • B. Option B: A generated image of the invoice with the key fields highlighted
    • C. Option C: A spoken audio summary of the invoice's totals and due date
    • D. Option D: Only a plain-text transcript, with no structure or scores included
    Show answer and source

    Answer: A. In code, result.contents holds markdown and fields for each content item.

    • A (correct): Correct. The output shows the markdown, fields, field data and confidence scores.
    • B (incorrect): Analyzers return JSON, not images.
    • C (incorrect): Analyzers don't produce audio.
    • D (incorrect): The result is structured JSON, not just plain text.

    Official sourcelearn.microsoft.com

    The resulting output is JSON that shows the extracted markdown, fields, data in the fields, and confidence score.

    learn.microsoft.com/en-us/training/modules/get-started-information-extraction/2-documents (official page)