AI-901 · Sample questions

AI-901 sample questions with answers and explanations

10 real questions from the CertifyDeck AI-901 bank, spread across the domains of the official Microsoft Certified: Azure AI Fundamentals skills outline, with the correct answer and a full explanation of why the other options are wrong. They are original questions written from the public syllabus, not exam dumps, and they are the same style you will find in the practice tests and mock exams.

1. Question 1Domain: Implement AI solutions by using Microsoft FoundryMedium

   A script calls project.agents.create_version(agent_name='MyAgent', definition=PromptAgentDefinition(model='gpt-demo', instructions='Answer briefly')). What happens if MyAgent already exists?

   • A.A new version of MyAgent is created with that definition
   • B.The call fails because agent names must be unique and fixed
   • C.The existing agent is deleted and recreated from scratch
   • D.The model deployment gpt-demo is replaced by a new one
   Show answer and explanation

   Correct answer: A. A new version of MyAgent is created with that definition

   Explanation

   create_version creates the agent, or a new version of it if it already exists, defined here by a PromptAgentDefinition with a model and instructions. It does not delete the agent or touch the model deployment, and it does not fail merely because the name exists.

   Official references

   • Quickstart: Create a prompt agent
2. Question 2Domain: Implement AI solutions by using Microsoft FoundryMedium

   A developer wants the model to tag support tickets with the team's own label style, but plain instructions keep producing different wording. Which technique shows the desired behavior inside the prompt?

   • A.Increasing the max output tokens
   • B.Adding a few example input and output pairs
   • C.Raising the temperature value
   • D.Placing the request at the end of a long prompt
   Show answer and explanation

   Correct answer: B. Adding a few example input and output pairs

   Explanation

   Few-shot prompting includes one or more input and output examples that condition the model to respond as desired for the current request, without changing the model. A higher temperature would make wording more random, and a larger token cap only allows longer replies.

   Official references

   • Prompt engineering techniques - Microsoft Foundry
3. Question 3Domain: Implement AI solutions by using Microsoft FoundryMedium

   A bank's procurement team requires catalog models hosted by Microsoft under Microsoft Product Terms, with enterprise-grade SLAs. Which Foundry model category fits?

   • A.Models from partners and community, hosted by their vendors
   • B.Fine-tuned models on the Developer deployment type
   • C.Models sold directly by Azure, hosted by Microsoft
   • D.Open-source models that run on managed compute
   Show answer and explanation

   Correct answer: C. Models sold directly by Azure, hosted by Microsoft

   Explanation

   Models sold directly by Azure are hosted by Microsoft under Microsoft Product Terms and come with enterprise-grade SLAs. Partner and community models are often vendor-hosted, and the Developer deployment type is for short-lived fine-tuned model evaluation and has no SLA.

   Official references

   • Generative AI models (Get started with generative AI and agents in Azure)
   • Understanding deployment types in Microsoft Foundry Models
4. Question 4Domain: Implement AI solutions by using Microsoft FoundryMedium

   A client runs: openai = project.get_openai_client(agent_name='HelpBot'); c = openai.conversations.create(); openai.responses.create(conversation=c.id, input='Hi'); openai.responses.create(conversation=c.id, input='And you?'). Why create a conversation?

   • A.To reserve extra token quota for the second request in the same session
   • B.To publish HelpBot as an application that has its own dedicated endpoint
   • C.So the follow-up request keeps the history of the earlier turn in the same session
   • D.To create a new version of HelpBot automatically with each message that is sent
   Show answer and explanation

   Correct answer: C. So the follow-up request keeps the history of the earlier turn in the same session

   Explanation

   Passing the same conversation id to each responses.create call lets the agent keep history across turns, so a follow-up question is understood in context. Creating agent versions and publishing are separate lifecycle operations, and quota is a deployment setting unrelated to conversations.

   Official references

   • Quickstart: Create a prompt agent
5. Question 5Domain: Implement AI solutions by using Microsoft FoundryMedium

   A news app wants each summary to be made only of sentences copied from the original article, with no rewritten wording. Which action should the app pass to begin_analyze_actions?

   • A.ExtractiveSummaryAction, which returns the highest ranked original sentences
   • B.detect_language, which returns the language that the article is written in
   • C.recognize_entities, which lists the people, places and dates that are mentioned
   • D.AbstractiveSummaryAction, which writes new sentences that paraphrase the main ideas
   Show answer and explanation

   Correct answer: A. ExtractiveSummaryAction, which returns the highest ranked original sentences

   Explanation

   Extractive summarization extracts the most salient original sentences from the source, ranked by relevance, which matches the requirement to use only copied sentences. Abstractive summarization generates new sentences that are not verbatim extracts, while recognize_entities and detect_language do not summarize.

   Official references

   • What is summarization?
6. Question 6Domain: Identify AI concepts and capabilitiesMedium

   Why does a transformer receive a positional encoding together with each token vector?

   • A.Because the order in which tokens appear affects how they relate to one another
   • B.To encrypt the training text before it reaches the attention layers of the model
   • C.To keep the vocabulary small by merging tokens that often appear next to each other
   • D.To store earlier chat turns so the model can recall them in later requests from the user
   Show answer and explanation

   Correct answer: A. Because the order in which tokens appear affects how they relate to one another

   Explanation

   The positional encoding tells the transformer where each token sits in the sequence, which matters because token order shapes how tokens relate. It does not shrink the vocabulary, encrypt data or hold conversation history, which the application resends in later prompts.

   Official references

   • Large language models (LLMs) - Training
   • Prompts - Training
7. Question 7Domain: Identify AI concepts and capabilitiesMedium

   A retailer's support chat must answer thousands of shoppers at the same time with very short waits. Which model characteristic matters most?

   • A.A reasoning model set to its highest thinking level for every reply
   • B.A model built only to return embeddings for similarity search
   • C.A model that is able to generate video from a written script
   • D.A chat model optimized for speed and low-latency inference
   Show answer and explanation

   Correct answer: D. A chat model optimized for speed and low-latency inference

   Explanation

   For real-time, high-volume chat, Learn favors models optimized for speed, efficiency and low-latency inference over reasoning-heavy models, which trade speed for accuracy. Video generation and embedding-only models do not produce chat replies.

   Official references

   • Generative AI models - Training
8. Question 8Domain: Identify AI concepts and capabilitiesMedium

   A team wants to see how failures are distributed across a model's test data and to find subsets of data with a higher error rate than the overall benchmark. Which Responsible AI dashboard component helps?

   • A.Counterfactual what-if
   • B.Causal inference
   • C.Model interpretability
   • D.Error analysis
   Show answer and explanation

   Correct answer: D. Error analysis

   Explanation

   Error analysis gives a deep understanding of how failure is distributed and identifies cohorts with a higher error rate than the benchmark, which supports reliability and safety. Interpretability explains predictions, counterfactual what-if shows reactions to feature changes, and causal inference estimates treatment effects.

   Official references

   • What is Responsible AI - Azure Machine Learning
9. Question 9Domain: Identify AI concepts and capabilitiesEasy

   A bank's AI agent interacts only through speech, and customers with a hearing impairment cannot use it. Which change best applies the inclusiveness principle?

   • A.Show a confidence score next to every spoken reply
   • B.Generate text captions of the agent's spoken responses
   • C.Encrypt the recordings of every customer call
   • D.Publish a description of the data used to train the agent
   Show answer and explanation

   Correct answer: B. Generate text captions of the agent's spoken responses

   Explanation

   An AI agent offering speech-based interaction should also generate text captions so it is not unusable for people with a hearing impairment, which is inclusiveness. Confidence scores relate to reliability, encrypting recordings to privacy and security, and describing training data to transparency.

   Official references

   • Responsible AI - Training
10. Question 10Domain: Identify AI concepts and capabilitiesEasy

    A legal firm wants to turn thousands of documents into numeric vectors so it can run semantic search over them. Which type of model fits this task?

    • A.A reasoning model tuned for step-by-step coding assistance in an editor
    • B.A video generation model that renders short clips from text prompts
    • C.An embedding model built for semantic vector representations of text
    • D.A text-to-image model that creates pictures from written descriptions
    Show answer and explanation

    Correct answer: C. An embedding model built for semantic vector representations of text

    Explanation

    Foundry's task guidance points to embedding models for semantic vector representations, which is what semantic search needs. Video generation, image generation and a coding-oriented reasoning model do not produce document vectors.

    Official references

    • Generative AI models - Training