Pass Azure AI Fundamentals (AI-901) — the First Time

1. A bank deploys a generative AI assistant that summarizes loan applications and recommends approve or decline. An audit shows that applicants from one neighborhood are recommended for decline far more often than applicants with identical income and credit history from another neighborhood. Which responsible AI principle is most directly at stake?

   • A. Accountability
   • B. Inclusiveness
   • C. Fairness
   • D. Transparency

   Answer: C. Fairness requires that an AI system treat people equitably and not produce different outcomes for similar people based on sensitive attributes or proxies for them. Neighborhood frequently acts as a proxy for race or income, so the outcome gap is a fairness problem. Transparency concerns explaining how the system works, inclusiveness concerns serving people of all abilities and backgrounds, and accountability concerns human ownership and governance, none of which describes the outcome disparity itself.

   Exam domain: Identify AI concepts and capabilities

2. A retailer launches a customer support chat experience powered by a large language model. Customers assume they are messaging a human agent, and nothing in the interface indicates otherwise. Which responsible AI principle is the retailer failing to uphold?

   • A. Reliability and safety
   • B. Transparency
   • C. Fairness
   • D. Privacy and security

   Answer: B. Transparency means people understand how and when AI is being used, which starts with disclosing that they are talking to an AI system rather than a person. Reliability and safety would concern whether the bot answers correctly and fails gracefully, privacy and security would concern what happens to the chat data, and fairness would concern whether some customers get worse service than others.

   Exam domain: Identify AI concepts and capabilities

3. A university's lecture capture service uses speech to text. Students who are deaf report that audience questions are never captioned, and students with strong regional accents report frequent transcription errors. Which principle should guide the fix?

   • A. Inclusiveness
   • B. Accountability
   • C. Transparency
   • D. Privacy and security

   Answer: A. Inclusiveness means AI systems should empower and engage everyone, including people with disabilities and people whose speech, language, or background differs from the majority in the training data. Accountability addresses who owns the outcome, transparency addresses explaining the system, and privacy and security addresses protection of the recordings, none of which is the barrier described here.

   Exam domain: Identify AI concepts and capabilities

4. A developer estimating the cost of a chat application reads that the model is billed per token. Which statement best describes what a token is in this context?

   • A. An authentication credential used when calling the endpoint
   • B. One single character of text sent to the model
   • C. A common chunk of text, a word or piece of a word
   • D. One complete sentence of text sent to the model at a time

   Answer: C. Language models split text into subword tokens, so a short common word is usually one token while a long or unusual word becomes several. A token is not a character and not a sentence. The word token is also used for access credentials in authentication, which is a different meaning entirely and a common source of confusion.

   Exam domain: Identify AI concepts and capabilities

5. A team wants their internal search to return documents that are conceptually related to a query even when the documents do not contain the query's exact words. Which representation of the text makes that possible?

   • A. A keyword inverted index built from the exact terms that appear in each document
   • B. A confusion matrix summarizing the classification results
   • C. A cryptographic hash computed from each of the documents
   • D. An embedding, a vector placing similar meanings close together

   Answer: D. Embeddings encode meaning as coordinates, so semantically similar text ends up near other similar text and can be found by vector similarity rather than exact word overlap. An inverted index matches literal terms and is exactly the limitation being solved, a hash is designed so similar inputs produce completely different outputs, and a confusion matrix is an evaluation artifact, not a text representation.

   Exam domain: Identify AI concepts and capabilities

6. A compliance team needs a model to produce policy summaries with wording that is as consistent as possible from run to run for the same input. Which configuration change best supports this?

   • A. Set temperature to 1.5
   • B. Set temperature to a value at or near 0
   • C. Increase the maximum output tokens
   • D. Increase the frequency penalty on each request

   Answer: B. Temperature controls how much randomness is introduced when picking the next token, so a value at or near 0 makes the model consistently choose the highest probability token and produces the most repeatable output. A high temperature does the opposite, max tokens only changes how long the response can be, and a frequency penalty only discourages the model from repeating tokens within a response, which does nothing to make output repeatable from run to run.

   Exam domain: Identify AI concepts and capabilities

7. A support team wants a solution that reads an incoming customer email and drafts a suggested reply for a human agent to review before sending. The solution never looks anything up and never updates another system. Which type of AI workload best fits this scenario?

   • A. An anomaly detection workload
   • B. A single call to a generative language model
   • C. An agentic AI solution that plans and calls tools
   • D. A speech synthesis workload

   Answer: B. The task is text in, text out, with all the needed context already supplied in the prompt, so one generative model call is sufficient. An agentic solution is only warranted when the system must call tools, take actions, or work through multiple dependent steps. Speech synthesis produces audio and anomaly detection flags unusual data points, neither of which is being asked for.

   Exam domain: Identify AI concepts and capabilities

8. A retailer wants a solution that handles a returns request end to end: it looks up the order in the order management system, decides whether the return falls inside policy, generates a shipping label, and updates the CRM record. Which workload type does this describe?

   • A. A single prompt to a generative model with a long system message
   • B. A text classification model applied to the request
   • C. A speech recognition workload
   • D. An agentic AI solution that can call tools and perform multiple steps

   Answer: D. The solution has to reach into external systems, act on the results, and take real actions, which is the defining shape of an agentic workload. A longer system message does not give a model the ability to query an order system or write to a CRM. Classification would only label the request, and speech recognition is irrelevant because the input is text.

   Exam domain: Identify AI concepts and capabilities

9. A developer is building a chat app on a model deployed in Microsoft Foundry. The assistant must always behave as a formal IT helpdesk agent, no matter what the user types. Which message role should carry those standing instructions?

   • A. The tool role
   • B. The system role
   • C. The assistant role
   • D. The user role

   Answer: B. The system message, sometimes called the developer message, carries persistent behavior, persona, and rules that apply to the whole conversation. The user role carries each individual request, the assistant role holds the model's own previous replies, and the tool role returns output from a tool call back to the model.

   Exam domain: Implement AI solutions by using Microsoft Foundry

10. A prompt sent to a model deployed in Foundry looks like this:Classify the ticket sentiment as Positive, Neutral, or Negative.Ticket: "Shipping was fast and the item works great." -> Positive Ticket: "It arrived, but the box was dented." -> Neutral Ticket: "Third failure this month, I want a refund." -> Negative Ticket: "The app crashes every time I open it." ->Which prompt engineering technique is being used?

    • A. Retrieval augmented generation
    • B. Zero-shot prompting
    • C. Few-shot prompting
    • D. Chain-of-thought prompting

    Answer: C. Supplying several worked input and output pairs before the real input is few-shot prompting, which teaches the model the label set and the exact output shape by example. Zero-shot gives instructions with no examples, chain-of-thought asks the model to reason step by step, and retrieval augmented generation injects retrieved source documents into the prompt.

    Exam domain: Implement AI solutions by using Microsoft Foundry

Want more? Finish the quiz above to unlock 10 more free questions, or open the full AI-901 practice app.