Azure AI Fundamentals (AI-901) · 57% practice allocation

Azure AI-901: Implement AI solutions by using Microsoft Foundry Practice Questions

5 free sample questions from the Implement AI solutions by using Microsoft Foundry domain, with answers and explanations. The full AI-901 bank has 105 questions across 2 domains. Questions are scenario-based and written from the official Azure exam guide.

1.

Your Foundry SDK chat client needs to show the assistant's reply appearing word-by-word as it's generated, instead of waiting for the whole response. What should you use?

  • A. Multiple separate short prompts stitched together after each finishes
  • B. A batch inference endpoint intended for offline scoring
  • C. A synchronous call that blocks until the entire response is generated, then reprints it letter by letter client-side
  • D. A streaming response call that yields partial output chunks as they are produced✓ Correct
Explanation

Streaming is the SDK pattern for incremental output: the client receives and displays chunks as the model generates them, giving a live typing effect. Faking it by re-printing a completed response letter by letter is not real streaming, chaining separate prompts changes the task, and batch endpoints are built for offline, non-interactive scoring rather than live chat.

2.

You are building a lightweight application that lets a user upload a scanned form and immediately see the extracted fields displayed in a table in the UI. What is the correct general flow?

  • A. Send the document to a text-to-speech endpoint and play the resulting audio
  • B. Convert the form to a generative image and display that as if it were the extracted data
  • C. Upload the document, call Content Understanding via the SDK/API with the appropriate analyzer, then parse the structured result and render it in the UI✓ Correct
  • D. Call Content Understanding to extract the fields, then display the raw unparsed JSON response directly in the UI instead of mapping it into the table's columns
Explanation

The correct flow uploads the document, submits it to Content Understanding through the SDK/API with a relevant analyzer, and then parses the returned structured result into the app's own table columns. Calling the service correctly but then dumping the raw JSON straight into the UI skips the necessary parsing/mapping step, so the user sees an unreadable blob instead of a table — the right service call, at the wrong final step. Sending a document to text-to-speech produces audio rather than structured data, and image generation is unrelated to reading an existing form.

3.

An application built with Azure Speech needs to convert a live customer call into text as the conversation happens, not after it ends. Which capability should it use?

  • A. Text-to-speech synthesis
  • B. Batch transcription designed for large offline archives
  • C. Real-time / streaming speech-to-text transcription✓ Correct
  • D. Image captioning
Explanation

Real-time (streaming) speech-to-text is designed to transcribe audio as it is spoken, which fits a live call scenario. Batch transcription is intended for processing recorded audio files after the fact, text-to-speech goes in the opposite direction (text to audio), and image captioning is unrelated to audio at all.

4.

Your information-extraction application calls Content Understanding and gets back a JSON result, but a few fields have low confidence scores. What is the most appropriate way to handle this in the application?

  • A. Flag low-confidence fields for human review rather than silently trusting them as fully accurate✓ Correct
  • B. Always discard the entire document if even one field has a low confidence score
  • C. Ignore confidence scores entirely and always treat every extracted value as certain
  • D. Automatically publish all extracted values to the customer immediately with no review, regardless of confidence
Explanation

Confidence scores exist so applications can route uncertain extractions to a human reviewer rather than blindly trusting them, which is standard practice for extraction pipelines. Discarding whole documents over one low-confidence field is wasteful, ignoring confidence scores defeats their purpose, and auto-publishing uncertain data to customers risks propagating errors.

5.

In a client application built against a Foundry agent, a run occasionally returns a status indicating the agent is waiting on a tool call to complete. What should the client application do?

  • A. Poll or await the run status until it completes, then retrieve the resulting messages once finished✓ Correct
  • B. Treat the in-progress status the same as a completed run and immediately display whatever message text is already available
  • C. Restart the entire conversation thread from scratch
  • D. Cancel the current run and immediately start a new run with the same user message, since the previous one appears stuck
Explanation

Agent runs are asynchronous and pass through in-progress states, including tool execution; a well-built client polls or awaits until the run reaches a terminal state before reading the final messages. Treating the in-progress status as if it were already 'completed' risks displaying an empty or partial message before the tool result is incorporated; restarting the thread throws away conversation history unnecessarily; and canceling and resubmitting a run that is simply mid-tool-call is a premature reaction to normal, expected latency rather than an actual failure.

Test yourself across every AI-901 domain

The free AI-901 practice test mixes about 20 questions weighted like the exam. The full 105-question bank adds instant grading, flashcards and a timed mock exam.

Take the free AI-901 practice testPractice the full bank

Study this domain

Implement AI solutions by using Microsoft Foundry study guide (AI-901) →

Other AI-901 domains

← All Azure AI Fundamentals study resources