Microsoft Azure AI Apps and Agents Developer Associate (AI-103) · 13% practice allocation

Microsoft AI-103: Implement text analysis solutions Practice Questions

5 free sample questions from the Implement text analysis solutions domain, with answers and explanations. The full AI-103 bank has 350 questions across 5 domains. Questions are scenario-based and written from the official Microsoft exam guide.

1.

A legal research tool must summarize case law opinions while preserving exact citation formats (e.g., "123 F.3d 456") without paraphrasing or altering them, even though the surrounding prose should be condensed. Which practice addresses this requirement?

  • A. Allowing the model to paraphrase citations freely since exact formatting rarely matters in legal contexts
  • B. Applying a post-processing step that reformats any citation-like text found in the generated summary, without instructing the summarization prompt to preserve citations verbatim
  • C. Replacing summarization with translation, since translation never alters any text
  • D. Instructing the summarization prompt to preserve citations verbatim while condensing surrounding explanatory text, as part of customizing the task for this domain✓ Correct
Explanation

Customizing the summarization instructions to explicitly preserve citation formatting verbatim while still condensing the surrounding prose directly addresses this domain-specific accuracy requirement. Option A is incorrect and risky — exact citation format is critical in legal research and must not be left to chance paraphrasing. Option C misunderstands the tools; translation converts language and does not perform condensation, so it does not meet the summarization goal. Option B is a right-idea-wrong-stage distractor: cleaning up citation formatting after generation can fix superficial spacing or punctuation, but it cannot recover a citation whose volume, reporter, or page number the model already altered or dropped during generation — preservation has to be instructed at the prompt stage, not patched afterward.

2.

A podcast platform wants to generate accurate, punctuated transcripts of episodes for search indexing, and separately wants a synthesized voice to read out auto-generated episode summaries for a voice assistant integration. Which TWO capabilities are needed? (Select TWO)

  • A. Speech-to-text to produce accurate transcripts for search indexing✓ Correct
  • B. Azure Translator alone, since translation inherently performs transcription
  • C. Text-to-speech to synthesize spoken audio of the episode summaries✓ Correct
  • D. Entity extraction alone, since it can produce audio output directly
Explanation

Speech-to-text produces the punctuated, searchable transcript from episode audio, and text-to-speech synthesizes spoken audio from the generated summaries — together covering both stated needs. Entity extraction (option D) is a text-analysis capability that does not produce audio output. Azure Translator (option B) converts between languages and does not perform speech transcription on its own.

3.

A compliance analytics team wants to scan thousands of chat transcripts for named entities (people, account numbers, organizations) to build an investigation graph, while also flagging any messages containing profanity or harassment language. Which TWO capabilities are needed together? (Select TWO)

  • A. Entity extraction to identify people, account numbers and organizations across the transcripts✓ Correct
  • B. Detection of sentiment, tone, safety issues and sensitive content to flag profanity/harassment✓ Correct
  • C. Text-to-speech to read the transcripts aloud for the investigation team
  • D. Image generation to visually represent each flagged conversation
Explanation

Entity extraction identifies the named entities needed for the investigation graph, while sensitive-content/safety detection flags profanity and harassment language — together covering both stated requirements. Text-to-speech (option C) only converts text to audio and adds no analytical value to this text-based investigation task. Image generation (option D) is an unrelated generative capability that does not extract entities or detect harmful language.

4.

A company operating in a specialized industry (e.g., aviation) finds that general-purpose speech-to-text frequently misrecognizes their domain-specific jargon and acronyms during pilot radio communications. Which capability addresses this accuracy gap?

  • A. A custom speech model trained or adapted to the domain's vocabulary and acoustic conditions✓ Correct
  • B. Text-to-speech tuned with a different voice persona
  • C. Azure Translator configured for a closely related language pair
  • D. Sentiment detection layered on top of the existing transcription output
Explanation

Custom speech models can be adapted to domain-specific vocabulary, acronyms, and acoustic conditions, directly addressing recognition accuracy gaps for specialized jargon that general-purpose models miss. Option B addresses voice output style, not input recognition accuracy. Option C is unrelated, since the problem is domain vocabulary recognition, not language translation. Option D operates on the already-inaccurate transcription and cannot fix upstream recognition errors.

5.

A support team wants to convert incoming customer emails into a structured JSON object with fields for issue category, product name, and urgency level, for automatic routing. Which approach BEST achieves this?

  • A. Speech-to-text transcription of the emails before any further processing
  • B. Sentiment detection alone, since urgency can be inferred purely from a positive/negative score
  • C. Azure Translator applied to every email regardless of source language
  • D. Generative prompting with a defined output schema (or a Foundry Tools extraction capability) instructing the model to return the fields as JSON✓ Correct
Explanation

Structured extraction via generative prompting with an explicit schema (or an equivalent Foundry Tools extraction capability) is designed exactly for turning unstructured text into defined structured fields like category, product, and urgency. Option B conflates sentiment (positive/negative tone) with urgency, which is a different and broader signal than sentiment alone can reliably capture. Option A is unnecessary since the input is already text, not audio. Option C applies translation to every email even when unneeded, and translation alone does not produce structured routing fields.

Test yourself across every AI-103 domain

The free AI-103 practice test mixes about 20 questions weighted like the exam. The full 350-question bank adds instant grading, flashcards and a timed mock exam.

Take the free AI-103 practice testPractice the full bank

Study this domain

Implement text analysis solutions study guide (AI-103) →

Other AI-103 domains

← All Microsoft Azure AI Apps and Agents Developer Associate study resources