ExamSprintLab

AI-103 text analysis solutions explained

Updated September 20, 2026

Implement text analysis solutions is worth 10–15% of AI-103 and splits cleanly in two: language model text analysis, and speech. The shift from earlier Azure AI exams matters here. You are no longer expected to reach for a dedicated endpoint for every task — a fair number of these jobs are now done with generative prompting and Foundry Tools, and part of what the exam tests is knowing when a purpose-built service is still the better answer.

Language model text analysis

The objectives cover extracting entities, topics, summaries and structured JSON output through generative prompting and Foundry Tools. Structured output is the detail worth underlining: an agent cannot reason reliably over prose, so a step that turns free text into a fixed schema is a common building block.

Alongside extraction sit:

  • Detection of sentiment, tone, safety issues and sensitive content. Note that sensitive content detection — personal data appearing where it should not — is a separate concern from sentiment, and the exam treats them as different requirements.
  • Translation, either with Azure Translator in Foundry Tools or through a model-powered flow. Translator is the answer when you need broad language coverage, predictable cost and consistency at volume; a model flow is the answer when translation has to be context-aware or combined with another task in one step.
  • Customising output for domain tasks, such as compliance summarisation or extracting terms specific to your industry.

A reliable exam heuristic: if the requirement is a high-volume, well-defined, repeatable operation, the purpose-built tool wins on cost and consistency. If the requirement is nuanced, combined or domain-specific, the model wins.

Speech

Speech is tested as an agent modality rather than as a standalone product:

  • Converting speech to text and text to speech for agentic interactions.
  • Integrating speech into an agent, including custom speech models — which exist to handle vocabulary a general model mishears, such as product names, drug names or part numbers.
  • Multimodal reasoning from audio input, where the model works from the audio rather than from a transcript.
  • Translating speech into other languages with models and Foundry Tools.

Know the difference between speech recognition (what was said), speaker recognition (who said it) and custom speech (recognising your vocabulary correctly). Questions in this area are frequently decided on exactly that distinction.

Sample questions

Question 1. A logistics company transcribes support calls. The transcripts are accurate except that its own product codes, such as 'XR-14 Duotrack', come out wrong almost every time. What should you implement?

  • A. Speaker recognition on each call
  • B. A custom speech model trained on the company’s product vocabulary
  • C. Speech translation into a second language and back
  • D. A higher audio sample rate on the recording device
Show answer

Answer: B

A custom speech model is trained on domain vocabulary so that product names and codes are recognised correctly. Speaker recognition identifies who is talking, translation changes the language, and raising audio quality does not teach the model unfamiliar words.

Question 2. Incoming emails must be turned into records with fields for customer, product, issue type and urgency, so that a downstream agent can route them. What is the most suitable design?

  • A. Run sentiment analysis and route by the sentiment score
  • B. Generate a free-text summary of each email
  • C. Use generative prompting to produce structured JSON matching a defined schema
  • D. Index the emails for keyword search and let the agent search them
Show answer

Answer: C

Prompting the model to return a fixed JSON schema gives the downstream agent predictable fields to act on. Sentiment alone yields one score, a summary is still prose, and full-text keyword search does not produce structured fields.

Question 3. A retailer must translate two million short product descriptions into eight languages, with consistent wording and predictable cost. Quality requirements are standard. Which approach fits best?

  • A. Azure Translator in Foundry Tools as a batch translation pipeline
  • B. A large language model prompted once per description per language
  • C. Human translators for all eight languages
  • D. Client-side browser translation for each visitor
Show answer

Answer: A

A dedicated translation service is built for high-volume, repeatable translation and gives consistent results at a predictable price. A model-powered flow is better when translation needs context or must be combined with another task, but at this volume it costs more and varies more. Manual translation and per-user machine translation in the browser do not meet the requirement.

What to practise

Take one paragraph of messy real text and run it four ways: entity extraction, sentiment, a structured JSON extraction against a schema you define, and a translation. Then record ten seconds of audio containing a word no general model would know, transcribe it, and see it fail. Those two exercises give you the intuition behind nearly every question in this domain.