AI-103 Sample Questions With Answers

About these questions

These five questions are original. We wrote them to match the style of the AI-103 skills measured, one per domain. They are not real exam questions, and the real exam will word things differently, mix in other formats, and may include interactive components. Our how we write our questions page explains our approach.

How to use these

Try each question without looking at your notes, and pick an answer before you open the explanation. Then read the explanation even when you were right, because the reason the closest wrong option fails is where the learning is. Note any domain where you slipped and revisit it on the domains explained page.

Question 1 (Plan and manage an Azure AI solution)

A company runs a customer-support chatbot built on a Foundry model deployment. The operations team is building a dashboard from traces and telemetry exported from the application. Which two signals belong on that monitoring dashboard rather than coming from the content filtering system? Choose two.

A. The latency of each tool call compared with the model call in the same request
B. The severity level assigned to a user’s prompt in the violence category
C. Whether a response was stopped because it matched a custom blocklist
D. Token usage per request over time
E. The harm category reported in the content filter results of a response

Show the answer and explanation

Correct answers: A and D. Latency breakdowns and token analytics come from traces and telemetry, and they answer operational questions: where time goes and how much each request costs. Options B, C, and E are all outputs of content filtering, which decides at request time whether a prompt or completion may pass. The closest wrong option is C, because a blocklist stop is visible in the application and could be logged, but it is still a filtering outcome and not an operational signal. Mixing the two leads to dashboards that cannot answer either question well.

Question 2 (Implement generative AI and agentic solutions)

You build an agent with two tools: one searches the company policy library, and one creates support tickets. For requests that your application has already classified as policy questions, the agent must always call the search tool and never skip it. Which request setting should you use?

A. Set parallel tool calls to false
B. Set the tool choice to the named search tool
C. Provide the search index name in the tool resources
D. Request a JSON response format

Show the answer and explanation

Correct answer: B. Setting the tool choice to a specific named tool forces the model to call that tool. By default the model decides for itself whether to call a tool and which one. The closest wrong option is C, because tool resources hold configuration the tool needs, such as which index to query, but they do not make the model choose the tool. Parallel tool calls only control whether several calls can happen at once, and a response format shapes the output, not tool selection.

Question 3 (Implement computer vision solutions)

A marketing team wants to replace only the background of a product photo using an image editing request with a mask. How should the mask mark the background so it is regenerated?

A. Make the background pixels fully opaque in a mask image of any size
B. List the background region as coordinates in the prompt text and skip the mask
C. Make the background pixels fully transparent in a PNG mask that has the same pixel dimensions as the photo
D. Paint the background black in a JPEG mask

Show the answer and explanation

Correct answer: C. In inpainting, the instruction is carried by the alpha channel of the mask, which is why it must be a PNG. The mask is applied pixel for pixel, so its dimensions must match the photo, and fully transparent pixels mark the area the model should regenerate. Option A is the closest wrong answer because it inverts the meaning: opaque pixels are preserved. A JPEG has no alpha channel, and the prompt alone cannot define an exact region.

Question 4 (Implement text analysis solutions)

A restaurant group collects thousands of guest reviews. Managers want to know that guests praise the staff but complain about waiting time, within the same review. Which capability fits best?

A. Key phrase extraction
B. Opinion mining
C. Language detection
D. Named entity recognition

Show the answer and explanation

Correct answer: B. Opinion mining is aspect-based sentiment analysis. It links an opinion to the specific thing it is about, so one review can be positive about staff and negative about waiting time. Key phrase extraction, the closest wrong option, would list phrases such as “staff” and “waiting time” but would not say how the guest felt about each. Named entity recognition finds people, places, and organizations, and language detection only identifies which language the review is written in.

Question 5 (Implement information extraction solutions)

An insurer receives scanned claim packets that mix repair estimates, medical invoices, and signed consent forms. Each document type has its own extraction model. What should the pipeline run first so each page reaches the right model?

A. A layout analysis step
B. A custom extraction model for every document type, run on every page
C. A text translation step
D. A document classification step

Show the answer and explanation

Correct answer: D. Classification reports which type a document is, and the pipeline then calls the extraction model built for that type. Layout analysis, the closest wrong option, returns structure such as tables and text regions but does not decide what kind of document it is. Running every extraction model on every page also works in principle, but it wastes compute and produces confusing results on documents that do not match. Translation has nothing to do with routing.

More guides for AI-103

Next step

If you missed a question, the domains explained page shows where it fits. For more questions with detailed feedback, see the AI-103 practice tests.

Try the AI-103 practice exams