MICROSOFT · AI-103

Developing AI Apps and Agents on Azure practice questions

Developers building AI apps and agents

Pass mark700 / 1000
Practice questions here40

What the AI-103 exam covers

Skills as of 16 Apr 2026
Plan and manage an Azure AI solution25–30%
Implement generative AI and agentic solutions30–35%
Implement computer vision solutions10–15%
Implement text analysis solutions10–15%
Implement information extraction solutions10–15%

Sample AI-103 questions

Five of the 40 practice questions, each with its answer and explanation.

Plan and manage an Azure AI solution

An application should call Foundry models without storing API keys. What is the recommended approach?

  1. Put the key in source code
  2. Use a managed identity with Microsoft Entra ID and assign it the required role
  3. Email the key to each developer
  4. Turn off authentication
Show the answer

B. Use a managed identity with Microsoft Entra ID and assign it the required role

Keyless access through managed identity removes stored secrets and lets access be controlled with role assignments.

Plan and manage an Azure AI solution

At peak times, requests to a model deployment fail with HTTP 429. What is the most likely cause?

  1. The model has been deleted
  2. The deployment is hitting its rate limit or quota
  3. The prompt contains an image
  4. The content filter is switched off
Show the answer

B. The deployment is hitting its rate limit or quota

429 signals throttling. Typical responses are retrying with backoff, raising quota or spreading load across deployments.

Plan and manage an Azure AI solution

You must block harmful content in both prompts and model responses. What should you configure?

  1. A larger context window
  2. Content filters, also called guardrails, on the deployment
  3. A higher temperature
  4. A second search index
Show the answer

B. Content filters, also called guardrails, on the deployment

Content filtering screens inputs and outputs for categories of harmful content.

Implement generative AI and agentic solutions

A chatbot must answer from policy documents that change every week, without retraining a model. Which pattern fits?

  1. Fine-tuning every week
  2. Retrieval-augmented generation
  3. A higher temperature
  4. A longer system prompt containing no documents
Show the answer

B. Retrieval-augmented generation

Retrieval-augmented generation fetches current content at question time and grounds the answer in it.

Implement generative AI and agentic solutions

An agent needs live order status from an internal system. How is this normally implemented?

  1. Hope the model already knows the status
  2. Define a function tool with a schema that the agent can call and the application executes
  3. Paste every order into the system prompt
  4. Retrain the model each night
Show the answer

B. Define a function tool with a schema that the agent can call and the application executes

Function calling lets the model request a specific operation with structured arguments, and the application runs it and returns the result.