Developing AI Apps and Agents on Azure โ€” Question Bank

MS_AI_103
AssociateVersion AI-103Official exam guide โ†—

Ready to test yourself?

A timed, blueprint-proportional exam drawn fresh from this bank โ€” with a per-domain score report.

๐Ÿ”’ Unlock the simulation โ†’

๐Ÿ”“ Free preview: showing 10 of 60 questions. Unlock the full bank โ€” every question, explanation, and reference.

Unlock all 60 questions โ†’

10 questions across 5 topics. Choose an answer for each question, then check it to see the correct answer and explanation.

Filter by topic

All 10 questions

PLAN_MANAGE

Plan and manage an Azure AI solution

10 questions in topic

Choosing appropriate models and Foundry services for generative, grounding, vector search, agent and multimodal tasks; designing infrastructure, deployment options and CI/CD integration; managing quota, scaling, rate limits and cost; monitoring performance, drift, safety and grounding quality; securing with managed identity and private networking; and implementing responsible AI through filters, evaluators, auditing and agent governance.

1
Single choice~90s

A deployment is created successfully but every inference request returns 401. The application's managed identity holds Contributor on the Foundry resource. What is the cause?

2
Single choice~90s

A team wants to reference a model in application code by a stable name that does not change when the underlying model version is upgraded. What should the code call?

3
Multiple selectSelect 2 answers~130s

An application must call a Foundry deployment with no stored credential and no public network exposure. Which TWO configurations are required? (Choose two.)

4
Single choice~90s

A nightly job processes 2 million documents through a model. Results are needed by morning and cost must be minimised. Which deployment type fits?

5
Single choice~90s

Which observation indicates that a deployment's quota is genuinely insufficient rather than that its traffic is bursty?

6
Single choice~90s

A Foundry project's flows must move through development, test, and production with the same definition and environment-specific targets. What should be parameterised?

7
Single choice~115s

A model version pinned to a production deployment reaches its retirement date and the deployment stops serving requests. Which upgrade policy would have prevented the outage without introducing an unvalidated version earlier?

8
Single choice~90s

Which control prevents an agent from taking an irreversible action without human authorisation?

9
Single choice~90s

A RAG assistant is blocked by the default content filter when summarising legitimate clinical documents. What is the correct remedy?

10
Multiple selectSelect 2 answers~105s

Which TWO signals belong to monitoring the quality of an AI system rather than its infrastructure? (Choose two.)