AI-103 Practice Questions: Plan and manage an Azure AI solution Domain

Correct answer (B): This scenario has two different needs: grounded answers over a knowledge base and deterministic interaction with a live business system. RAG is the right fit for the knowledge content, while tool calling or orchestrated workflows are more reliable for checking live status and creating tickets.

Why the other options are wrong:
- Option A: Prompt-only behavior is not reliable for deterministic external actions or querying live systems.
- Option C: Historical tickets cannot guarantee current live status, and fine-tuning is not the best method for deterministic system actions.
- Option D: Content safety filters do not provide the workflow reliability, authorization, or action design needed for live ITSM operations.

Correct answer (C): Submitting refunds is a high-risk action, so it should not be left to a fully autonomous agent without guardrails. The best design is to combine least-privilege tool access with a human approval step before the final action is executed.

Why the other options are wrong:
- Option A: Content filters can reduce harmful outputs, but they do not provide the operational control needed for financial approval actions.
- Option B: Grounding helps explain a recommendation, but citations alone do not replace approval controls for high-risk transactions.
- Option D: Fine-tuning historical behavior does not remove the need for governance and approval controls over risky business actions.

Correct answer (B): Separate environments are the best planning choice because they improve version control, rollback, policy isolation, and identity boundaries. They reduce the risk that experimental changes directly affect production workloads.

Why the other options are wrong:
- Option A: Better documentation helps, but it does not provide the isolation needed for safe rollout and rollback.
- Option C: Model size does not address environment isolation or deployment discipline.
- Option D: Content filters do not prevent configuration drift or protect production from untested changes.

Correct answer (B): High-risk financial actions should not be fully autonomous. Human approval provides an appropriate control point, and limiting tool access supports least privilege. Auditable tool calls also help the organization track who initiated actions and how the agent executed them.

Why the other options are wrong:
- Option A: Content filters do not make sensitive business actions safe to automate without review, and they do not replace approval controls.
- Option C: Fine-tuning does not solve the governance problem, and broad permissions conflict with least-privilege design.
- Option D: A larger model does not remove the need for approval and audit controls on sensitive financial operations.

Correct answer (B): Sensitive enterprise actions should be controlled through explicit tool definitions, scoped permissions, and approval steps. Prompt instructions alone are not an enforceable control boundary, and content safety filters do not replace access control or approval workflows.

Why the other options are wrong:
- Option A: Prompts can guide behavior, but they are not sufficient for permission enforcement or approval requirements on high-impact actions.
- Option C: Content safety filtering helps with unsafe content, not with enforcing business approvals or limiting ERP permissions.
- Option D: A larger model may change behavior, but it does not provide enforceable authorization boundaries or approval controls.

Correct answer (B): When a RAG system returns outdated or incomplete answers after source changes, retrieval quality is the best place to investigate first. Index freshness, chunking strategy, and citation validation directly affect whether the latest documents are retrieved and grounded correctly.

Why the other options are wrong:
- Option A: A larger model is not the best first fix when the problem points to stale or poor retrieval. Grounding quality depends heavily on the retrieval pipeline.
- Option C: Fine-tuning is not the primary remedy for frequently changing private knowledge. Retrieval freshness and source quality are more relevant.
- Option D: Content filters do not fix stale indexes, poor chunks, or missing citations.

Correct answer (B): RAG is the best fit for frequently changing enterprise knowledge. By indexing the current policy documents and grounding responses on retrieved content, the assistant can use up-to-date information without retraining whenever the manuals change.

Why the other options are wrong:
- Option A: Fine-tuning is not the best choice for knowledge that changes weekly. It adds retraining overhead and does not naturally keep answers aligned to the newest source documents.
- Option C: A larger model does not solve the core problem of keeping answers current and grounded. Long prompts are also not a good substitute for retrieval over changing documents.
- Option D: Content safety filters help with unsafe content, but they do not provide grounding on current policies or solve stale knowledge.

Correct answer (C): A model deployment only exposes a model endpoint. The application still needs its own deployment and operational planning, including environment-specific configuration, identity, networking, monitoring, and release controls. Separate dev, test, and production processes reduce accidental exposure and support controlled releases.

Why the other options are wrong:
- Option A: A model endpoint is not the entire application deployment. The app still needs its own secure and observable production rollout.
- Option B: Sharing the same credentials and configuration across environments increases risk and weakens environment separation.
- Option D: Fine-tuning per environment does not address the actual gap, which is application deployment, environment separation, and operational readiness.

Correct answer (B): The symptoms point to retrieval design problems, not just model quality. Outdated citations suggest stale indexing, and incomplete coverage of long procedures suggests chunking or embedding issues. Fixing the retrieval pipeline and refreshing the index is the most direct and appropriate response.

Why the other options are wrong:
- Option A: A larger model does not reliably fix stale indexes, weak chunking, or unsuitable embeddings. Those are retrieval-layer problems.
- Option C: Content filters address harmful or prohibited content, not relevance, groundedness, or index freshness.
- Option D: Fine-tuning on older documents would move the system further away from the requirement to use current policy content.

Correct answer (C): RAG is the best fit when answers must come from changing enterprise content. By retrieving current policy documents at runtime and grounding the response on those sources, the assistant can answer from the latest information instead of relying on static model knowledge.

Why the other options are wrong:
- Option A: A larger model does not solve the core problem of needing current private knowledge. Pretraining alone is not a reliable source for frequently changing enterprise policies.
- Option B: Fine-tuning can shape behavior or style, but it is not the primary solution for rapidly changing private documents. Retrieval quality and freshness are more relevant here.
- Option D: Content filters can help with harmful content, but they do not provide factual grounding or ensure that answers reflect the latest documents.