Free AI-103 Generative AI and Agentic Solutions Practice Questions

The Implement generative AI and agentic solutions domain makes up 30–35% of the Microsoft AI-103 exam. Practice free questions covering Microsoft Foundry model deployment, prompt engineering, RAG patterns, AI agent orchestration, and fine-tuning — each with the correct answer and a detailed explanation.

AI-103 Generative AI and Agentic Solutions Sample Questions with Answers

Sample Question 1 — Implement generative AI and agentic solutions

You are building a Microsoft Foundry employee policy copilot. HR policies change weekly, and the app must answer only from approved policy documents and include citations. Users currently receive confident answers that are not supported by the source documents. What is the best design change?

  1. A. Move to a larger model deployment and keep all policy knowledge only in the system prompt.
  2. B. Fine-tune the model on the policy documents after each policy update.
  3. C. Use retrieval-augmented generation with Azure AI Search over the approved documents and return citations. (Correct answer)
  4. D. Enable stricter content filtering so unsupported answers are automatically blocked.

Correct answer: C

Explanation: Correct answer (C): RAG with Azure AI Search is the best fit because the requirement is current, enterprise-specific, source-backed answers from approved documents. RAG grounds the model on retrieved content, helps reduce hallucinations, and supports citations for traceability and trust. This is a better Azure-native pattern than relying on a larger model or repeatedly fine-tuning for frequently changing policies. Why the other options are wrong: - Option A: A larger model can still hallucinate and does not by itself provide grounded, source-backed answers from current enterprise documents. - Option B: Fine-tuning is not the default solution for frequently changing business knowledge. Searchable external knowledge sources are usually better for dynamic content. - Option D: Content filtering can reduce unsafe outputs, but it does not guarantee factual grounding, current knowledge, or answers restricted to approved documents.

Sample Question 2 — Implement generative AI and agentic solutions

A product-manual assistant uses Azure AI Search. Users search for exact part numbers such as PX-4400, but they also ask conceptual questions phrased differently from the source text. You need retrieval that handles both exact lexical matches and embedding-based similarity. Which search approach is the best fit?

  1. A. Vector search only
  2. B. Keyword search only
  3. C. Hybrid search (Correct answer)
  4. D. Semantic search only

Correct answer: C

Explanation: Correct answer (C): Hybrid search is the best answer because the scenario requires both exact lexical matching for identifiers like part numbers and embedding-based similarity for conceptually similar phrasing. The blueprint specifically identifies hybrid retrieval as the common choice when both lexical and vector-style matching are needed. Why the other options are wrong: - Option A: Vector search is useful for semantic similarity, but by itself it is not the best fit when exact lexical matches are also critical. - Option B: Keyword search can help with exact identifiers, but it may miss semantically similar content when wording differs from the source documents. - Option D: Semantic search can improve meaning-aware retrieval or ranking, but the scenario explicitly calls for both lexical and embedding-based matching, which is why hybrid is the best choice here.

Sample Question 3 — Implement generative AI and agentic solutions

A Microsoft Foundry RAG assistant answers questions from policy PDFs stored in Azure AI Search. The team recently ingested each PDF as one very large chunk, and some updated files were not successfully reindexed. Users now report unsupported answers, missing citations, and stale content. A project manager suggests switching to a larger model deployment first. What should the engineering team inspect before changing models?

  1. A. The retrieval pipeline, including chunking, index freshness, citations, and grounding instructions (Correct answer)
  2. B. Only the content filter settings, because stale answers are usually caused by safety configuration
  3. C. Only the web app CPU utilization, because unsupported answers usually come from slow page rendering
  4. D. Only the model size, because retrieval problems are usually secondary in RAG applications

Correct answer: A

Explanation: Correct answer (A): The best first step is to inspect retrieval quality. The blueprint states that poor RAG quality is often caused by chunking strategy, embedding alignment, missing filters, or stale indexes rather than model choice. Missing citations and stale answers strongly point to retrieval health issues. In RAG scenarios, engineers should inspect retrieval quality, index freshness, citations, and grounding instructions before assuming a larger model is needed. Why the other options are wrong: - Option B: Content filters address unsafe outputs, not the core grounding, freshness, or citation problems described in the scenario. - Option C: Web app CPU utilization may affect UI performance, but it does not explain stale retrieved content or unsupported answers with missing citations. - Option D: Model size is not the best first area to inspect when the scenario already indicates chunking and reindexing problems in a RAG system.

Sample Question 4 — Implement generative AI and agentic solutions

An operations agent must open incidents in an external ticketing API. Each request must include fields such as severity, assetId, and summary. The current prompt asks the model to 'write the API request as text,' and the app frequently sends malformed requests. What is the best improvement?

  1. A. Ask the model for a more detailed paragraph before the app submits the request
  2. B. Add Azure AI Search so the model can read ticketing manuals before posting incidents
  3. C. Use structured tool or function calling with an explicit schema, and validate parameters before execution (Correct answer)
  4. D. Fine-tune the model on historical tickets so it memorizes the endpoint format

Correct answer: C

Explanation: Correct answer (C): Structured tool or function calling is the best choice because the requirement is reliable action execution against a business system. The blueprint states that tool calling is more reliable than expecting free-text output for external actions, and that tool schemas should be explicit while the application enforces authorization and parameter validation before executing sensitive operations. Why the other options are wrong: - Option A: A more detailed paragraph is still free text and does not solve the reliability problem of generating machine-executable requests. - Option B: Retrieval can help the model answer questions about the ticketing process, but it is not the best solution for producing reliable, structured API calls. - Option D: Fine-tuning is not the default answer for action reliability and would not replace the need for structured tool invocation and parameter validation.

Sample Question 5 — Implement generative AI and agentic solutions

A procurement agent can submit purchase requests to an internal system. Finance policy requires a person to review any order before it is placed, and the agent should not have broader purchasing permissions than needed. What should you implement?

  1. A. Allow autonomous submissions, but rely on content filtering to block risky requests
  2. B. Require human approval before submission and limit the agent's tool permissions to only the necessary purchasing action (Correct answer)
  3. C. Give the agent full purchasing API access so it can correct mistakes without waiting for a reviewer
  4. D. Switch to a larger model so it can understand finance policy well enough to act independently

Correct answer: B

Explanation: Correct answer (B): This is the best design because the scenario combines two requirements: human review for a high-risk action and least-privilege access for the agent's tool. The blueprint explicitly states that high-risk actions should use human-in-the-loop review or explicit confirmation, and that least-privilege access applies to agent tools as well as Azure resources. Why the other options are wrong: - Option A: Content filtering helps reduce unsafe outputs, but it does not replace authorization controls or a required human approval step for sensitive transactions. - Option C: Full purchasing API access violates least-privilege design and increases the blast radius of mistakes or misuse. - Option D: A larger model does not remove the need for workflow controls, human approval, or scoped permissions for sensitive actions.

Sample Question 6 — Implement generative AI and agentic solutions

Your team changed an agent's prompt template and orchestration logic in a test environment, but the underlying model deployment remains unchanged. Before release, the manager asks whether this change still needs versioning and a rollback plan. What is the best answer?

  1. A. No, because only model deployment changes can affect application behavior in production
  2. B. No, because prompt and workflow changes are text changes and do not require release controls
  3. C. Yes, because application workflow, prompts, indexes, and model deployments are separate change areas, and app changes should still be versioned with rollback planning (Correct answer)
  4. D. Yes, but only if every application change is paired with a new model deployment

Correct answer: C

Explanation: Correct answer (C): This is correct because model deployment and application deployment are separate concerns. The blueprint also states that prompt, workflow, index, and model changes can all affect behavior, so environment separation, versioning, and rollback planning are part of production readiness even when the underlying model deployment does not change. Why the other options are wrong: - Option A: Application behavior can change significantly through prompt, workflow, and index updates even if the model deployment stays the same. - Option B: Prompt and orchestration changes can alter quality, safety, latency, and business behavior, so they still need release controls. - Option D: Application changes do not require a new model deployment every time. These are separate change areas.

Sample Question 7 — Implement generative AI and agentic solutions

In production, an agent sometimes responds slowly, occasionally returns answers with no citations, and intermittently fails after attempting a tool call. You need to determine whether the primary problem is model latency, retrieval behavior, or tool execution. What should you add or review?

  1. A. Traces and logs that capture model calls, retrieval steps, tool invocations, failures, and latency breakdowns (Correct answer)
  2. B. Only the final chat transcript shown to end users, because intermediate steps are unnecessary
  3. C. Only Azure resource tags, because tags identify the responsible service owners
  4. D. Only content filter event counts, because all major failures appear there

Correct answer: A

Explanation: Correct answer (A): The best answer is end-to-end observability with traces and logs across model calls, retrieval, tool usage, failures, and latency. That is the only option that lets you distinguish whether the issue is in generation, search, or tool execution. The blueprint specifically calls for observability that makes those intermediate steps diagnosable in production. Why the other options are wrong: - Option B: Final chat transcripts are useful but insufficient. They do not show where latency or failures occurred inside retrieval and tool execution paths. - Option C: Resource tags help governance, not root-cause analysis of model, retrieval, or tool failures. - Option D: Content filter counts cover a narrow class of events and do not diagnose missing citations, retrieval problems, or tool-call failures.

Sample Question 8 — Implement generative AI and agentic solutions

A customer service assistant answers questions from warranty guides stored in a search index. Customers also ask, 'Has my replacement order shipped yet?' Order status changes throughout the day in an authoritative transaction system. What is the best design?

  1. A. Use the document index for product guidance and add a tool call to the transaction system for live order status (Correct answer)
  2. B. Use only the document index, because retrieval replaces the need for direct system access
  3. C. Fine-tune the model on yesterday's order export so it remembers the current statuses
  4. D. Expand the system prompt with a list of recent shipments and let the model infer the current answer

Correct answer: A

Explanation: Correct answer (A): This scenario has two different needs: document-grounded answers from manuals and live transactional data from an authoritative system. The blueprint states that if a scenario requires current business data or transaction status, tool access to authoritative systems may be more appropriate than retrieval from a static document index alone. Why the other options are wrong: - Option B: A static document index is not the best source for rapidly changing order status in an authoritative transaction system. - Option C: Fine-tuning on yesterday's export does not provide current order status and is not the default solution for dynamic business data. - Option D: Putting recent shipments into the prompt is not a reliable or scalable way to answer current transaction questions from an authoritative system.

Keep Practicing

Take the 10-question AI-103 quick-start test across all 5 domains, or return to the AI-103 practice test hub for 200+ questions, exam details, and a study plan.