Developing AI Apps and Agents on Azure (AI-103) practice questions

Domain coverage 5 domains

• Plan and manage an Azure AI solution
• Implement generative AI and agentic solutions
• Implement computer vision solutions
• Implement text analysis solutions
• Implement information extraction solutions

Sample questions 9 of 88 shown

1. Q1 Plan and Manage an Azure AI Solution (25–30%)

   A Python support agent deploys to pre-production and production environments through the same pipeline. Each environment has its own Foundry project and model deployments, but both deployments use the same model family. Release gates must verify that production traffic uses the production project and expected deployment. What should you implement?

   1. Emit traces with the Foundry project endpoint and deployment name on each call
   2. Run safety evaluations on production responses
   3. Monitor aggregate token usage by model family
   4. Compare generated answers against a fixed golden dataset

   The goal is to verify connection context, not model quality or safety. Trace logs that record the Foundry project endpoint and deployment name used by each request directly prove the application is using the expected production context.
2. Q2 Plan and Manage an Azure AI Solution (25–30%)

   A Foundry project using CI/CD deploys an enterprise procurement agent. The approved baseline requires business system tools to use user-assigned managed identity and least-privilege role policies, and must prohibit storing API keys. A drift scan shows the erpLookup tool in production now authenticates via an API key stored in app configuration. What should you do?

   1. Add a prompt guardrail telling the model not to leak secrets
   2. Disable the erpLookup tool until the next monthly release
   3. Redeploy the approved managed identity tool configuration and remove the API key
   4. Leave the setting unchanged since endpoints and tool schema match

   Drift occurred on the tool's authentication setting, not the tool schema. Restoring the approved managed identity and role policy eliminates the stored secret risk while not blocking authorized procurement order queries.
3. Q3 Plan and Manage an Azure AI Solution (25–30%)

   A Foundry project hosts a support agent that uses Azure AI Search for knowledge retrieval and calls a refund-eligibility function. Before production, stakeholders ask if the agent gives correct policy answers, cites retrieved sources, and blocks unsafe refund guidance. Which evaluation plan best meets this request?

   1. Track token spend per conversation and select the lowest-cost model
   2. Compare p95 latency across different model deployments
   3. Run concurrent user load tests against the agent endpoint
   4. Evaluate traces for groundedness, answer correctness, safety, and tool call correctness

   The scenario asks about output quality and agent behavior, not operational efficiency. A suitable Foundry evaluation plan should use conversation traces to test correctness, groundedness, safety, and whether the refund function was called appropriately.
4. Q4 Plan and Manage an Azure AI Solution (25–30%)

   You maintain a Foundry RAG application that answers HR policy questions from documents indexed in Azure AI Search. The team changes retrieval from top-3 semantic hybrid results to top-5 and modifies the base prompt. You must verify this change does not reduce answer grounding or citation accuracy on known issues. What should you do?

   1. Run a baseline-candidate regression evaluation with groundedness checks
   2. Replace hybrid search with vector-only search for all queries
   3. Increase chunk size and redeploy after smoke testing
   4. Monitor production user feedback after deployment

   Prompt and retrieval changes alter which passages are used and how answers cite them. Regression evaluation uses the same test cases against the current and candidate workflows to detect regressions in groundedness or citation before deployment.
5. Q5 Plan and Manage an Azure AI Solution (25–30%)

   A Foundry project uses CI/CD to deploy an enterprise procurement agent. A drift scan reveals the erpLookup tool authenticates via API key instead of managed identity. The endpoint and tool schema remain unchanged. What should you do?

   1. Add a prompt guardrail
   2. Disable erpLookup
   3. Redeploy the approved managed identity tool configuration and remove the API key
   4. Leave as-is since endpoints match

   The drift is in authentication settings, not tool schema. Redeploying the approved managed identity configuration restores compliance without changing functionality.
6. Q6 Plan and Manage an Azure AI Solution (25–30%)

   You are designing a Microsoft Foundry HR benefits agent. The agent uses Azure AI Search for handbook knowledge retrieval and a payroll tool that can submit compensation change requests. The solution must block unsafe or prompt-injected input, avoid retrieval based on restricted content, limit tool calls to approved actions, prevent unsafe responses, and require HR approval before any payroll change is submitted. Which guardrail placement strategy should you use?

   1. Use system prompts to describe all restrictions and rely on conversation memory to avoid unsafe requests
   2. Run one safety check on the final model response only and send payroll changes to HR review after execution
   3. Guard input before retrieval, filter content after retrieval, enforce tool policy before execution, scan output before returning, and gate approval before submission
   4. Require HR approval for every user conversation turn and skip retrieval and tool guardrails

   Guardrails should be placed as close to the risk they control as possible. Each checkpoint in this workflow requires a separate guardrail so unsafe content or unauthorized actions are blocked before they affect the next stage.
7. Q7 Plan and Manage an Azure AI Solution (25–30%)

   A team in a Microsoft Foundry project builds a RAG-based enterprise assistant. New policy PDFs sync from SharePoint, undergo OCR/layout extraction, chunking, embedding, and indexing into Azure AI Search. Users report that some newly uploaded policies are not being referenced. The team has evaluated generated answer quality and needs a data ingestion quality observability signal. What should they monitor?

   1. Search click-through rate for referenced documents
   2. Source-to-index completeness and freshness per ingestion run
   3. LLM token usage and response latency per conversation
   4. Safety evaluator scores on generated responses

   Ingestion quality monitoring should verify the path from source content to searchable grounded data. A useful signal is whether each source PDF is processed, chunked, embedded, indexed, and refreshed on time.
8. Q8 Plan and Manage an Azure AI Solution (25–30%)

   A Foundry project support agent sees intermittent latency spikes during morning login. Model deployment traces show a 429 RateLimitExceeded pattern occurring in 2-minute bursts. The application currently retries up to three times immediately on failed calls. What is the best next fix?

   1. Increase the model temperature for the agent deployment
   2. Increase retry concurrency to clear the backlog
   3. Throttle and queue model calls respecting Retry-After headers
   4. Rebuild the search index with larger vectors

   The symptom is a model deployment rate limit. The 429 status with token-per-minute limit message and retry-after-ms header indicates the application should slow down and schedule requests rather than retrying immediately.
9. Q9 Plan and Manage an Azure AI Solution (25–30%)

   You are evaluating a Microsoft Foundry RAG assistant that uses Azure AI Search for knowledge retrieval to answer HR policy questions. Groundedness scores are low, but trace logs show answers are consistent with retrieved chunks. For failing questions, the relevant policy pages exist in the source files but are not in the top retrieval results. What should you prioritize?

   1. Run content safety evaluations on failing answers
   2. Increase system prompt verbosity and retest for fluency
   3. Compare answer length against model token limits
   4. Measure retrieval recall@k against the marked relevant policy pages

   Trace evidence points to a retrieval failure, not a generation failure. If the model is answering from retrieved context but the correct source passages are missing, measuring retrieval recall against marked relevant content is the key verification.

Answer all 88 questions with instant feedback →

Reader reviews 124 ratings

• Solid Microsoft prep material. The AI-103 questions are well-written and the answer rationales are thorough.

  ★★★★★★ Madison T.

• Got the AI-103 certification after three weeks of self-study with this question bank. Very happy with the purchase.

  ★★★★★★ Victoria K.

• Lifetime access was the selling point for me. Got sidetracked for a few months and came back — everything was still there for AI-103.

  ★★★★★★ Layla S.

• Couldn’t have passed the AI-103 exam without this. The questions are challenging, the explanations are thorough, and the value is unbeatable.

  ★★★★★ Hannah L.

• Best investment for Microsoft certification prep. The question bank for AI-103 is comprehensive and mirrors the real exam perfectly.

  ★★★★★ Sarah M.

• The progress tracking feature for AI-103 really motivated me. Seeing my improvement over time was incredibly satisfying.

  ★★★★★★ Daniel H.

Frequently asked questions

Free study resources

Community-verified analysis of 65 questions from real test-taker discussions — 20 deep analyses and 0 FAQs.

More Microsoft exams