McAfee-Secured Website

Exam Code: AI-103

Exam Name: Developing AI Apps and Agents on Azure

Certification Provider: Microsoft

Corresponding Certification: Microsoft Certified: Azure AI Apps and Agents Developer Associate

100% Microsoft AI-103 Guaranteed Success With Testking Training Materials

Get Real AI-103 Exam Questions & Expert Verified Answers as Experienced on the Real Exam!

187 Questions & Answers with Testing Engine

"Developing AI Apps and Agents on Azure Exam", also known as AI-103 exam, is a Microsoft certification exam.

Pass your tests with the always up-to-date AI-103 Exam Engine. Your AI-103 training materials keep you at the head of the pack!

Satisfaction Guaranteed

Money Back Guarantee

Testking provides hassle-free money back guarantee with our products. That is because we have 100% trust in the abilities of our professional and experience product team, and our record is a proof of that.

99.6% PASS RATE
Was: $137.49
Now: $124.99

Microsoft AI-103 Practice Test 9

Question 80

A team protects production resources with locks but its normal deployment service cannot update a resource during planned releases. What lock design should it review?

  1. Use delete-protecting locks without blocking authorized update operations.
  2. Change the model deployment name whenever a lock denies a management request.
  3. Apply delete/read-only locks indiscriminately to every dependent resource and ignore release failures.
  4. Remove all RBAC access control because resource locks provide complete identity security.

Explanation

Correct Answer

A:

Azure locks can interfere with legitimate management operations; scope and release processes must be planned.

Incorrect Answers

B:

Renaming client configuration does not remove the management lock that blocked the operation.

C:

Overbroad locks may block required write actions and create operational deadlocks.

D:

Locks and role-based authorization solve different problems and cannot replace each other.

Question 81

A CI pipeline can update Foundry prompts but must not delete compliance evidence or diagnostic storage. Which boundary should be defined?

  1. Make the pipeline subscription Owner to avoid occasional permission errors.
  2. Move audit logs into the agent's editable conversation memory for convenience.
  3. Scope CI/CD identity to minimum deployment actions; deny audit edits.
  4. Put a warning in pipeline documentation but retain universal delete permissions.

Explanation

Correct Answer

C:

Scoped RBAC separates deployment automation from retention and evidence-management permissions.

Incorrect Answers

A:

Owner would allow actions on protected audit systems that the requirement explicitly forbids.

B:

Agent-accessible memory is not an independent protected logging system.

D:

Documentation does not enforce a technical least-privilege boundary.

Question 82

An upcoming launch triples request volume, but the model's assigned quota remains at its current test level. Which readiness check should precede promotion?

  1. Use additional trace logging instead of modeling request and token rates.
  2. Raise max output length for every request to make more capacity available.
  3. Check deployment quota before promoting a high-traffic model.
  4. Assume a successful one-user test proves capacity for launch traffic.

Explanation

Correct Answer

C:

Quota constraints can cause 429 throttling even when application code is otherwise healthy.

Incorrect Answers

A:

Traces aid diagnosis but do not allocate inference throughput.

B:

Longer completions may consume more tokens and worsen quota exhaustion.

D:

Low-load functional tests do not validate sustained peak capacity.

Question 83

A prototype relies on a preview agent feature that may change before the product's regulated launch. What should the release plan explicitly capture?

  1. Document preview-feature dependencies before production rollout.
  2. Add a resource tag named ProductionReady as proof of general availability.
  3. Treat every preview API as contractually identical to a stable service release.
  4. Hide the preview dependency in documentation so reviewers do not delay deployment.

Explanation

Correct Answer

A:

Preview dependencies require explicit risk acceptance and contingency design, not an assumption of GA stability.

Incorrect Answers

B:

Tags cannot change vendor support lifecycle or preview guarantees.

C:

Preview interfaces and service guarantees can change, affecting production compatibility.

D:

Concealment prevents informed risk assessment and leaves operators without recovery guidance.

Question 84

An agent submits an approved travel reimbursement request to a company ledger API. New developers confuse the Foundry inference endpoint with the ledger transaction endpoint. Which separation is required?

  1. Post reimbursement transactions to the chat-completions endpoint's model parameter.
  2. Separate application business API from the agent execution endpoint.
  3. Embed the full finance admin key into the assistant's natural-language instructions.
  4. Use the ledger REST endpoint as the model deployment's inference URL.

Explanation

Correct Answer

B:

Agent reasoning and external business side effects have distinct interfaces, identities and authorization checks.

Incorrect Answers

A:

A model route is not the financial system's transactional API.

C:

Prompt text is not a secure substitute for a scoped authenticated integration.

D:

A business API does not implement the model's inference request contract.

Question 85

An app sends small chat requests to a deployed model. Calls start failing with an HTTP 429 during a burst, although individual prompts are short. Which diagnosis should the operator prioritize?

  1. The input exceeded the model context window in every request.
  2. All caller credentials were revoked during the burst.
  3. The search-index analyzer cannot recognize the user language.
  4. The deployment exceeded a request or token rate limit.

Explanation

Correct Answer

D:

429 throttling is consistent with quota pressure; compare request and token metrics against deployment limits.

Incorrect Answers

A:

A context-size failure is not established by a 429 during a burst of individually short requests.

B:

Authentication failures typically present authorization responses, not a capacity-rate throttle signal.

C:

Search-language analysis is a different subsystem and does not explain an inference deployment throttling response.

Question 86

An agent completes only a few long requests per minute. Each contains a very large retrieved context, and the model reaches its quota despite low call counts. Which quota dimension is most relevant?

  1. Tokens-per-minute quota exhaustion.
  2. Number of separate employees who have bookmarked the web application.
  3. Maximum Azure subscription tag count.
  4. Number of text files stored outside the retrieval index.

Explanation

Correct Answer

A:

Long prompt contexts can exhaust tokens-per-minute capacity despite a low requests-per-minute count.

Incorrect Answers

B:

A bookmark count is unrelated to inference token consumption until those users send actual requests.

C:

Resource tags are management metadata and do not express the per-minute model token budget.

D:

Unused external files do not directly determine token consumption on these requests.

Question 87

A support chatbot intermittently receives throttling responses from a deployed model during high traffic. The response indicates a temporary rate limit. How should the client behave?

  1. Immediately retry without a delay in an unlimited loop until one request succeeds.
  2. Convert the failure into a success response containing a fabricated answer.
  3. Retry a limited number of times using server guidance or exponential backoff, then surface failure if necessary.
  4. Delete the source search index whenever a model invocation is throttled.

Explanation

Correct Answer

C:

Bounded delay-based retries reduce thundering-herd effects and retain an explicit failure path.

Incorrect Answers

A:

Aggressive unbounded retries can intensify throttling and consume unnecessary system resources.

B:

Faking success hides operational failure and undermines answer trustworthiness.

D:

Index deletion is destructive and irrelevant to the short-term model service capacity signal.

Question 88

Several services share one model deployment. A nightly process suddenly launches hundreds of parallel generation requests and crowds out interactive traffic. What client-side change most directly limits interference?

  1. Increase the nightly process prompt size so each request consumes more tokens.
  2. Apply client-side concurrency limits to protect a shared deployment.
  3. Remove telemetry and stop recording the source of the requests.
  4. Give every service an identical request ID so they share a single execution.

Explanation

Correct Answer

B:

A bounded request rate protects the shared deployment and improves fairness for latency-sensitive consumers.

Incorrect Answers

A:

Larger prompts can worsen the shared quota contention rather than mitigate it.

C:

Suppressing observation does not reduce real request volume or contention.

D:

Correlating requests with identical identifiers does not merge or throttle the underlying inference calls.

Question 89

Two product groups share one model deployment. An overnight document classifier starves the customer chatbot during business hours. What should operations plan?

  1. Allocate or isolate quota by workload priority and measure actual request and token demand.
  2. Use the same unrestricted shared quota and ask both teams to retry aggressively.
  3. Increase temperature on the chatbot to make throttling less visible.
  4. Remove usage telemetry so neither group can see the other team's traffic.

Explanation

Correct Answer

A:

Capacity partitioning or separate deployments protects interactive workloads from batch consumption peaks.

Incorrect Answers

B:

Uncoordinated retries can amplify contention without protecting customer-facing traffic.

C:

Sampling parameters do not reserve or allocate requests-per-minute capacity.

D:

Without workload measurements the contention cannot be diagnosed or fairly managed.

Question 90

A finance team compares token expenses for prompts containing 20,000 tokens versus responses containing 500 tokens. The price schedule differentiates input and output. What calculation should it use?

  1. Use the response character count as the entire bill without measuring billed token units.
  2. Charge only on the number of HTTP requests, ignoring token usage.
  3. Price input and output tokens separately and sum the charges.
  4. Multiply all tokens by the output rate without distinguishing categories.

Explanation

Correct Answer

C:

Prompt and completion tokens can carry different unit pricing; both terms contribute to total cost.

Incorrect Answers

A:

Characters do not map directly to billable input/output token consumption.

B:

Request count alone omits variable prompt and completion volume.

D:

Applying one rate to both categories can misstate costs under differential pricing.

Developing AI Apps and Agents on Azure (AI-103) Practice Tests