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 3

Question 20

A legal analysis agent must reconcile several conflicting contract provisions and explain multi-step exceptions. Small baselines repeatedly fail this evaluation despite adequate latency. What model-selection approach fits?

  1. Force the cheapest small model into production without revisiting the failed benchmark.
  2. Use lexical search ranking as the entire contract-reconciliation explanation.
  3. Choose a large reasoning model for complex multi-constraint analysis.
  4. Use only text embeddings and return their raw floating-point vectors to reviewers.

Explanation

Correct Answer

C:

The observed complex reasoning failures justify testing greater reasoning capability rather than optimizing latency alone.

Incorrect Answers

A:

The documented accuracy failure makes that option unsuitable even if inference is inexpensive.

B:

Retrieval alone does not perform the required cross-clause interpretation and synthesis.

D:

Vectors are representations for similarity, not a legal reasoning narrative.

Question 21

A developer tool must explain the effect of a patch across Python modules and unit tests. Business-oriented summaries miss dependency and symbol relationships. What should be considered?

  1. Use a receipt-extraction analyzer for the repository source files.
  2. Use a general-purpose summarizer that reads only commit messages, not modified code.
  3. Rank files only by their creation timestamps and report the newest as the explanation.
  4. Select a code-capable model for repository analysis.

Explanation

Correct Answer

D:

Code-aware evaluation checks whether the selected model understands syntax, references and test implications.

Incorrect Answers

A:

Invoice/document field extraction does not establish code dependency reasoning.

B:

Commit summaries cannot demonstrate effects across changed module symbols, call sites and tests without inspecting the diff itself.

C:

File age does not explain execution flow or semantic effects of a patch.

Question 22

Field engineers upload photographed handwritten maintenance notes mixed with equipment diagrams and ask for a diagnosis summary. A text-only deployment currently receives filenames. Which input capability is missing?

  1. Run OCR on the photo and discard any non-text visual context before interpreting the maintenance note.
  2. Generate a new maintenance photograph and analyze that instead of the submitted evidence.
  3. Send the photo's byte count as the sole evidence for troubleshooting.
  4. Use a vision-capable multimodal model with the actual photos and textual question.

Explanation

Correct Answer

D:

Image input permits grounded interpretation of handwritten marks and spatial diagram context alongside the scenario.

Incorrect Answers

A:

OCR can extract visible words, but it cannot provide non-text visual context when the task requires joint interpretation.

B:

A synthetic replacement may differ materially from the engineer's observed failure.

C:

File size does not reveal technical observations or component relationships.

Question 23

A search page already retrieves the three most relevant policy paragraphs, but employees expect one coherent answer explaining how the clauses interact. What additional component is needed?

  1. Return only cosine similarity scores and treat them as a policy explanation.
  2. Synthesize a cited answer from retrieved passages with a generative model.
  3. Index the user question as a new source document and omit the retrieved clauses.
  4. Disable the search index and answer from the model's original training memory.

Explanation

Correct Answer

B:

Retrieval selects the passages; generation can synthesize their interaction into a readable response.

Incorrect Answers

A:

Similarity values rank documents but do not explain requirements or exceptions.

C:

The scenario is not authoritative evidence for the policy answer.

D:

The requested answer depends on private policy provisions that were actually retrieved.

Question 24

An approval workflow must always verify a case ID, retrieve a fixed policy, and invoke a model for a short justification. An autonomous agent sometimes skips the mandatory lookup. What orchestration is better?

  1. Delete the approval step from the agent instruction and trust the model's internal judgment.
  2. Use a deterministic workflow engine that enforces required steps and includes the model only where needed.
  3. Increase agent temperature so it may choose more varied step orders.
  4. Let the model produce arbitrary tool names and rely on text parsing to execute them.

Explanation

Correct Answer

B:

Explicit ordered orchestration enforces compliance checkpoints while still using generation for the interpretive step.

Incorrect Answers

A:

This removes a business requirement and fails to enforce independent control.

C:

Higher randomness makes a mandatory ordered workflow less dependable.

D:

Unvalidated free-form tool commands do not guarantee the mandatory lookup or safe execution path.

Question 25

An agent requests a warehouse count. Developers currently scrape pseudo-command strings from natural-language responses, occasionally parsing an unintended write operation. How should calls be represented?

  1. Search every response for a verb and run the first shell command that follows it.
  2. Allow the model to change database schema when its answer mentions missing inventory.
  3. Use the model's supported structured function/tool calling with a constrained read-only operation schema.
  4. Treat the model's description of the warehouse count as proof the API was invoked.

Explanation

Correct Answer

C:

Typed calls separate tool invocation intent from free-form assistant prose and permit application-side validation.

Incorrect Answers

A:

Text heuristics can misinterpret harmless text or execute an unapproved action.

B:

A live inventory lookup requires read access, not unrestricted data definition privileges.

D:

A language response is not a successful authorized tool result.

Question 26

An order agent calls a fulfillment API that accepts `action`, `sku`, and `quantity`; accidental strings like 'remove all' must never become executable actions. What contract reduces risk?

  1. Accept every action name that the model can produce as long as it is grammatical.
  2. Validate only response length after the API has already committed the order.
  3. Permit a single free-text payload and trust the agent to infer the correct JSON shape.
  4. Use schema-constrained tool arguments for reliable API integrations.

Explanation

Correct Answer

D:

Schema constraints narrow the accepted arguments and reject unsupported operations at the execution boundary.

Incorrect Answers

A:

Grammatical text does not prove the action belongs to the permitted business API.

B:

Post-execution length checks cannot prevent malformed or unauthorized business actions.

C:

Free-form payloads bypass the declared action vocabulary and complicate reliable validation.

Question 27

A production Foundry client currently embeds a long-lived service key in a public container image. The runtime has an assigned workload identity and appropriate permissions. Which authentication change is preferred?

  1. Use managed identity or equivalent keyless Entra authentication with least-privilege access to the service.
  2. Store the API key in every source-control branch with a restrictive README instruction.
  3. Disable authentication on the model deployment because the container image is private.
  4. Obfuscate the same API key in a client-side JavaScript constant.

Explanation

Correct Answer

A:

Scoped workload identity avoids distributing reusable embedded credentials while retaining auditable access control.

Incorrect Answers

B:

Versioned copies broaden key exposure and do not replace appropriate secret management.

C:

Private image storage alone does not authorize inference requests or replace endpoint authentication.

D:

Obfuscation does not protect secrets delivered to the client and does not provide per-workload identity.

Question 28

A public chat app must screen harmful requests before invoking an expensive generative endpoint. Developers propose adding only an after-generation warning label. Which control is suitable?

  1. Set the model's token limit to one and treat short output as equivalent to safe output.
  2. Use search result ranking to decide whether the user's request contains disallowed content.
  3. Pick content moderation tooling for pre-generation safety checks.
  4. Rely exclusively on a user-facing label displayed after every generated answer.

Explanation

Correct Answer

C:

Pre-generation moderation can reject disallowed requests before potentially unsafe or costly output is produced.

Incorrect Answers

A:

A shorter response may still violate the policy; token budgets are not moderation controls.

B:

Retrieval relevance ranking does not implement an input content-safety classification policy.

D:

A post hoc label does not prevent prohibited content from being processed or returned.

Question 29

A team has an approved Azure AI Search service. Two Foundry projects need access, but developers currently paste search admin keys into every agent prompt. What should they configure?

  1. Configure an approved project-scoped search connection.
  2. Re-create a separate unmanaged search service for each individual conversation.
  3. Write the admin API key in the agent's system message for every session.
  4. Use the generative deployment name in place of a search connection identifier.

Explanation

Correct Answer

A:

Managed project connections make the intended resource and permissions explicit without leaking credentials into model text.

Incorrect Answers

B:

Per-conversation copies increase management overhead and fragment the approved searchable corpus.

C:

System prompt text is not a secure secret store and grants unnecessarily powerful access.

D:

A model deployment identifier does not authenticate or select the search data source.

Question 30

An opt-in study assistant should remember each student's preferred explanation language after they log out. Chat threads expire after inactivity. Where should the preference live?

  1. Embed all user preferences into a global system prompt shared by every account.
  2. Assume the model will infer the language from training data without any profile state.
  3. Choose durable storage for customer opt-in memory across sessions.
  4. Store the preference only in the most recent in-memory message object.

Explanation

Correct Answer

C:

Durable scoped storage can retrieve a returning user's preferences without preserving unrelated transient chat content.

Incorrect Answers

A:

A shared prompt risks cross-user leakage and lacks user-level consent and deletion boundaries.

B:

Pretraining cannot reliably remember a new personal opt-in preference between independent sessions.

D:

Transient thread content can disappear when the session ends and cannot provide cross-session continuity.

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