cert
cert-1
cert-2

Pass Microsoft AI-103 Exam in First Attempt Guaranteed!

Get 100% Latest Exam Questions, Accurate & Verified Answers to Pass the Actual Exam!
30 Days Free Updates, Instant Download!

cert-5
cert-6
AI-103 Exam - Verified By Experts
AI-103 Premium File

AI-103 Premium File

$89.99
$98.99
  • Premium File 187 Questions & Answers. Last Update: Nov 15, 2026

Whats Included:

  • Latest Questions
  • 100% Accurate Answers
  • Fast Exam Updates
 
$98.99
$89.99
accept 397 downloads in the last 7 days
block-screenshots
Microsoft AI-103 Practice Test 21

Question 200

A multi-environment Foundry app must avoid embedded keys, select the right production project and use an unattended identity. Which THREE settings directly meet those goals? Choose THREE.

  1. Configure the application's correct Foundry project connection and deployment mapping
  2. Hard-code the developer's personal interactive model session as the only credential for the unattended production workload.
  3. Use an authorized production managed identity instead of a personal developer session
  4. Store only nonsecret endpoints in configuration and use keyless token authentication
  5. Choose production project and tenant from display-name similarity alone, without validating the actual resource or tenant identity.
  6. Use one shared global connection string containing an embedded secret for staging and production, without an environment-specific deployment mapping.

Explanation

Correct Answers

A:

Applications need the intended project scope and deployment route, not just an arbitrary model family name. Explicit resource mapping, managed identity and token authentication address routing and secure access.

C:

Unattended services need their own governed principal with revocable scoped access.

D:

Keyless credentials prevent leaked static secrets from becoming service access paths.

Incorrect Answers

B:

A developer session is not a noninteractive production identity and cannot provide reliable keyless service authorization after the developer signs out.

E:

Display names can overlap across environments. Production connections need their real resource identifiers, scoped authorization and controlled configuration.

F:

A shared secret and ambiguous connection violate the separate production context and unattended keyless-authentication requirements.

Question 201

An agent prompt says only 'be helpful', causing it to initiate unrelated purchases. What should the agent specification contain?

  1. Tell the agent it may purchase any item whenever doing so seems helpful, without any authorized-task boundary.
  2. Define a bounded role, explicit task goals and permitted action scope
  3. Maintain a short ordered conversation and approval state tied to the task
  4. Expire stale task context and retrieve current policy before sensitive decisions

Explanation

Correct Answer

B:

Agent autonomy needs a precise responsibility and limits so unrelated actions are outside its authority. Here the relevant capability must let the team constrain an agent to an authorized task. An alternative targeting collect missing mandatory task context solves another problem.

Incorrect Answers

A:

Open-ended helpfulness is not permission for financial side effects. The agent specification needs explicit goals, tool restrictions, confirmation and authorization before account-changing actions.

C:

Relevant conversation state keeps authorization and progress coherent across tool cycles. That capability is appropriate for retain task-scoped user decisions during turns, not this separate task of finding how to constrain an agent to an authorized task.

D:

Short-lived task facts should not override newer authoritative policy evidence. Here the needed outcome is to constrain an agent to an authorized task; an option designed to limit stale memory from overruling current rules would not directly provide it.

Question 202

Model deployments are configured by hand and often differ between test and production. What pipeline improvement helps?

  1. Run post-deployment connectivity and permission smoke tests for critical tools
  2. Attach prompt, model and application version identifiers to release and trace records
  3. Require a recorded protected-environment approval gate for safety changes
  4. Declare supported AI infrastructure and deployment settings as version-controlled automation

Explanation

Correct Answer

D:

Infrastructure-as-code reduces drift and makes resource changes reviewable and reproducible. Here the relevant capability must let the team reproduce deployment infrastructure across environments. An alternative targeting validate actual deployed tool dependencies solves another problem.

Incorrect Answers

A:

Build tests cannot validate environment-specific tool networking and authorization. This helps when a team needs to validate actual deployed tool dependencies; it does not establish how to reproduce deployment infrastructure across environments.

B:

Version provenance makes regression triage and controlled rollback possible. The function serves to trace behavior to the precise released configuration, whereas the actual requirement is to reproduce deployment infrastructure across environments.

C:

Release approvals must be checked by the pipeline, not just requested in a prompt. That capability is appropriate for enforce auditable approval of safety configuration, not this separate task of finding how to reproduce deployment infrastructure across environments.

Question 203

A customer uploads an image of two nearly identical components. The assistant needs to compare visible distinguishing marks. How should analysis be structured?

  1. Exclude unsupported orientation claims unless grounded in trusted context
  2. Request evidence-grounded side-by-side comparison of marked features
  3. Send the image and question to a multimodal visual-understanding model
  4. Give a generic impression that the components look similar without identifying the observable differences.

Explanation

Correct Answer

B:

A comparative visual task should identify observed differences rather than assert unsupported product identity. This resolves the specific need to compare two visually similar components carefully, while a measure that helps prevent unsupported visual-context inferences addresses a different requirement.

Incorrect Answers

A:

A visual image often cannot establish compass directions without additional evidence. Its primary purpose is to prevent unsupported visual-context inferences; the stated constraint instead demands a way to compare two visually similar components carefully.

C:

The model must inspect the visual scene rather than answer from the user's words alone. This helps when a team needs to answer an image-grounded visual count; it does not establish how to compare two visually similar components carefully.

D:

The user requests comparative evidence, not an overall similarity label. Identify the visible, testable differences between the two components and avoid inventing features hidden by the photograph.

Question 204

A generative assistant must translate a contract clause but must not explain or summarize it. What prompt/output rule helps?

  1. Paraphrase the contract with extra legal interpretation whenever a term is ambiguous
  2. Request faithful translation only and validate that no extra commentary was added
  3. Use Azure Translator with a managed terminology strategy where supported
  4. Preserve and compare numeric, currency and identifier tokens across translations

Explanation

Correct Answer

B:

Translation output should preserve meaning and scope rather than introduce advisory material. The practical requirement is to avoid additions when translating legal text; a solution built to select the correct source language automatically would not meet that same need.

Incorrect Answers

A:

The requirement is translation, not freeform legal analysis. Extra interpretation can alter obligations; terminology and intended meaning must be preserved and uncertain clauses flagged.

C:

A dedicated translation flow can provide multilingual conversion while preserving approved terminology. The function serves to translate support text with consistent terminology, whereas the actual requirement is to avoid additions when translating legal text.

D:

Translation must not silently alter financial facts or reference numbers. That capability is appropriate for protect exact business values during translation, not this separate task of finding how to avoid additions when translating legal text.

Question 205

A vector index and its query API both use 1,536-dimensional arrays, but queries changed to a new embedding-model family and search relevance collapsed. What compatibility issue must the engineer resolve?

  1. Deduplicate or diversify ranked results while retaining strongest evidence
  2. Use compatible embedding-model versions for indexed documents and queries, re-embedding the corpus when the representation changes.
  3. Use hybrid keyword and vector retrieval in one query workflow
  4. Assume equal vector dimensionality guarantees semantic compatibility across any embedding models.

Explanation

Correct Answer

B:

The index field dimensions already match the query payload. The failure is semantic compatibility: vector distances are meaningful only when indexed documents and incoming queries share the same embedding space. Re-embed the index or restore the matching query encoder. Equal dimensionality does not make different embedding models’ coordinate spaces compatible; both query and document vectors must use a compatible representation.

Incorrect Answers

A:

Source diversity and duplicate filtering improve usable coverage for grounded generation. This is a separate capability and does not directly keep indexed document and query embeddings in a compatible semantic vector space, beyond matching shape.

C:

Hybrid retrieval covers rare identifiers and semantic descriptions that neither method handles perfectly alone. This is a separate capability and does not directly keep indexed document and query embeddings in a compatible semantic vector space, beyond matching shape.

D:

The same numeric vector length is necessary for storage but insufficient to preserve semantic distance meaning across incompatible embedding models.

Question 206

A coordinator receives a specialist's final answer but none of the supporting tool evidence. What should the delegated result contract include?

  1. Return structured findings with source references, uncertainty and the specialist's task outcome.
  2. Return only a confidence score without any source or completed task description.
  3. Return the last tool's raw output with no structured summary or status.
  4. Provide only the specialist's name and execution duration.

Explanation

Correct Answer

A:

A bare conclusion is hard to audit. Passing cited evidence and confidence enables the coordinator to judge and synthesize responsibly.

Incorrect Answers

B:

A score cannot show which findings are supported or what work the specialist performed.

C:

Raw tool data can omit uncertainty, relevance and usable source references for the coordinator.

D:

Attribution and runtime do not establish the substance or evidence of delegated findings.

Question 207

A production model update should be blocked if quality regressions exceed an approved threshold. What CI gate is needed?

  1. Run a representative evaluation suite with defined failure thresholds before promotion
  2. Accept any newly trained model whenever its deployment succeeds, regardless of evaluation-set failures or regression thresholds.
  3. Use workload identity or managed federation rather than embedded pipeline secrets
  4. Require a recorded protected-environment approval gate for safety changes

Explanation

Correct Answer

A:

A measurable evaluation gate prevents known regressions from reaching users. This directly supports the need to test model behavior before promotion, unlike a mechanism primarily intended to restore a known working prompt-model pair.

Incorrect Answers

B:

A deployable artifact can still regress on safety, correctness or task quality. A release gate must compare reproducible evaluation results to approved thresholds before promotion.

C:

Federated or managed identities avoid long-lived credentials committed to repositories. Here the needed outcome is to test model behavior before promotion; an option designed to remove static deployment pipeline secrets would not directly provide it.

D:

Release approvals must be checked by the pipeline, not just requested in a prompt. Its primary purpose is to enforce auditable approval of safety configuration; the stated constraint instead demands a way to test model behavior before promotion.

Question 208

A multimodal inspector analyzes several damaged-package images and must avoid mixing evidence, inventing obscured text or ignoring scale labels. Which THREE guardrails help? Choose THREE.

  1. Read and preserve visible scale labels, measurement units and reference markings in each photograph.
  2. Provide all permitted images and request structured detailed observations per image
  3. Infer the direction and scale of every photograph from unverified visual cues even when orientation and measurement markings are absent.
  4. Merge all photographs into one short unlabeled caption and omit which image or scale label supports each reported defect.
  5. Express uncertainty when the image lacks sufficient resolution to support recognition
  6. Merge every photograph into one unlabeled observation without identifying which image supports a finding.

Explanation

Correct Answers

A:

Measurement labels and reference scale markings give important context when comparing damage and should be preserved rather than inferred. The inspector should preserve each image’s source identity and visible measurements, and explicitly acknowledge unreadable text instead of inventing it. Those controls support trustworthy comparisons of damaged packaging.

B:

The combined inspection must preserve which image supplied each damage observation, respect visible scale units, and explicitly flag details that are unreadable. These constraints stop evidence from one package being attributed to another or measured without adequate pixels.

E:

Visual evidence limits must govern claims; hallucinating obscured text can cause operational errors.

Incorrect Answers

C:

Unverified orientation claims are unsupported evidence. Each photo should be interpreted only within its visible markings and the approved contextual information.

D:

An unlabelled summary can conflate damage from different photographs and omit scale. Structured observations must preserve image identity and visible measurements.

F:

Losing the image-to-evidence mapping makes it impossible to distinguish damage visible in different photos.

Question 209

An LLM summarizes a meeting but invents a decision nobody made. How should the summary be reviewed?

  1. Trust any plausible summary statement if its prose is grammatical.
  2. Require each reported action item to match the provided transcript span or record uncertainty.
  3. Compare the summary against unrelated prior meeting agendas.
  4. Ask the model to assert certainty even when no transcript evidence exists.

Explanation

Correct Answer

B:

Grounded summaries should distinguish actual commitments from plausible inferred next steps.

Incorrect Answers

A:

Well-written action items can still invent approvals or commitments absent from the meeting.

C:

An agenda for another session cannot confirm what was actually decided in this transcript.

D:

Expressing confidence in an invented meeting decision does not make it true. Require a matching transcript passage, attribution to the participants and a review path for claims without evidence.

Question 210

A document index uses metadata-based access filters, but a query forgets to send the tenant filter. What enforcement is missing?

  1. Require mandatory authorization filters at the search service boundary, not as an optional prompt instruction.
  2. Let each calling application choose whether to include a tenant filter.
  3. Filter only the final answer after retrieving documents from every tenant.
  4. Rely on a system prompt saying not to mention other tenants.

Explanation

Correct Answer

A:

An omitted filter can expose unauthorized text to the model; server-side enforcement should reject unscoped searches.

Incorrect Answers

B:

An optional client filter can be omitted, allowing cross-tenant documents into retrieval results.

C:

The unauthorized source has already entered processing before answer-level filtering.

D:

A prompt instruction about keeping tenants separate is not an enforceable search authorization filter. The retrieval tool must apply trusted tenant and principal claims to every indexed-document query before returning snippets.

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