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 38

Question 370

A semi-autonomous workflow may create draft emails but must never send them without review. What state machine pattern fits?

  1. Transition from draft to send only after recorded authorized approval
  2. Permit outbound message-send tools whenever a draft was saved successfully for the same tenant.
  3. Enforce an iteration/time budget with no-progress detection and explicit failure
  4. Allow bounded read retries but require idempotency for writes

Explanation

Correct Answer

A:

A code-enforced approval transition prevents premature outbound communication. The practical requirement is to separate drafting from authorized sending; a solution built to constrain tools to a user's authorized tenant would not meet that same need.

Incorrect Answers

B:

Creating a draft and authorizing delivery are separate transitions. Hold the workflow in awaiting-review and invoke the send tool only after the approved action.

C:

Bounded tool execution prevents runaway use even when the agent cannot self-recognize a loop. The function serves to terminate runaway autonomous actions, whereas the actual requirement is to separate drafting from authorized sending.

D:

Read-only and side-effecting operations have different replay risks and need distinct handling. That capability is appropriate for differentiate retry safety between reads and writes, not this separate task of finding how to separate drafting from authorized sending.

Question 371

A managed identity has access to the Foundry project but calls a protected Search index and receives forbidden. What should be inspected?

  1. Verify separate Azure AI Search data-plane authorization for that identity
  2. Grant the identity additional access to the Foundry project and assume those privileges automatically authorize every Azure AI Search index.
  3. Reduce the principal's permissions to model-inference roles at the smallest applicable scope
  4. Remove its identity role assignments and disable the unused workload identity

Explanation

Correct Answer

A:

Project access does not automatically imply permissions to query another service's data. This directly supports the need to authorize search data access separately from project access, unlike a mechanism primarily intended to replace embedded credentials with workload identity.

Incorrect Answers

B:

Foundry project access does not automatically confer Search data-plane rights. The identity needs appropriate search-index permissions and should be checked at the correct Search resource scope.

C:

Broad Owner grants are unrelated to the specific model call and unnecessarily enable resource changes. Here the needed outcome is to authorize search data access separately from project access; an option designed to enforce least privilege for model inference would not directly provide it.

D:

Deleting only an application configuration file does not necessarily revoke the workload's authorization. Its primary purpose is to revoke retired workload access completely; the stated constraint instead demands a way to authorize search data access separately from project access.

Question 372

A model deployment's answer quality drops after a backend revision, but the prompt text is unchanged. What release evidence should be compared?

  1. Assume identical prompt text means model behavior must be identical.
  2. Compare only the application's public URL without checking deployed model revisions.
  3. Compare model-version manifests and regression evaluation results for the two deployments.
  4. Disable the regression tests after the release to reduce telemetry noise.

Explanation

Correct Answer

C:

Model revisions can change behavior independently of prompts; pinning and evaluation help isolate the regression.

Incorrect Answers

A:

Model versions and deployment configurations can change outputs even with unchanged prompts.

B:

An endpoint URL is not enough to identify the exact backend model and policy configuration.

D:

Removing evaluation would conceal the behavior change instead of explaining its cause.

Question 373

A contractor's model-calling service is decommissioned. What access revocation is necessary?

  1. Keep the decommissioned contractor workload's role assignments active in case its model API is restarted.
  2. Authenticate the workload with managed identity and supported Azure RBAC
  3. Reduce the principal's permissions to model-inference roles at the smallest applicable scope
  4. Remove its identity role assignments and disable the unused workload identity

Explanation

Correct Answer

D:

Deleting only an application configuration file does not necessarily revoke the workload's authorization. The decisive difference is whether the design can revoke retired workload access completely; the competing function serves to authenticate automated deployments without client secrets.

Incorrect Answers

A:

Retired workload identities must not retain standing resource grants. Revoke applicable roles and disable or delete the unused identity under the approved decommissioning procedure.

B:

Managed identity issues scoped tokens without storing resource keys in the app environment. Here the needed outcome is to revoke retired workload access completely; an option designed to replace embedded credentials with workload identity would not directly provide it.

C:

Broad Owner grants are unrelated to the specific model call and unnecessarily enable resource changes. Its primary purpose is to enforce least privilege for model inference; the stated constraint instead demands a way to revoke retired workload access completely.

Question 374

A task requires repeating harmless reads but external writes could duplicate bills if retried. How should retry policies differ?

  1. Require a verified human approval artifact at the transfer tool boundary
  2. Retry all external billing writes as freely as read-only tenant-scoped queries, without a stable operation identifier.
  3. Allow bounded read retries but require idempotency for writes
  4. Stop side effects while retaining governed logs and trace evidence

Explanation

Correct Answer

C:

Read-only and side-effecting operations have different replay risks and need distinct handling. Here the relevant capability must let the team differentiate retry safety between reads and writes. An alternative targeting gate high-value transfers on human authorization solves another problem.

Incorrect Answers

A:

Approval must be enforced by the application handling the real transfer, not only by prompt text. This helps when a team needs to gate high-value transfers on human authorization; it does not establish how to differentiate retry safety between reads and writes.

B:

Reads can often be retried safely, while writes risk duplicate bills. Use idempotency keys and status checks for state-changing calls regardless of tenant scope.

D:

Safe termination should preserve information necessary to diagnose and review the violation. Here the needed outcome is to differentiate retry safety between reads and writes; an option designed to halt unsafe actions without destroying evidence would not directly provide it.

Question 375

A model-update evaluation passes fluency tests but fails because it fabricates references. Which release gate is missing?

  1. Approve on the strength of fluent language and high response speed alone.
  2. Require groundedness and citation-validity checks on held-out tasks before deployment approval.
  3. Check only that every answer contains at least one citation-looking string.
  4. Evaluate only spelling and punctuation in the generated answers.

Explanation

Correct Answer

B:

Fluent text is not evidence-backed text. A release gate should test whether each material claim follows retrieved authorized sources.

Incorrect Answers

A:

Well-formed prose does not demonstrate that cited material supports generated factual claims.

C:

A citation's presence does not establish that its span exists or entails the claim.

D:

Correct spelling and punctuation can coexist with nonexistent or irrelevant citations. The release gate must verify that each referenced source exists and that its passage actually supports the generated claim.

Question 376

An agent is allowed to access one customer's workspace but its tools can search all tenants. What restriction must be implemented?

  1. Bind each tool request to the caller's tenant and permission context
  2. Allow every tool to run with workspace-wide credentials although this agent is scoped to one customer
  3. Stop side effects while retaining governed logs and trace evidence
  4. Enforce an iteration/time budget with no-progress detection and explicit failure

Explanation

Correct Answer

A:

Tool authorization needs to be enforced on every request independently of model instructions. This directly supports the need to constrain tools to a user's authorized tenant, unlike a mechanism primarily intended to differentiate retry safety between reads and writes.

Incorrect Answers

B:

The controller must enforce customer-level tool authorization. Broad credentials can expose other tenants even if the prompt mentions only one workspace.

C:

Safe termination should preserve information necessary to diagnose and review the violation. Here the needed outcome is to constrain tools to a user's authorized tenant; an option designed to halt unsafe actions without destroying evidence would not directly provide it.

D:

Bounded tool execution prevents runaway use even when the agent cannot self-recognize a loop. Its primary purpose is to terminate runaway autonomous actions; the stated constraint instead demands a way to constrain tools to a user's authorized tenant.

Question 377

A retrieval agent must call a model and read grounding blobs without storing keys. Which THREE measures jointly support that goal? Choose THREE.

  1. Use one shared static administrator key for model inference, blob reads and agent orchestration across all environments.
  2. Reduce the principal's permissions to model-inference roles at the smallest applicable scope
  3. Remove its identity role assignments and disable the unused workload identity
  4. Authenticate the workload with managed identity and supported Azure RBAC
  5. Assign a narrowly scoped Storage Blob Data Reader role to its workload identity
  6. Verify separate Azure AI Search data-plane authorization for that identity

Explanation

Correct Answers

B:

Broad Owner grants are unrelated to the specific model call and unnecessarily enable resource changes. Managed identity provides secretless authentication; blob data access and narrowly scoped model inference privileges authorize only the necessary operations.

D:

Managed identity issues scoped tokens without storing resource keys in the app environment.

E:

Model invocation permission is distinct from reading blobs through the storage data plane.

Incorrect Answers

A:

A broad shared key defeats keyless least privilege and creates cross-resource exposure. Use workload identity, distinct minimal data-plane roles and controlled runtime authorization for the three required capabilities.

C:

Deleting only an application configuration file does not necessarily revoke the workload's authorization. That mechanism concerns revoke retired workload access completely, whereas this case tests how to provide keyless identity plus least-privilege model and blob access.

F:

Project access does not automatically imply permissions to query another service's data. Here, however, the stated objective is to provide keyless identity plus least-privilege model and blob access; this option focuses on authorize search data access separately from project access.

Question 378

A controller catches a policy violation, stops the agent and then deletes all traces. What incident practice is better?

  1. Clear all trace IDs and delete the triggering tool records immediately after stopping the agent.
  2. Allow bounded read retries but require idempotency for writes
  3. Require a verified human approval artifact at the transfer tool boundary
  4. Stop side effects while retaining governed logs and trace evidence

Explanation

Correct Answer

D:

Safe termination should preserve information necessary to diagnose and review the violation. The decisive difference is whether the design can halt unsafe actions without destroying evidence; the competing function serves to terminate runaway autonomous actions.

Incorrect Answers

A:

Stopping unsafe activity should preserve the investigation evidence, not erase it. Retain governed, access-controlled traces while blocking further unauthorized side effects.

B:

Read-only and side-effecting operations have different replay risks and need distinct handling. Here the needed outcome is to halt unsafe actions without destroying evidence; an option designed to differentiate retry safety between reads and writes would not directly provide it.

C:

Approval must be enforced by the application handling the real transfer, not only by prompt text. Its primary purpose is to gate high-value transfers on human authorization; the stated constraint instead demands a way to halt unsafe actions without destroying evidence.

Question 379

An application in an isolated Azure VNet must resolve and reach its supported Azure AI model endpoint over a private IP rather than the service’s public hostname route. What connectivity configuration is required?

  1. Use an appropriate private endpoint and private DNS resolution for the service
  2. Disable public ingress on the model resource without configuring any private endpoint or resolving the service hostname to a private address.
  3. Renumber unrelated VNets but continue sending model requests through the public inference endpoint.
  4. Review the public endpoint firewall logs but make no Private Link endpoint or private DNS zone configuration for the model service.

Explanation

Correct Answer

A:

Private Link and correct name resolution keep service access on controlled private networking. This resolves the specific need to route AI endpoint access via private DNS and endpoint, while a measure that helps deny public ingress to a private AI service addresses a different requirement.

Incorrect Answers

B:

Disabling public access can be an important defense-in-depth step, but by itself it supplies no private IP or name resolution to the application. Private endpoint and private DNS configuration are necessary for this question.

C:

Eliminating CIDR overlaps can help connectivity planning, but it does not create private model access. Configure the supported private endpoint and DNS/routing path to the intended AI service.

D:

Firewall review explains denials but does not create the required private address path. Configure the supported private endpoint, associate correct private DNS resolution and verify network reachability from the VNet.

Question 380

A planner chooses an image tool when the user asked only for a cost calculation. What decision boundary should correct this?

  1. Invoke the image tool because a financial report contains the word picture.
  2. Give the planner permission to call every available tool regardless of task.
  3. Replace the computation with a generated chart image without checking arithmetic.
  4. Select tools from the user's actual authorized task and tool descriptions, not from superficial keyword matches.

Explanation

Correct Answer

D:

Tool use should be relevant to a justified action; unrelated modalities waste resources and can create errors.

Incorrect Answers

A:

Superficial keywords do not authorize a visual operation for a pure cost calculation.

B:

Unnecessary broad tool scope increases error, expense and potential side effects.

C:

An image cannot substitute for a correct calculation and verified numerical answer.

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