Microsoft AI-103 Practice Test 39
Question 380
A planner chooses an image tool when the user asked only for a cost calculation. What decision boundary should correct this?
- Invoke the image tool because a financial report contains the word picture.
- Give the planner permission to call every available tool regardless of task.
- Replace the computation with a generated chart image without checking arithmetic.
- 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.
Question 381
A corporate network uses the same private address range as the planned AI application VNet, making peering and routing ambiguous. What should the network architect resolve first?
- Inspect service firewall rules, endpoint association and effective network policy
- Peer the networks with overlapping private CIDRs and assume identical address ranges provide unambiguous routes.
- Plan nonoverlapping private address ranges and unambiguous routes before connecting networks
- Use private network connectivity for both search and storage dependencies
Explanation
Correct Answer
C:Overlapping address spaces make routing unreliable; choosing distinct CIDR ranges is a design prerequisite before adding private service connections. The practical requirement is to avoid ambiguous routing caused by overlapping virtual network CIDRs; a solution built to diagnose network admission on AI resources would not meet that same need.
Incorrect Answers
A:Service network ACLs and private endpoint scopes control allowed clients independently of model quota. That is relevant to diagnose network admission on AI resources, but the present scenario requires an ability to avoid ambiguous routing caused by overlapping virtual network CIDRs.
B:Overlapping address space prevents correct route selection in supported VNet connectivity patterns. Plan nonoverlapping CIDRs before peering.
D:A privately reachable model alone does not protect grounding data sent through public retrieval dependencies. That capability is appropriate for protect the RAG data path, not only the model, not this separate task of finding how to avoid ambiguous routing caused by overlapping virtual network CIDRs.
Question 382
An autonomous assistant must avoid repeated runaway calls, prevent unauthorized data scope and preserve investigation evidence when halted. Which THREE controls belong in the controller? Choose THREE.
- Enforce an iteration/time budget with no-progress detection and explicit failure
- Stop side effects while retaining governed logs and trace evidence
- Allow the agent to retry every read operation forever, provided that database writes are individually idempotent.
- Transition from draft to send only after recorded authorized approval
- Bind each tool request to the caller's tenant and permission context
- Require explicit confirmed user intent before privileged destructive actions
Explanation
Correct Answers
A:Bounded tool execution prevents runaway use even when the agent cannot self-recognize a loop. Execution budgets, tenant-bound authorization and retained traces address three separate autonomy risks.
B:Safe termination should preserve information necessary to diagnose and review the violation.
E:Tool authorization needs to be enforced on every request independently of model instructions.
Incorrect Answers
C:Write idempotency cannot stop an unbounded read loop or excessive tool consumption. Enforce total iterations, elapsed time and no-progress detection in the controller.
D:A code-enforced approval transition prevents premature outbound communication. Here, however, the stated objective is to bound calls, authorize tools and preserve trace evidence; this option focuses on separate drafting from authorized sending.
F:Irreversible effects should not rely on assumptions about an underspecified user request. This addresses confirm intent before destructive operations. The scenario instead requires a way to bound calls, authorize tools and preserve trace evidence.
Question 383
An organization wants to prevent accidental public access to its managed AI resource after private connectivity is working. What setting should be enforced?
- Use an appropriate private endpoint and private DNS resolution for the service
- Leave the AI resource public endpoint open because a private endpoint was also provisioned.
- Validate VNet connectivity, routes, DNS and network security rules end to end
- Disable public network access when the resource and service support it
Explanation
Correct Answer
D:Turning off public ingress closes the unwanted alternative path around private connectivity. Here the relevant capability must let the team deny public ingress to a private AI service. An alternative targeting route AI endpoint access via private DNS and endpoint solves another problem.
Incorrect Answers
A:Private Link and correct name resolution keep service access on controlled private networking. This helps when a team needs to route AI endpoint access via private DNS and endpoint; it does not establish how to deny public ingress to a private AI service.
B:A private endpoint does not necessarily disable public network access automatically. Explicitly restrict or disable the public path when supported.
C:Private endpoint reachability depends on both network paths and resolving the correct private address. That capability is appropriate for troubleshoot cross-network private connectivity, not this separate task of finding how to deny public ingress to a private AI service.
Question 384
A user gets an incorrect answer after a tool timeout but the final response looks normal. Which trace fields reveal the cause?
- Keep only aggregate token totals and omit the timed-out tool's request ID and failure status from logs.
- Record each branch's start, end, dependency waits and tool durations
- Inspect function-choice traces with tool descriptions and schema versions
- Capture the original tool request, its failure status and correlated model continuation
Explanation
Correct Answer
D:A complete trajectory shows where tool errors turned into unsupported model output. This resolves the specific need to trace tool failure through model continuation, while a measure that helps identify expensive agent steps addresses a different requirement.
Incorrect Answers
A:Token usage does not reveal the critical authorization or timeout failure. Correlated tool-call status and the subsequent model continuation are required to explain the misleading normal-looking answer.
B:Span timelines reveal the true critical path and wasted sequencing. That is relevant to profile multi-agent critical-path latency, but the present scenario requires an ability to trace tool failure through model continuation.
C:Tool-selection observability reveals ambiguous descriptions and version drift. This helps when a team needs to diagnose wrong function selection; it does not establish how to trace tool failure through model continuation.
Question 385
A deployed agent must call an external enterprise API reachable only through approved egress paths. What architecture should be established?
- Permit outbound HTTPS to every external address rather than enforcing the approved enterprise API destination and egress rules.
- Disable public network access when the resource and service support it
- Use an appropriate private endpoint and private DNS resolution for the service
- Route tool calls through governed egress controls and approved network connectivity
Explanation
Correct Answer
D:Tool access can require separate routing and firewall allowances beyond model serving connectivity. This directly supports the need to control outbound traffic to enterprise tool APIs, unlike a mechanism primarily intended to diagnose network admission on AI resources.
Incorrect Answers
A:Unrestricted egress bypasses the stated connectivity controls and can expose data to unapproved services. Route calls through a governed network path with destination policy enforcement.
B:Turning off public ingress closes the unwanted alternative path around private connectivity. That capability is appropriate for deny public ingress to a private AI service, not this separate task of finding how to control outbound traffic to enterprise tool APIs.
C:Private Link and correct name resolution keep service access on controlled private networking. Here the needed outcome is to control outbound traffic to enterprise tool APIs; an option designed to route AI endpoint access via private DNS and endpoint would not directly provide it.
Question 386
A web assistant returns partially generated markdown but the web client needs incremental tokens. Which integration mode supports progressive display?
- Consume a supported streaming response while handling termination and error states.
- Wait for the entire response body to finish before updating the client.
- Increase the model's maximum context window without changing transport mode.
- Split the final completed Markdown string into characters after generation finishes.
Explanation
Correct Answer
A:Streaming improves perceived responsiveness but requires clients to account for unfinished responses and tool calls.
Incorrect Answers
B:A nonstreamed response cannot provide the requested progressive token display.
C:Context capacity does not cause partial tokens to reach the browser incrementally.
D:Simulated post-completion chunks do not reduce the delay until the model's full response arrives.
Question 387
Developers in an allowed subnet are rejected by a model endpoint while other traffic succeeds. What should they investigate?
- Validate VNet connectivity, routes, DNS and network security rules end to end
- Use an appropriate private endpoint and private DNS resolution for the service
- Inspect service firewall rules, endpoint association and effective network policy
- Increase the model context window while leaving subnet allowlists and private endpoint routing unexamined.
Explanation
Correct Answer
C:Service network ACLs and private endpoint scopes control allowed clients independently of model quota. The decisive difference is whether the design can diagnose network admission on AI resources; the competing function serves to troubleshoot cross-network private connectivity.
Incorrect Answers
A:Private endpoint reachability depends on both network paths and resolving the correct private address. That capability is appropriate for troubleshoot cross-network private connectivity, not this separate task of finding how to diagnose network admission on AI resources.
B:Private Link and correct name resolution keep service access on controlled private networking. Here the needed outcome is to diagnose network admission on AI resources; an option designed to route AI endpoint access via private DNS and endpoint would not directly provide it.
D:Prompt size cannot repair a network authorization denial. Check effective firewall rules, endpoint associations and the permitted path.
Question 388
An agent uses far more tokens than expected on one path. What profiling helps?
- Break down prompt and completion tokens by model invocation and agent step
- Profile only how frequently the model selected each tool name, ignoring input and output token counts by stage.
- Correlate retrieved document IDs and ranking scores with agent output quality
- Measure unauthorized tool-attempt and policy-denial events alongside task success
Explanation
Correct Answer
A:Step-level token accounting identifies costly loops and unnecessary context. The practical requirement is to identify expensive agent steps; a solution built to diagnose wrong function selection would not meet that same need.
Incorrect Answers
B:Tool frequencies alone do not explain token cost spikes. Break down prompt, completion and retry token consumption per model invocation and agent step.
C:Coupled retrieval and answer telemetry isolates the evidence-selection stage. The function serves to connect search relevance problems to agent responses, whereas the actual requirement is to identify expensive agent steps.
D:Outcome success can hide risky intermediate attempts that require security review. That capability is appropriate for monitor agent policy violations independently of success, not this separate task of finding how to identify expensive agent steps.
Question 389
Customer prompts may include personal information, and the team wants evaluators to see only what is necessary. Which data handling step belongs in evaluation?
- Store every raw prompt and identifying field permanently for reproducibility.
- Give all evaluators direct access to production customer transcripts.
- Redact or pseudonymize personal fields in evaluation traces and restrict retention and reviewer access.
- Hash only record IDs while leaving personal names and addresses in traces.
Explanation
Correct Answer
C:Evaluation observability must preserve enough context for quality checks while minimizing exposure of regulated identifiers.
Incorrect Answers
A:Permanent retention of every raw user prompt and identifying field exceeds the stated minimization need. Redact or pseudonymize sensitive fields, restrict evaluator access and apply a bounded retention policy.
B:Broad access to full personal data is unnecessary for most evaluation questions.
D:Partial obfuscation does not minimize the many other identifiable fields evaluators can see.
Question 390
A production agent has a good final success rate but sometimes violates approval constraints. Which monitor is needed?
- Report the percentage of tasks completed and hide approval-policy denials so success looks higher.
- Keep redacted trace spans with stable identifiers and secured evidence references
- Capture the original tool request, its failure status and correlated model continuation
- Measure unauthorized tool-attempt and policy-denial events alongside task success
Explanation
Correct Answer
D:Outcome success can hide risky intermediate attempts that require security review. Here the relevant capability must let the team monitor agent policy violations independently of success. An alternative targeting identify expensive agent steps solves another problem.
Incorrect Answers
A:Task success is not proof of compliant execution. Instrument unauthorized tool attempts, policy denials and the corresponding oversight outcomes alongside completion metrics.
B:Structured privacy-controlled telemetry balances reproducible investigation with data minimization. The function serves to investigate using privacy-governed traces, whereas the actual requirement is to monitor agent policy violations independently of success.
C:A complete trajectory shows where tool errors turned into unsupported model output. That capability is appropriate for trace tool failure through model continuation, not this separate task of finding how to monitor agent policy violations independently of success.