Practice Exams:

Microsoft AI-103 Dumps

EXAM: Developing AI Apps and Agents on Azure
Verified By Experts
Premium Bundle
$49.99

AI-103 Premium ETE File


  • Premium File 187 Questions & Answers
Whats Included:
  • Latest Questions
  • 100% Accurate Answers
  • Fast Exam Updates

Microsoft AI-103 Practice Test 1

Question 1

At a regional hospital, the AI platform engineer is preparing a staged service rollout. The concrete requirement is: An internal classification endpoint has strict CPU budget and bounded label space. Which design action BEST addresses the stated need?

  1. Assume the largest available LLM is always the cheapest production choice
  2. Compare representative label accuracy and p95 response time against the baseline
  3. Use scoped conversation memory keyed to an authorized session or tenant
  4. Benchmark a compact task-capable model against accuracy and latency targets before deployment

Explanation

Correct Answer

D:

The proposed implementation is to benchmark a compact task-capable model against accuracy and latency targets before deployment; this matches the operational requirement that an internal classification endpoint has strict CPU budget and bounded label space. The team can compare representative label accuracy and p95 response time against the baseline. The team should also enforce accuracy and latency acceptance thresholds in model selection.

Incorrect Answers

A:

Choosing to assume the largest available LLM is always the cheapest production choice cannot directly replace the specified action to benchmark a compact task-capable model against accuracy and latency targets before deployment when an internal classification endpoint has strict CPU budget and bounded label space. The actual implementation must benchmark a compact task-capable model against accuracy and latency targets before deployment, because the stated service condition is that an internal classification endpoint has strict CPU budget and bounded label space.

B:

The action to compare representative label accuracy and p95 response time against the baseline can contribute to small-model latency tradeoff as a targeted acceptance check step; here, an internal classification endpoint has strict CPU budget and bounded label space. Choosing to compare representative label accuracy and p95 response time against the baseline cannot directly replace the specified action to benchmark a compact task-capable model against accuracy and latency targets before deployment when an internal classification endpoint has strict CPU budget and bounded label space.

C:

Choosing to use scoped conversation memory keyed to an authorized session or tenant would direct effort toward agent memory scope, whose characteristic issue is that an assistant must recall preferences within a permitted client session but isolate different clients. The relevant corrective or preventive action here remains to benchmark a compact task-capable model against accuracy and latency targets before deployment, with its own validation criteria.

Question 2

After a release acceptance review at a insurance platform, the team observes this failure: An internal classification endpoint has strict CPU budget and bounded label space. Which flawed assumption or implementation choice would MOST DIRECTLY explain why the intended outcome was not achieved?

  1. Route to an untested model without checking tool or modality support
  2. Benchmark a compact task-capable model against accuracy and latency targets before deployment
  3. Compare representative label accuracy and p95 response time against the baseline
  4. Assume the largest available LLM is always the cheapest production choice

Explanation

Correct Answer

D:

This design mistake is to assume the largest available LLM is always the cheapest production choice; it conflicts with the actual small-model latency tradeoff behavior needed when an internal classification endpoint has strict CPU budget and bounded label space. The insurance platform should instead benchmark a compact task-capable model against accuracy and latency targets before deployment, followed by a check to compare representative label accuracy and p95 response time against the baseline.

Incorrect Answers

A:

This insurance platform failure-analysis alternative of route to an untested model without checking tool or modality support does not diagnose the particular small-model latency tradeoff defect where an internal classification endpoint has strict CPU budget and bounded label space; the keyed mistake is assume the largest available LLM is always the cheapest production choice. The alternative may be useful in its own context, but it does not directly supply assume the largest available LLM is always the cheapest production choice for the stated decision.

B:

The alternative to benchmark a compact task-capable model against accuracy and latency targets before deployment is a implementation step rather than the faulty assumption being diagnosed. This insurance platform failure-analysis alternative of benchmark a compact task-capable model against accuracy and latency targets before deployment does not diagnose the particular small-model latency tradeoff defect where an internal classification endpoint has strict CPU budget and bounded label space; the keyed mistake is assume the largest available LLM is always the cheapest production choice.

C:

The alternative to compare representative label accuracy and p95 response time against the baseline is a targeted acceptance check step rather than the faulty assumption being diagnosed. This insurance platform failure-analysis alternative of compare representative label accuracy and p95 response time against the baseline does not diagnose the particular small-model latency tradeoff defect where an internal classification endpoint has strict CPU budget and bounded label space; the keyed mistake is assume the largest available LLM is always the cheapest production choice.

Question 3

The media platform is responding to a design meeting. A product support lead discovers this issue: An internal classification endpoint has strict CPU budget and bounded label space. What corrective step should the team take to restore the intended functionality most directly?

  1. Reselect an available deployment after checking the regional catalog and quota
  2. Assume the largest available LLM is always the cheapest production choice
  3. Compare representative label accuracy and p95 response time against the baseline
  4. Route classification requests to a measured smaller model and retest difficult classes

Explanation

Correct Answer

D:

The direct repair is to route classification requests to a measured smaller model and retest difficult classes. Once the condition is corrected, the media platform should compare representative label accuracy and p95 response time against the baseline. To reduce recurrence, it should also enforce accuracy and latency acceptance thresholds in model selection; those steps provide different recovery and prevention functions.

Incorrect Answers

A:

During media platform remediation, the proposed step to reselect an available deployment after checking the regional catalog and quota is not the direct restoration action for small-model latency tradeoff when an internal classification endpoint has strict CPU budget and bounded label space; the keyed repair is route classification requests to a measured smaller model and retest difficult classes. If the team chose it instead, the targeted check to compare representative label accuracy and p95 response time against the baseline could still fail because the underlying required action is to route classification requests to a measured smaller model and retest difficult classes.

B:

During media platform remediation, the proposed step to assume the largest available LLM is always the cheapest production choice is not the direct restoration action for small-model latency tradeoff when an internal classification endpoint has strict CPU budget and bounded label space; the keyed repair is route classification requests to a measured smaller model and retest difficult classes. After remediation the team must compare representative label accuracy and p95 response time against the baseline; the action in this option does not itself establish that target recovery outcome.

C:

The action to compare representative label accuracy and p95 response time against the baseline can contribute to small-model latency tradeoff as a targeted acceptance check step; here, an internal classification endpoint has strict CPU budget and bounded label space. During media platform remediation, the proposed step to compare representative label accuracy and p95 response time against the baseline is not the direct restoration action for small-model latency tradeoff when an internal classification endpoint has strict CPU budget and bounded label space; the keyed repair is route classification requests to a measured smaller model and retest difficult classes.

Question 4

A data engineer at a customer support center wants the following issue to be prevented in later deployments: An internal classification endpoint has strict CPU budget and bounded label space. Which enduring control or engineering guardrail BEST prevents recurrence?

  1. Schedule source freshness checks and remove stale indexed copies
  2. Enforce accuracy and latency acceptance thresholds in model selection
  3. Assume the largest available LLM is always the cheapest production choice
  4. Route classification requests to a measured smaller model and retest difficult classes

Explanation

Correct Answer

B:

Standardizing the boundary to enforce accuracy and latency acceptance thresholds in model selection is preventive because the same small-model latency tradeoff failure could otherwise return: an internal classification endpoint has strict CPU budget and bounded label space. For this customer support center, the service should also benchmark a compact task-capable model against accuracy and latency targets before deployment, and tests should compare representative label accuracy and p95 response time against the baseline so that the ongoing requirement is measurable.

Incorrect Answers

A:

The action to schedule source freshness checks and remove stale indexed copies addresses managed search versus flat prompts rather than the decisive small-model latency tradeoff condition. The customer support center should not infer that fixing one adjacent AI service validates its unrelated workflow; the requested action is to enforce accuracy and latency acceptance thresholds in model selection.

C:

For prevention at the customer support center, the measure to assume the largest available LLM is always the cheapest production choice is weaker or aimed at another issue than the small-model latency tradeoff guardrail to enforce accuracy and latency acceptance thresholds in model selection needed when an internal classification endpoint has strict CPU budget and bounded label space. The lasting guardrail for this specific threat is to enforce accuracy and latency acceptance thresholds in model selection, not simply to complete an adjacent service task.

D:

The action to route classification requests to a measured smaller model and retest difficult classes can contribute to small-model latency tradeoff as a corrective remediation step; here, an internal classification endpoint has strict CPU budget and bounded label space. For prevention at the customer support center, the measure to route classification requests to a measured smaller model and retest difficult classes is weaker or aimed at another issue than the small-model latency tradeoff guardrail to enforce accuracy and latency acceptance thresholds in model selection needed when an internal classification endpoint has strict CPU budget and bounded label space.

Question 5

During a release acceptance review, a service administrator must accept a change for small-model latency tradeoff. The issue to be addressed is: An internal classification endpoint has strict CPU budget and bounded label space. Which concrete verification action demonstrates that the relevant behavior works?

  1. Assume the largest available LLM is always the cheapest production choice
  2. Update a source paragraph and confirm retrieval reflects the new revision
  3. Benchmark a compact task-capable model against accuracy and latency targets before deployment
  4. Compare representative label accuracy and p95 response time against the baseline

Explanation

Correct Answer

D:

The targeted acceptance test is to compare representative label accuracy and p95 response time against the baseline. The design should first benchmark a compact task-capable model against accuracy and latency targets before deployment. A robust manufacturing group deployment also needs to enforce accuracy and latency acceptance thresholds in model selection; passing this check provides relevant evidence but does not certify every possible input or security risk.

Incorrect Answers

A:

In the manufacturing group acceptance review, trying to assume the largest available LLM is always the cheapest production choice does not test the exact small-model latency tradeoff condition where an internal classification endpoint has strict CPU budget and bounded label space; the relevant test is to compare representative label accuracy and p95 response time against the baseline. Acceptance should instead demonstrate that the service can compare representative label accuracy and p95 response time against the baseline, using the described failure as the test criterion.

B:

For managed search versus flat prompts, it can be sensible to update a source paragraph and confirm retrieval reflects the new revision if a service assistant needs frequently changing knowledge across thousands of files. Without the actual change to compare representative label accuracy and p95 response time against the baseline, the answer lacks the correct operational scope even if the alternate action is technically valid elsewhere.

C:

The action to benchmark a compact task-capable model against accuracy and latency targets before deployment can contribute to small-model latency tradeoff as a implementation step; here, an internal classification endpoint has strict CPU budget and bounded label space. In the manufacturing group acceptance review, trying to benchmark a compact task-capable model against accuracy and latency targets before deployment does not test the exact small-model latency tradeoff condition where an internal classification endpoint has strict CPU budget and bounded label space; the relevant test is to compare representative label accuracy and p95 response time against the baseline.

Question 6

At a travel support provider, the cloud developer is preparing a quality regression test. The concrete requirement is: An intake app must reason jointly about photographs and textual descriptions. Which design action BEST addresses the stated need?

  1. Submit paired images and descriptions and check image-grounded responses
  2. Select a multimodal-capable model that accepts both image and text inputs
  3. Treat a text-only model as though it can inspect image pixels
  4. Choose a supported video-understanding pipeline that retains temporal context

Explanation

Correct Answer

B:

Choosing to select a multimodal-capable model that accepts both image and text inputs directly targets the described multimodal support selection gap: an intake app must reason jointly about photographs and textual descriptions. As a lasting boundary, they also need to validate supported modalities and input limits before rollout, because an implementation by itself does not prove safe operation.

Incorrect Answers

A:

The action to submit paired images and descriptions and check image-grounded responses can contribute to multimodal support selection as a targeted acceptance check step; here, an intake app must reason jointly about photographs and textual descriptions. Choosing to submit paired images and descriptions and check image-grounded responses cannot directly replace the specified action to select a multimodal-capable model that accepts both image and text inputs when an intake app must reason jointly about photographs and textual descriptions.

C:

Choosing to treat a text-only model as though it can inspect image pixels cannot directly replace the specified action to select a multimodal-capable model that accepts both image and text inputs when an intake app must reason jointly about photographs and textual descriptions. The actual implementation must select a multimodal-capable model that accepts both image and text inputs, because the stated service condition is that an intake app must reason jointly about photographs and textual descriptions.

D:

Choosing to choose a supported video-understanding pipeline that retains temporal context would direct effort toward video context selection, whose characteristic issue is that a safety application must summarize events occurring across video frames. The relevant corrective or preventive action here remains to select a multimodal-capable model that accepts both image and text inputs, with its own validation criteria.

Question 7

After a production incident at a scientific publisher, the team observes this failure: An intake app must reason jointly about photographs and textual descriptions. Which flawed assumption or implementation choice would MOST DIRECTLY explain why the intended outcome was not achieved?

  1. Select a multimodal-capable model that accepts both image and text inputs
  2. Treat a text-only model as though it can inspect image pixels
  3. Assume a generic chat completion guarantees stable invoice line-item schemas
  4. Submit paired images and descriptions and check image-grounded responses

Explanation

Correct Answer

B:

The faulty assumption is treat a text-only model as though it can inspect image pixels. A functioning alternative would be to select a multimodal-capable model that accepts both image and text inputs, and the relevant validation would be to submit paired images and descriptions and check image-grounded responses.

Incorrect Answers

A:

The alternative to select a multimodal-capable model that accepts both image and text inputs is a implementation step rather than the faulty assumption being diagnosed. This scientific publisher failure-analysis alternative of select a multimodal-capable model that accepts both image and text inputs does not diagnose the particular multimodal support selection defect where an intake app must reason jointly about photographs and textual descriptions; the keyed mistake is treat a text-only model as though it can inspect image pixels.

C:

The option to assume a generic chat completion guarantees stable invoice line-item schemas would normally serve a different service choice for structured receipts requirement, where an accounts service needs invoice fields with coordinates and confidence evidence. The alternative may be useful in its own context, but it does not directly supply treat a text-only model as though it can inspect image pixels for the stated decision.

D:

The alternative to submit paired images and descriptions and check image-grounded responses is a targeted acceptance check step rather than the faulty assumption being diagnosed. This scientific publisher failure-analysis alternative of submit paired images and descriptions and check image-grounded responses does not diagnose the particular multimodal support selection defect where an intake app must reason jointly about photographs and textual descriptions; the keyed mistake is treat a text-only model as though it can inspect image pixels.

Question 8

The nonprofit organization is responding to a data protection inspection. A security engineer discovers this issue: An intake app must reason jointly about photographs and textual descriptions. What corrective step should the team take to restore the intended functionality most directly?

  1. Treat a text-only model as though it can inspect image pixels
  2. Switch the deployment to a supported multimodal model and verify image handling
  3. Ingest current documents into the index and reconnect the application retriever
  4. Submit paired images and descriptions and check image-grounded responses

Explanation

Correct Answer

B:

The appropriate remediation is to switch the deployment to a supported multimodal model and verify image handling, because the failure concerns multimodal support selection: an intake app must reason jointly about photographs and textual descriptions. A meaningful verification should submit paired images and descriptions and check image-grounded responses, and the durable operating rule should validate supported modalities and input limits before rollout after the immediate fix.

Incorrect Answers

A:

During nonprofit organization remediation, the proposed step to treat a text-only model as though it can inspect image pixels is not the direct restoration action for multimodal support selection when an intake app must reason jointly about photographs and textual descriptions; the keyed repair is switch the deployment to a supported multimodal model and verify image handling. After remediation the team must submit paired images and descriptions and check image-grounded responses; the action in this option does not itself establish that target recovery outcome.

C:

This approach relates to managed search versus flat prompts: a service assistant needs frequently changing knowledge across thousands of files. If the team chose it instead, the targeted check to submit paired images and descriptions and check image-grounded responses could still fail because the underlying required action is to switch the deployment to a supported multimodal model and verify image handling.

D:

The action to submit paired images and descriptions and check image-grounded responses can contribute to multimodal support selection as a targeted acceptance check step; here, an intake app must reason jointly about photographs and textual descriptions. During nonprofit organization remediation, the proposed step to submit paired images and descriptions and check image-grounded responses is not the direct restoration action for multimodal support selection when an intake app must reason jointly about photographs and textual descriptions; the keyed repair is switch the deployment to a supported multimodal model and verify image handling.

Question 9

A agent designer at a municipal services team wants the following issue to be prevented in later deployments: An intake app must reason jointly about photographs and textual descriptions. Which enduring control or engineering guardrail BEST prevents recurrence?

  1. Restrict tool permissions and validate arguments before every call
  2. Validate supported modalities and input limits before rollout
  3. Switch the deployment to a supported multimodal model and verify image handling
  4. Treat a text-only model as though it can inspect image pixels

Explanation

Correct Answer

B:

The sustained prevention measure is to validate supported modalities and input limits before rollout. The implementation can still use the separate action to select a multimodal-capable model that accepts both image and text inputs, but a guardrail constrains future changes. The municipal services team can measure whether the rule works by arranging to submit paired images and descriptions and check image-grounded responses.

Incorrect Answers

A:

The action to restrict tool permissions and validate arguments before every call addresses tool calling versus freeform text rather than the decisive multimodal support selection condition. The municipal services team should not infer that fixing one adjacent AI service validates its unrelated workflow; the requested action is to validate supported modalities and input limits before rollout.

C:

The action to switch the deployment to a supported multimodal model and verify image handling can contribute to multimodal support selection as a corrective remediation step; here, an intake app must reason jointly about photographs and textual descriptions. For prevention at the municipal services team, the measure to switch the deployment to a supported multimodal model and verify image handling is weaker or aimed at another issue than the multimodal support selection guardrail to validate supported modalities and input limits before rollout needed when an intake app must reason jointly about photographs and textual descriptions.

D:

For prevention at the municipal services team, the measure to treat a text-only model as though it can inspect image pixels is weaker or aimed at another issue than the multimodal support selection guardrail to validate supported modalities and input limits before rollout needed when an intake app must reason jointly about photographs and textual descriptions. The lasting guardrail for this specific threat is to validate supported modalities and input limits before rollout, not simply to complete an adjacent service task.

Question 10

During a production incident, a responsible-AI reviewer must accept a change for multimodal support selection. The issue to be addressed is: An intake app must reason jointly about photographs and textual descriptions. Which concrete verification action demonstrates that the relevant behavior works?

  1. Simulate two tenant sessions and verify no preference crosses the boundary
  2. Submit paired images and descriptions and check image-grounded responses
  3. Select a multimodal-capable model that accepts both image and text inputs
  4. Treat a text-only model as though it can inspect image pixels

Explanation

Correct Answer

B:

To verify this change, the payment services team should submit paired images and descriptions and check image-grounded responses. Operators still need the implementation to select a multimodal-capable model that accepts both image and text inputs and the preventive requirement to validate supported modalities and input limits before rollout.

Incorrect Answers

A:

For agent memory scope, it can be sensible to simulate two tenant sessions and verify no preference crosses the boundary if an assistant must recall preferences within a permitted client session but isolate different clients. Without the actual change to submit paired images and descriptions and check image-grounded responses, the answer lacks the correct operational scope even if the alternate action is technically valid elsewhere.

C:

The action to select a multimodal-capable model that accepts both image and text inputs can contribute to multimodal support selection as a implementation step; here, an intake app must reason jointly about photographs and textual descriptions. In the payment services team acceptance review, trying to select a multimodal-capable model that accepts both image and text inputs does not test the exact multimodal support selection condition where an intake app must reason jointly about photographs and textual descriptions; the relevant test is to submit paired images and descriptions and check image-grounded responses.

D:

In the payment services team acceptance review, trying to treat a text-only model as though it can inspect image pixels does not test the exact multimodal support selection condition where an intake app must reason jointly about photographs and textual descriptions; the relevant test is to submit paired images and descriptions and check image-grounded responses. Acceptance should instead demonstrate that the service can submit paired images and descriptions and check image-grounded responses, using the described failure as the test criterion.

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