Microsoft AI-103 Dumps
Microsoft AI-103 Practice Test 23
Question 220
During a source update, a responsible-AI reviewer must accept a change for access token authorization. The issue to be addressed is: An inference app succeeds locally using developer credentials but fails under deployment. Which concrete verification action demonstrates that the relevant behavior works?
Explanation
Correct Answer
To verify this change, the travel support provider should invoke the app under its actual service principal rather than a local user. Operators still need the implementation to grant the workload identity the required role and validate token audience and the preventive requirement to use automated role permission tests for runtime identities.
Incorrect Answers
The action to use automated role permission tests for runtime identities can contribute to access token authorization as a ongoing preventive rule step; here, an inference app succeeds locally using developer credentials but fails under deployment. In the travel support provider acceptance review, trying to use automated role permission tests for runtime identities does not test the exact access token authorization condition where an inference app succeeds locally using developer credentials but fails under deployment; the relevant test is to invoke the app under its actual service principal rather than a local user.
In the travel support provider acceptance review, trying to assume personal developer sign-in roles transfer to deployed managed identities does not test the exact access token authorization condition where an inference app succeeds locally using developer credentials but fails under deployment; the relevant test is to invoke the app under its actual service principal rather than a local user. Acceptance should instead demonstrate that the service can invoke the app under its actual service principal rather than a local user, using the described failure as the test criterion.
Choosing to ask an agent to retrieve forbidden document IDs and confirm access is blocked would direct effort toward data-connector permissions, whose characteristic issue is that an agent can read documents from projects it was never assigned. The relevant corrective or preventive action here remains to invoke the app under its actual service principal rather than a local user, with its own validation criteria.
The action to grant the workload identity the required role and validate token audience can contribute to access token authorization as a implementation step; here, an inference app succeeds locally using developer credentials but fails under deployment. In the travel support provider acceptance review, trying to grant the workload identity the required role and validate token audience does not test the exact access token authorization condition where an inference app succeeds locally using developer credentials but fails under deployment; the relevant test is to invoke the app under its actual service principal rather than a local user.
Question 221
At a regional hospital, the solution architect is preparing a security tabletop exercise. The concrete requirement is: A team cannot explain which prompt version produced a degraded answer. Which design action BEST addresses the stated need?
Explanation
Correct Answer
The proposed implementation is to version prompt templates model IDs datasets and evaluation results in each release; this matches the operational requirement that a team cannot explain which prompt version produced a degraded answer. The team can compare traced prompt and model versions for an observed regression. The team should also tag every trace with deployment and prompt template identifiers.
Incorrect Answers
This approach relates to model output drift: an assistant becomes less accurate after a deployment or prompt change. If the team chose it instead, the targeted check to compare traced prompt and model versions for an observed regression could still fail because the underlying required action is to version prompt templates model IDs datasets and evaluation results in each release.
The action to compare traced prompt and model versions for an observed regression can contribute to experiment reproducibility as a targeted acceptance check step; here, a team cannot explain which prompt version produced a degraded answer. Choosing to compare traced prompt and model versions for an observed regression cannot directly replace the specified action to version prompt templates model IDs datasets and evaluation results in each release when a team cannot explain which prompt version produced a degraded answer.
The action to tag every trace with deployment and prompt template identifiers can contribute to experiment reproducibility as a ongoing preventive rule step; here, a team cannot explain which prompt version produced a degraded answer. Choosing to tag every trace with deployment and prompt template identifiers cannot directly replace the specified action to version prompt templates model IDs datasets and evaluation results in each release when a team cannot explain which prompt version produced a degraded answer.
Choosing to rely on a screenshot with no versioned deployment metadata cannot directly replace the specified action to version prompt templates model IDs datasets and evaluation results in each release when a team cannot explain which prompt version produced a degraded answer. The actual implementation must version prompt templates model IDs datasets and evaluation results in each release, because the stated service condition is that a team cannot explain which prompt version produced a degraded answer.
The action to reconstruct a tested release from saved configuration and rerun evaluations can contribute to experiment reproducibility as a corrective remediation step; here, a team cannot explain which prompt version produced a degraded answer. Choosing to reconstruct a tested release from saved configuration and rerun evaluations cannot directly replace the specified action to version prompt templates model IDs datasets and evaluation results in each release when a team cannot explain which prompt version produced a degraded answer.
Question 222
After a support escalation at a insurance platform, the team observes this failure: A team cannot explain which prompt version produced a degraded answer. Which flawed assumption or implementation choice would MOST DIRECTLY explain why the intended outcome was not achieved?
Explanation
Correct Answer
This design mistake is to rely on a screenshot with no versioned deployment metadata; it conflicts with the actual experiment reproducibility behavior needed when a team cannot explain which prompt version produced a degraded answer. The insurance platform should instead version prompt templates model IDs datasets and evaluation results in each release, followed by a check to compare traced prompt and model versions for an observed regression.
Incorrect Answers
The alternative to compare traced prompt and model versions for an observed regression is a targeted acceptance check step rather than the faulty assumption being diagnosed. This insurance platform failure-analysis alternative of compare traced prompt and model versions for an observed regression does not diagnose the particular experiment reproducibility defect where a team cannot explain which prompt version produced a degraded answer; the keyed mistake is rely on a screenshot with no versioned deployment metadata.
This insurance platform failure-analysis alternative of assume application authentication automatically restricts administrator permissions does not diagnose the particular experiment reproducibility defect where a team cannot explain which prompt version produced a degraded answer; the keyed mistake is rely on a screenshot with no versioned deployment metadata. The alternative may be useful in its own context, but it does not directly supply rely on a screenshot with no versioned deployment metadata for the stated decision.
The alternative to version prompt templates model IDs datasets and evaluation results in each release is a implementation step rather than the faulty assumption being diagnosed. This insurance platform failure-analysis alternative of version prompt templates model IDs datasets and evaluation results in each release does not diagnose the particular experiment reproducibility defect where a team cannot explain which prompt version produced a degraded answer; the keyed mistake is rely on a screenshot with no versioned deployment metadata.
Question 223
The media platform is responding to a service audit. A AI application owner discovers this issue: A team cannot explain which prompt version produced a degraded answer. What corrective step should the team take to restore the intended functionality most directly?
Explanation
Correct Answer
The direct repair is to reconstruct a tested release from saved configuration and rerun evaluations. Once the condition is corrected, the media platform should compare traced prompt and model versions for an observed regression. To reduce recurrence, it should also tag every trace with deployment and prompt template identifiers; those steps provide different recovery and prevention functions.
Incorrect Answers
The action to compare traced prompt and model versions for an observed regression can contribute to experiment reproducibility as a targeted acceptance check step; here, a team cannot explain which prompt version produced a degraded answer. During media platform remediation, the proposed step to compare traced prompt and model versions for an observed regression is not the direct restoration action for experiment reproducibility when a team cannot explain which prompt version produced a degraded answer; the keyed repair is reconstruct a tested release from saved configuration and rerun evaluations.
During media platform remediation, the proposed step to remove overbroad role grants and retest required developer workflows is not the direct restoration action for experiment reproducibility when a team cannot explain which prompt version produced a degraded answer; the keyed repair is reconstruct a tested release from saved configuration and rerun evaluations. If the team chose it instead, the targeted check to compare traced prompt and model versions for an observed regression could still fail because the underlying required action is to reconstruct a tested release from saved configuration and rerun evaluations.
During media platform remediation, the proposed step to rely on a screenshot with no versioned deployment metadata is not the direct restoration action for experiment reproducibility when a team cannot explain which prompt version produced a degraded answer; the keyed repair is reconstruct a tested release from saved configuration and rerun evaluations. After remediation the team must compare traced prompt and model versions for an observed regression; the action in this option does not itself establish that target recovery outcome.
Question 224
A quality assurance lead at a customer support center wants the following issue to be prevented in later deployments: A team cannot explain which prompt version produced a degraded answer. Which enduring control or engineering guardrail BEST prevents recurrence?
Explanation
Correct Answer
Standardizing the boundary to tag every trace with deployment and prompt template identifiers is preventive because the same experiment reproducibility failure could otherwise return: a team cannot explain which prompt version produced a degraded answer. For this customer support center, the service should also version prompt templates model IDs datasets and evaluation results in each release, and tests should compare traced prompt and model versions for an observed regression so that the ongoing requirement is measurable.
Incorrect Answers
For prevention at the customer support center, the measure to rely on a screenshot with no versioned deployment metadata is weaker or aimed at another issue than the experiment reproducibility guardrail to tag every trace with deployment and prompt template identifiers needed when a team cannot explain which prompt version produced a degraded answer. The lasting guardrail for this specific threat is to tag every trace with deployment and prompt template identifiers, not simply to complete an adjacent service task.
The action to reconstruct a tested release from saved configuration and rerun evaluations can contribute to experiment reproducibility as a corrective remediation step; here, a team cannot explain which prompt version produced a degraded answer. For prevention at the customer support center, the measure to reconstruct a tested release from saved configuration and rerun evaluations is weaker or aimed at another issue than the experiment reproducibility guardrail to tag every trace with deployment and prompt template identifiers needed when a team cannot explain which prompt version produced a degraded answer.
The action to test cross-project denial and log knowledge-source permission changes addresses data-connector permissions rather than the decisive experiment reproducibility condition. The customer support center should not infer that fixing one adjacent AI service validates its unrelated workflow; the requested action is to tag every trace with deployment and prompt template identifiers.
Question 225
During a support escalation, a AI platform engineer must accept a change for experiment reproducibility. The issue to be addressed is: A team cannot explain which prompt version produced a degraded answer. Which concrete verification action demonstrates that the relevant behavior works?
Explanation
Correct Answer
The targeted acceptance test is to compare traced prompt and model versions for an observed regression. The design should first version prompt templates model IDs datasets and evaluation results in each release. A robust manufacturing group deployment also needs to tag every trace with deployment and prompt template identifiers; passing this check provides relevant evidence but does not certify every possible input or security risk.
Incorrect Answers
For model request quota, it can be sensible to replay the peak request mix and verify stable throughput under actual quota if a product launch drives higher token-per-minute demand than the configured quota. Without the actual change to compare traced prompt and model versions for an observed regression, the answer lacks the correct operational scope even if the alternate action is technically valid elsewhere.
The action to version prompt templates model IDs datasets and evaluation results in each release can contribute to experiment reproducibility as a implementation step; here, a team cannot explain which prompt version produced a degraded answer. In the manufacturing group acceptance review, trying to version prompt templates model IDs datasets and evaluation results in each release does not test the exact experiment reproducibility condition where a team cannot explain which prompt version produced a degraded answer; the relevant test is to compare traced prompt and model versions for an observed regression.
In the manufacturing group acceptance review, trying to rely on a screenshot with no versioned deployment metadata does not test the exact experiment reproducibility condition where a team cannot explain which prompt version produced a degraded answer; the relevant test is to compare traced prompt and model versions for an observed regression. Acceptance should instead demonstrate that the service can compare traced prompt and model versions for an observed regression, using the described failure as the test criterion.
Question 226
At a travel support provider, the compliance specialist is preparing a staged service rollout. The concrete requirement is: A public assistant occasionally produces disallowed harmful instructions. Which design action BEST addresses the stated need?
Explanation
Correct Answer
Choosing to configure content filters and evaluate representative unsafe request cases directly targets the described unsafe content output gap: a public assistant occasionally produces disallowed harmful instructions. As a lasting boundary, they also need to require safety evaluation before and after deployment changes, because an implementation by itself does not prove safe operation.
Incorrect Answers
Choosing to use source-grounded evaluation to flag unsupported numeric assertions would direct effort toward hallucination evaluation, whose characteristic issue is that a financial assistant invents currency figures absent from source statements. The relevant corrective or preventive action here remains to configure content filters and evaluate representative unsafe request cases, with its own validation criteria.
The action to submit a categorized safety test set and verify handling by category can contribute to unsafe content output as a targeted acceptance check step; here, a public assistant occasionally produces disallowed harmful instructions. Choosing to submit a categorized safety test set and verify handling by category cannot directly replace the specified action to configure content filters and evaluate representative unsafe request cases when a public assistant occasionally produces disallowed harmful instructions.
Choosing to assume a polite system prompt alone reliably prevents all prohibited outputs cannot directly replace the specified action to configure content filters and evaluate representative unsafe request cases when a public assistant occasionally produces disallowed harmful instructions. The actual implementation must configure content filters and evaluate representative unsafe request cases, because the stated service condition is that a public assistant occasionally produces disallowed harmful instructions.
Question 227
After a release acceptance review at a scientific publisher, the team observes this failure: A public assistant occasionally produces disallowed harmful instructions. Which flawed assumption or implementation choice would MOST DIRECTLY explain why the intended outcome was not achieved?
Explanation
Correct Answer
The faulty assumption is assume a polite system prompt alone reliably prevents all prohibited outputs. A functioning alternative would be to configure content filters and evaluate representative unsafe request cases, and the relevant validation would be to submit a categorized safety test set and verify handling by category.
Incorrect Answers
This scientific publisher failure-analysis alternative of disable all safety filtering in production to eliminate benign rejections does not diagnose the particular unsafe content output defect where a public assistant occasionally produces disallowed harmful instructions; the keyed mistake is assume a polite system prompt alone reliably prevents all prohibited outputs. The alternative may be useful in its own context, but it does not directly supply assume a polite system prompt alone reliably prevents all prohibited outputs for the stated decision.
The alternative to submit a categorized safety test set and verify handling by category is a targeted acceptance check step rather than the faulty assumption being diagnosed. This scientific publisher failure-analysis alternative of submit a categorized safety test set and verify handling by category does not diagnose the particular unsafe content output defect where a public assistant occasionally produces disallowed harmful instructions; the keyed mistake is assume a polite system prompt alone reliably prevents all prohibited outputs.
The alternative to configure content filters and evaluate representative unsafe request cases is a implementation step rather than the faulty assumption being diagnosed. This scientific publisher failure-analysis alternative of configure content filters and evaluate representative unsafe request cases does not diagnose the particular unsafe content output defect where a public assistant occasionally produces disallowed harmful instructions; the keyed mistake is assume a polite system prompt alone reliably prevents all prohibited outputs.
Question 228
The nonprofit organization is responding to a design meeting. A data engineer discovers this issue: A public assistant occasionally produces disallowed harmful instructions. What corrective step should the team take to restore the intended functionality most directly?
Explanation
Correct Answer
The appropriate remediation is to strengthen the relevant filter thresholds and test blocked outputs, because the failure concerns unsafe content output: a public assistant occasionally produces disallowed harmful instructions. A meaningful verification should submit a categorized safety test set and verify handling by category, and the durable operating rule should require safety evaluation before and after deployment changes after the immediate fix.
Incorrect Answers
The action to submit a categorized safety test set and verify handling by category can contribute to unsafe content output as a targeted acceptance check step; here, a public assistant occasionally produces disallowed harmful instructions. During nonprofit organization remediation, the proposed step to submit a categorized safety test set and verify handling by category is not the direct restoration action for unsafe content output when a public assistant occasionally produces disallowed harmful instructions; the keyed repair is strengthen the relevant filter thresholds and test blocked outputs.
During nonprofit organization remediation, the proposed step to assume a polite system prompt alone reliably prevents all prohibited outputs is not the direct restoration action for unsafe content output when a public assistant occasionally produces disallowed harmful instructions; the keyed repair is strengthen the relevant filter thresholds and test blocked outputs. After remediation the team must submit a categorized safety test set and verify handling by category; the action in this option does not itself establish that target recovery outcome.
This approach relates to data retention for agents: a conversation store preserves customer chats past their approved retention period. If the team chose it instead, the targeted check to submit a categorized safety test set and verify handling by category could still fail because the underlying required action is to strengthen the relevant filter thresholds and test blocked outputs.
Question 229
A service administrator at a municipal services team wants the following issue to be prevented in later deployments: A public assistant occasionally produces disallowed harmful instructions. Which enduring control or engineering guardrail BEST prevents recurrence?
Explanation
Correct Answer
The sustained prevention measure is to require safety evaluation before and after deployment changes. The implementation can still use the separate action to configure content filters and evaluate representative unsafe request cases, but a guardrail constrains future changes. The municipal services team can measure whether the rule works by arranging to submit a categorized safety test set and verify handling by category.
Incorrect Answers
For prevention at the municipal services team, the measure to assume a polite system prompt alone reliably prevents all prohibited outputs is weaker or aimed at another issue than the unsafe content output guardrail to require safety evaluation before and after deployment changes needed when a public assistant occasionally produces disallowed harmful instructions. The lasting guardrail for this specific threat is to require safety evaluation before and after deployment changes, not simply to complete an adjacent service task.
The action to strengthen the relevant filter thresholds and test blocked outputs can contribute to unsafe content output as a corrective remediation step; here, a public assistant occasionally produces disallowed harmful instructions. For prevention at the municipal services team, the measure to strengthen the relevant filter thresholds and test blocked outputs is weaker or aimed at another issue than the unsafe content output guardrail to require safety evaluation before and after deployment changes needed when a public assistant occasionally produces disallowed harmful instructions.
The action to require auditable input versions scores and appropriate uncertainty notes addresses model decision transparency rather than the decisive unsafe content output condition. The municipal services team should not infer that fixing one adjacent AI service validates its unrelated workflow; the requested action is to require safety evaluation before and after deployment changes.
Question 230
During a release acceptance review, a cloud developer must accept a change for unsafe content output. The issue to be addressed is: A public assistant occasionally produces disallowed harmful instructions. Which concrete verification action demonstrates that the relevant behavior works?
Explanation
Correct Answer
To verify this change, the payment services team should submit a categorized safety test set and verify handling by category. Operators still need the implementation to configure content filters and evaluate representative unsafe request cases and the preventive requirement to require safety evaluation before and after deployment changes.
Incorrect Answers
The action to configure content filters and evaluate representative unsafe request cases can contribute to unsafe content output as a implementation step; here, a public assistant occasionally produces disallowed harmful instructions. In the payment services team acceptance review, trying to configure content filters and evaluate representative unsafe request cases does not test the exact unsafe content output condition where a public assistant occasionally produces disallowed harmful instructions; the relevant test is to submit a categorized safety test set and verify handling by category.
For source attribution controls, it can be sensible to change the handbook and test that old citations cannot silently support new claims if an agent quotes internal handbook statements without showing which version they came from. Without the actual change to submit a categorized safety test set and verify handling by category, the answer lacks the correct operational scope even if the alternate action is technically valid elsewhere.
In the payment services team acceptance review, trying to assume a polite system prompt alone reliably prevents all prohibited outputs does not test the exact unsafe content output condition where a public assistant occasionally produces disallowed harmful instructions; the relevant test is to submit a categorized safety test set and verify handling by category. Acceptance should instead demonstrate that the service can submit a categorized safety test set and verify handling by category, using the described failure as the test criterion.
Developing AI Apps and Agents on Azure (AI-103) Practice Tests
- Microsoft AI-103 Practice Test 21 - 10 Questions (Last Reviewed 01-Nov-2026)
- Microsoft AI-103 Practice Test 22 - 10 Questions (Last Reviewed 31-Oct-2026)
- Microsoft AI-103 Practice Test 23 - 10 Questions (Last Reviewed 30-Oct-2026)
- Microsoft AI-103 Practice Test 24 - 10 Questions (Last Reviewed 29-Oct-2026)
- Microsoft AI-103 Practice Test 25 - 10 Questions (Last Reviewed 28-Oct-2026)