Microsoft AB-731 Practice Test 4
Question 30
At Northlake Services, a released AI classifier becomes less accurate after a product expansion. Evidence: New request types were absent from evaluation data. The business goal is to monitor deployed models and refresh evaluation with new examples. The project owner requires a measurable acceptance criterion before signoff. Which evidence would most convincingly demonstrate that the requirement has been met?
- Unseen-case format gains outweigh training and hosting costs
- The model meets accepted quality on both old and new cohorts
- Alternative text conveys necessary chart meaning without inventions
- Requests route to intended topics without guessing sensitive details
Explanation
Correct Answer
B:
New request types were absent from evaluation data. The model meets accepted quality on both old and new cohorts. Previously unseen request types can degrade a model after deployment while old benchmark scores remain stable. Capture representative production failures and update the evaluation set before changes are promoted.
Incorrect Answers
A:
“Unseen-case format gains outweigh training and hosting costs” validates an outcome in a case where the evidence is: An existing prompt baseline still violates approved template constraints. This question instead requires an action that can monitor deployed models and refresh evaluation with new examples. New request types were absent from evaluation data. The model meets accepted quality on both old and new cohorts. Previously unseen request types can degrade a model after deployment while old benchmark scores remain stable. Capture representative production failures and update the evaluation set before changes are promoted.
C:
“Alternative text conveys necessary chart meaning without inventions” validates an outcome in a case where the evidence is: Some images contain diagrams whose meaning needs human review. This question instead requires an action that can monitor deployed models and refresh evaluation with new examples. New request types were absent from evaluation data. The model meets accepted quality on both old and new cohorts. Previously unseen request types can degrade a model after deployment while old benchmark scores remain stable. Capture representative production failures and update the evaluation set before changes are promoted.
D:
“Requests route to intended topics without guessing sensitive details” validates an outcome in a case where the evidence is: Different user intents need separate approved answer paths. This question instead requires an action that can monitor deployed models and refresh evaluation with new examples. New request types were absent from evaluation data. The model meets accepted quality on both old and new cohorts. Previously unseen request types can degrade a model after deployment while old benchmark scores remain stable. Capture representative production failures and update the evaluation set before changes are promoted.
Question 31
At Sapphire Mobility, an audit department queries a policy assistant that cites a superseded manual. Evidence: The knowledge index was last refreshed before the recent policy revision. The business goal is to maintain fresh retrieval sources rather than changing model weights. The sponsor needs an evidence-based diagnosis rather than a guessed fix. Which initial investigation best identifies the controlling constraint before any solution is chosen?
- Check index refresh time against the policy effective date
- Identify tasks where the answer may be incomplete or high impact
- Check data-to-index lag and deleted document handling
- Classify relevant content risks and existing moderation coverage
Explanation
Correct Answer
A:
The knowledge index was last refreshed before the recent policy revision. Check index refresh time against the policy effective date. A policy that changes after indexing will not be reflected in retrieved passages until the source is refreshed. Track source versions, ingestion success and index freshness instead of retraining the model for factual updates.
Incorrect Answers
B:
“Identify tasks where the answer may be incomplete or high impact” is a diagnostic step for the reported condition: Confidence language may conceal missing or stale evidence. This question instead requires an action that can maintain fresh retrieval sources rather than changing model weights. The knowledge index was last refreshed before the recent policy revision. Check index refresh time against the policy effective date. A policy that changes after indexing will not be reflected in retrieved passages until the source is refreshed. Track source versions, ingestion success and index freshness instead of retraining the model for factual updates.
C:
“Check data-to-index lag and deleted document handling” is a diagnostic step for the reported condition: Its vector index does not reflect new catalog documents. This question instead requires an action that can maintain fresh retrieval sources rather than changing model weights. The knowledge index was last refreshed before the recent policy revision. Check index refresh time against the policy effective date. A policy that changes after indexing will not be reflected in retrieved passages until the source is refreshed. Track source versions, ingestion success and index freshness instead of retraining the model for factual updates.
D:
“Classify relevant content risks and existing moderation coverage” is a diagnostic step for the reported condition: Unmoderated content could harm employees or customers. This question instead requires an action that can maintain fresh retrieval sources rather than changing model weights. The knowledge index was last refreshed before the recent policy revision. Check index refresh time against the policy effective date. A policy that changes after indexing will not be reflected in retrieved passages until the source is refreshed. Track source versions, ingestion success and index freshness instead of retraining the model for factual updates.
Question 32
At Ashwood Public Services, a compliance officer wants an AI assistant to issue final policy rulings. Evidence: Regulated interpretation requires designated decision-makers. The business goal is to use Copilot as evidence support rather than substitute policy authority. The project owner requires a measurable acceptance criterion before signoff. Which evidence would most convincingly demonstrate that the requirement has been met?
- Unsafe requests remain blocked without harming approved tasks
- The response traces conclusions to authorized workplace records
- The result excludes restricted relationships and identities
- The assistant informs but cannot unilaterally approve exceptions
Explanation
Correct Answer
D:
Regulated interpretation requires designated decision-makers. The assistant informs but cannot unilaterally approve exceptions. Copilot can summarize relevant evidence, but it does not inherit statutory or organizational approval authority. A designated reviewer must apply the governing policy to the specific case.
Incorrect Answers
A:
“Unsafe requests remain blocked without harming approved tasks” validates an outcome in a case where the evidence is: Performance alone does not demonstrate responsible deployment. This question instead requires an action that can use Copilot as evidence support rather than substitute policy authority. Regulated interpretation requires designated decision-makers. The assistant informs but cannot unilaterally approve exceptions. Copilot can summarize relevant evidence, but it does not inherit statutory or organizational approval authority. A designated reviewer must apply the governing policy to the specific case.
B:
“The response traces conclusions to authorized workplace records” validates an outcome in a case where the evidence is: The answer depends on authorized Microsoft 365 work context. This question instead requires an action that can use Copilot as evidence support rather than substitute policy authority. Regulated interpretation requires designated decision-makers. The assistant informs but cannot unilaterally approve exceptions. Copilot can summarize relevant evidence, but it does not inherit statutory or organizational approval authority. A designated reviewer must apply the governing policy to the specific case.
C:
“The result excludes restricted relationships and identities” validates an outcome in a case where the evidence is: Graph relationships may expose sensitive work patterns and restricted files. This question instead requires an action that can use Copilot as evidence support rather than substitute policy authority. Regulated interpretation requires designated decision-makers. The assistant informs but cannot unilaterally approve exceptions. Copilot can summarize relevant evidence, but it does not inherit statutory or organizational approval authority. A designated reviewer must apply the governing policy to the specific case.
Question 33
At Summit Manufacturing, a chief operating officer asks an assistant to create a transformation plan. Evidence: The request gives no objective, audience, boundaries or evidence. The business goal is to define goal, context, constraints and output expectations in prompts. Only non-sensitive synthetic data and approved sandbox systems may be used. Which bounded trial with synthetic records best tests the proposed approach?
- Compare a vague request with a structured version on a mock business unit
- Run a mock plan requesting a prohibited payment update
- Test retrieval of mock credentials by authorized and denied roles
- Test rare fictional boundary cases alongside sampled actual distributions
Explanation
Correct Answer
A:
The request gives no objective, audience, boundaries or evidence. Compare a vague request with a structured version on a mock business unit. Missing task boundaries invite plausible but unrequested content. Specifying the intended audience, evidence, constraints and output format gives the assistant testable acceptance conditions.
Incorrect Answers
B:
“Run a mock plan requesting a prohibited payment update” tests a proposed change for the condition: Some requests require separate integrations, authority and audited actions. This question instead requires an action that can define goal, context, constraints and output expectations in prompts. The request gives no objective, audience, boundaries or evidence. Compare a vague request with a structured version on a mock business unit. Missing task boundaries invite plausible but unrequested content. Specifying the intended audience, evidence, constraints and output format gives the assistant testable acceptance conditions.
C:
“Test retrieval of mock credentials by authorized and denied roles” tests a proposed change for the condition: The credentials enable unauthorized paid inference requests. This question instead requires an action that can define goal, context, constraints and output expectations in prompts. The request gives no objective, audience, boundaries or evidence. Compare a vague request with a structured version on a mock business unit. Missing task boundaries invite plausible but unrequested content. Specifying the intended audience, evidence, constraints and output format gives the assistant testable acceptance conditions.
D:
“Test rare fictional boundary cases alongside sampled actual distributions” tests a proposed change for the condition: Synthetic samples may reproduce omissions and hide rare real-world issues. This question instead requires an action that can define goal, context, constraints and output expectations in prompts. The request gives no objective, audience, boundaries or evidence. Compare a vague request with a structured version on a mock business unit. Missing task boundaries invite plausible but unrequested content. Specifying the intended audience, evidence, constraints and output format gives the assistant testable acceptance conditions.
Question 34
At Cedar Analytics, an international company wants a single AI adoption policy for all jurisdictions. Evidence: Availability, privacy law and employee consultation rules differ by region. The business goal is to adapt rollout controls to local obligations without lowering core safety. Only non-sensitive synthetic data and approved sandbox systems may be used. Which bounded trial with synthetic records best tests the proposed approach?
- Pilot AI access against synthetic classified and public documents
- Test two fictional regions with differing data rules
- Create a fictional system card for an FAQ assistant
- Run a fictional workshop demonstrating assisted but reviewed work
Explanation
Correct Answer
B:
Availability, privacy law and employee consultation rules differ by region. Test two fictional regions with differing data rules. A feature or data practice acceptable in one jurisdiction may require different notice, consultation or deployment controls elsewhere. Review local obligations while preserving consistent minimum safeguards.
Incorrect Answers
A:
“Pilot AI access against synthetic classified and public documents” tests a proposed change for the condition: Unreviewed permissions could expose confidential material to broader employees. This question instead requires an action that can adapt rollout controls to local obligations without lowering core safety. Availability, privacy law and employee consultation rules differ by region. Test two fictional regions with differing data rules. A feature or data practice acceptable in one jurisdiction may require different notice, consultation or deployment controls elsewhere. Review local obligations while preserving consistent minimum safeguards.
C:
“Create a fictional system card for an FAQ assistant” tests a proposed change for the condition: Users and auditors cannot see intended scope and failure conditions. This question instead requires an action that can adapt rollout controls to local obligations without lowering core safety. Availability, privacy law and employee consultation rules differ by region. Test two fictional regions with differing data rules. A feature or data practice acceptable in one jurisdiction may require different notice, consultation or deployment controls elsewhere. Review local obligations while preserving consistent minimum safeguards.
D:
“Run a fictional workshop demonstrating assisted but reviewed work” tests a proposed change for the condition: Leadership communication focuses only on headcount cost savings. This question instead requires an action that can adapt rollout controls to local obligations without lowering core safety. Availability, privacy law and employee consultation rules differ by region. Test two fictional regions with differing data rules. A feature or data practice acceptable in one jurisdiction may require different notice, consultation or deployment controls elsewhere. Review local obligations while preserving consistent minimum safeguards.
Question 35
At Willow Creek Finance, a board sponsor assumes one language model can execute all enterprise transactions. Evidence: Some requests require separate integrations, authority and audited actions. The business goal is to separate model text ability from granted system-action capability. The project owner requires a measurable acceptance criterion before signoff. Which evidence would most convincingly demonstrate that the requirement has been met?
- The returned wording matches the approved source and its citation
- No unauthorized instruction is followed and the suspicious text is flagged
- Extracted amounts and row associations match source documents
- The plan blocks unapproved actions and offers authorized alternatives
Explanation
Correct Answer
D:
Some requests require separate integrations, authority and audited actions. The plan blocks unapproved actions and offers authorized alternatives. Text generation does not grant an agent permission to read external records or execute transactions. Integrations, scoped credentials and action approvals determine what the system can actually do.
Incorrect Answers
A:
“The returned wording matches the approved source and its citation” validates an outcome in a case where the evidence is: Paraphrasing would change obligations and cannot replace the signed source. This question instead requires an action that can separate model text ability from granted system-action capability. Some requests require separate integrations, authority and audited actions. The plan blocks unapproved actions and offers authorized alternatives. Text generation does not grant an agent permission to read external records or execute transactions. Integrations, scoped credentials and action approvals determine what the system can actually do.
B:
“No unauthorized instruction is followed and the suspicious text is flagged” validates an outcome in a case where the evidence is: The page is untrusted external content rather than a user-approved command. This question instead requires an action that can separate model text ability from granted system-action capability. Some requests require separate integrations, authority and audited actions. The plan blocks unapproved actions and offers authorized alternatives. Text generation does not grant an agent permission to read external records or execute transactions. Integrations, scoped credentials and action approvals determine what the system can actually do.
C:
“Extracted amounts and row associations match source documents” validates an outcome in a case where the evidence is: Corrupt source extraction can make even a grounded answer wrong. This question instead requires an action that can separate model text ability from granted system-action capability. Some requests require separate integrations, authority and audited actions. The plan blocks unapproved actions and offers authorized alternatives. Text generation does not grant an agent permission to read external records or execute transactions. Integrations, scoped credentials and action approvals determine what the system can actually do.
Question 36
At Summit Manufacturing, a hospital scheduling workflow sends full patient histories for appointment summaries. Evidence: Only scheduling constraints and contact details are needed. The business goal is to minimize personal information in AI inputs. Only non-sensitive synthetic data and approved sandbox systems may be used. Which bounded trial with synthetic records best tests the proposed approach?
- Run a mock appointment request with extra sensitive attributes
- Compare retrieval on test questions before and after chunk changes
- Model low and high adoption scenarios using fictional finance inputs
- Test synthetic labeled shelf images under varied lighting
Explanation
Correct Answer
A:
Only scheduling constraints and contact details are needed. Run a mock appointment request with extra sensitive attributes. Collect only the personal fields necessary to arrange the appointment. Removing unrelated identifiers reduces exposure without preventing the requested scheduling operation.
Incorrect Answers
B:
“Compare retrieval on test questions before and after chunk changes” tests a proposed change for the condition: Large irrelevant text consumes context and obscures the answer. This question instead requires an action that can minimize personal information in AI inputs. Only scheduling constraints and contact details are needed. Run a mock appointment request with extra sensitive attributes. Collect only the personal fields necessary to arrange the appointment. Removing unrelated identifiers reduces exposure without preventing the requested scheduling operation.
C:
“Model low and high adoption scenarios using fictional finance inputs” tests a proposed change for the condition: Integration, human review, logging and employee training are omitted. This question instead requires an action that can minimize personal information in AI inputs. Only scheduling constraints and contact details are needed. Run a mock appointment request with extra sensitive attributes. Collect only the personal fields necessary to arrange the appointment. Removing unrelated identifiers reduces exposure without preventing the requested scheduling operation.
D:
“Test synthetic labeled shelf images under varied lighting” tests a proposed change for the condition: Visual interpretation matters more than long-form text creativity. This question instead requires an action that can minimize personal information in AI inputs. Only scheduling constraints and contact details are needed. Run a mock appointment request with extra sensitive attributes. Collect only the personal fields necessary to arrange the appointment. Removing unrelated identifiers reduces exposure without preventing the requested scheduling operation.
Question 37
At Prairie Health, a prototype team grants every collaborator administrator privileges in a shared AI project. Evidence: Excess access can lead to unreviewed deployments and data exposure. The business goal is to apply least-privilege project roles and ownership. Only non-sensitive synthetic data and approved sandbox systems may be used. Which bounded trial with synthetic records best tests the proposed approach?
- Compare old and new models on an approved regression set
- Test a fictional contract email with an unsupported guarantee
- Test mock contributor and reviewer identities against deployment actions
- Test synthetic product photos paired with conflicting notes
Explanation
Correct Answer
C:
Excess access can lead to unreviewed deployments and data exposure. Test mock contributor and reviewer identities against deployment actions. Broad project permissions can authorize unrelated writes and deployments. Assign only the roles required for the named task, bind them to accountable owners and periodically review effective access.
Incorrect Answers
A:
“Compare old and new models on an approved regression set” tests a proposed change for the condition: The successor may change behavior and cost. This question instead requires an action that can apply least-privilege project roles and ownership. Excess access can lead to unreviewed deployments and data exposure. Test mock contributor and reviewer identities against deployment actions. Broad project permissions can authorize unrelated writes and deployments. Assign only the roles required for the named task, bind them to accountable owners and periodically review effective access.
B:
“Test a fictional contract email with an unsupported guarantee” tests a proposed change for the condition: Generated text could contain binding promises beyond approval. This question instead requires an action that can apply least-privilege project roles and ownership. Excess access can lead to unreviewed deployments and data exposure. Test mock contributor and reviewer identities against deployment actions. Broad project permissions can authorize unrelated writes and deployments. Assign only the roles required for the named task, bind them to accountable owners and periodically review effective access.
D:
“Test synthetic product photos paired with conflicting notes” tests a proposed change for the condition: Text-only models cannot inspect image evidence. This question instead requires an action that can apply least-privilege project roles and ownership. Excess access can lead to unreviewed deployments and data exposure. Test mock contributor and reviewer identities against deployment actions. Broad project permissions can authorize unrelated writes and deployments. Assign only the roles required for the named task, bind them to accountable owners and periodically review effective access.
Question 38
At Northlake Services, a department wants agents published directly from personal test environments. Evidence: Unreviewed deployment could expose confidential data or unauthorized tools. The business goal is to establish managed development, test and production controls. The operating team needs an ongoing check after the initial rollout. Which post-rollout signal is most important for detecting regression or governance drift?
- Monitor privilege creep and unreviewed role changes
- Monitor unmanaged environments and unexpected publication
- Monitor stale citations and missed owner review cycles
- Monitor new anonymous links and inheritance changes
Explanation
Correct Answer
B:
Unreviewed deployment could expose confidential data or unauthorized tools. Monitor unmanaged environments and unexpected publication. A prototype without separated environments and promotion approval can expose real data or invoke live tools. Distinct test/production scopes and controlled release gates contain those changes.
Incorrect Answers
A:
“Monitor privilege creep and unreviewed role changes” monitors this signal: Excess access can lead to unreviewed deployments and data exposure. This question instead requires an action that can establish managed development, test and production controls. Unreviewed deployment could expose confidential data or unauthorized tools. Monitor unmanaged environments and unexpected publication. A prototype without separated environments and promotion approval can expose real data or invoke live tools. Distinct test/production scopes and controlled release gates contain those changes.
C:
“Monitor stale citations and missed owner review cycles” monitors this signal: The source library was not refreshed after recent rule changes. This question instead requires an action that can establish managed development, test and production controls. Unreviewed deployment could expose confidential data or unauthorized tools. Monitor unmanaged environments and unexpected publication. A prototype without separated environments and promotion approval can expose real data or invoke live tools. Distinct test/production scopes and controlled release gates contain those changes.
D:
“Monitor new anonymous links and inheritance changes” monitors this signal: Copilot follows existing file permissions, including mistaken grants. This question instead requires an action that can establish managed development, test and production controls. Unreviewed deployment could expose confidential data or unauthorized tools. Monitor unmanaged environments and unexpected publication. A prototype without separated environments and promotion approval can expose real data or invoke live tools. Distinct test/production scopes and controlled release gates contain those changes.
Question 39
At Prairie Health, a payroll unit proposes rewriting numerical reconciliation checks as chatbot reasoning. Evidence: The reconciliation is deterministic arithmetic with known control rules. The business goal is to avoid generative substitution for reliable rule-based computation. Only non-sensitive synthetic data and approved sandbox systems may be used. Which bounded trial with synthetic records best tests the proposed approach?
- Search for a fictional policy using both phrase and paraphrase
- Apply a fictional scoring rubric to three competing proposals
- Test an unmatched balance that must fail arithmetic checks
- Label fictional tickets including ambiguous mixed-urgency cases
Explanation
Correct Answer
C:
The reconciliation is deterministic arithmetic with known control rules. Test an unmatched balance that must fail arithmetic checks. Known reconciliation rules are deterministic, so they should be implemented as repeatable arithmetic with explicit exception handling. A generative response can explain results but should not substitute for controlled calculation.
Incorrect Answers
A:
“Search for a fictional policy using both phrase and paraphrase” tests a proposed change for the condition: Matching literal keywords alone fails for synonymous wording. This question instead requires an action that can avoid generative substitution for reliable rule-based computation. The reconciliation is deterministic arithmetic with known control rules. Test an unmatched balance that must fail arithmetic checks. Known reconciliation rules are deterministic, so they should be implemented as repeatable arithmetic with explicit exception handling. A generative response can explain results but should not substitute for controlled calculation.
B:
“Apply a fictional scoring rubric to three competing proposals” tests a proposed change for the condition: Ideas differ in readiness, benefits and regulatory exposure. This question instead requires an action that can avoid generative substitution for reliable rule-based computation. The reconciliation is deterministic arithmetic with known control rules. Test an unmatched balance that must fail arithmetic checks. Known reconciliation rules are deterministic, so they should be implemented as repeatable arithmetic with explicit exception handling. A generative response can explain results but should not substitute for controlled calculation.
D:
“Label fictional tickets including ambiguous mixed-urgency cases” tests a proposed change for the condition: The desired answer is a stable label rather than a free-form narrative. This question instead requires an action that can avoid generative substitution for reliable rule-based computation. The reconciliation is deterministic arithmetic with known control rules. Test an unmatched balance that must fail arithmetic checks. Known reconciliation rules are deterministic, so they should be implemented as repeatable arithmetic with explicit exception handling. A generative response can explain results but should not substitute for controlled calculation.
Question 40
At Monarch Energy, a finance team needs the variance between actual and budget categories. Evidence: The original workbook contains numbers and formulas requiring verification. The business goal is to use workbook-aware analysis and check calculations. The sponsor needs an evidence-based diagnosis rather than a guessed fix. The numbers are drawn from one workbook with no SQL join step. Which initial inspection best verifies the selected worksheet ranges, reporting periods, and missing input rows before calculating actual-versus-budget variance?
- Inspect selected table ranges, periods and missing rows
- Confirm recurrence, source authority, owner and license
- Assess image permission, lighting and expected visual categories
- Inspect join logic, missing records and reporting periods
Explanation
Correct Answer
A:
The original workbook contains numbers and formulas requiring verification. Inspect selected table ranges, periods and missing rows. Workbook formulas, references and numeric precision must survive any analysis. Inspect actual cell ranges and recompute representative totals before treating a generated summary as evidence.
Incorrect Answers
B:
“Confirm recurrence, source authority, owner and license” is a diagnostic step for the reported condition: An ordinary chat response does not itself establish future recurring delivery. This question instead requires an action that can use workbook-aware analysis and check calculations. The original workbook contains numbers and formulas requiring verification. Inspect selected table ranges, periods and missing rows. Workbook formulas, references and numeric precision must survive any analysis. Inspect actual cell ranges and recompute representative totals before treating a generated summary as evidence.
C:
“Assess image permission, lighting and expected visual categories” is a diagnostic step for the reported condition: Text-only signals cannot describe the damage pattern. This question instead requires an action that can use workbook-aware analysis and check calculations. The original workbook contains numbers and formulas requiring verification. Inspect selected table ranges, periods and missing rows. Workbook formulas, references and numeric precision must survive any analysis. Inspect actual cell ranges and recompute representative totals before treating a generated summary as evidence.
D:
Inspecting join logic would be appropriate for a reconciliation that combines relational datasets. This problem explicitly uses one workbook and has no join step; selected ranges, period alignment, and omitted rows are the direct first checks for that calculation.