Microsoft AB-731 Practice Test 81
Question 800
At Sapphire Mobility, a bank proposes testing credit-risk fairness only on AI-generated applicants. Evidence: Synthetic samples may reproduce omissions and hide rare real-world issues. The business goal is to distinguish synthetic test utility from representative production evidence. The project owner requires a measurable acceptance criterion before signoff. Which evidence would most convincingly demonstrate that the requirement has been met?
- Users get safe failure messages and no duplicate high-impact actions
- Performance claims explicitly separate synthetic results from real validation
- Chosen approach meets quality targets at lower cost and acceptable latency
- The plan blocks unapproved actions and offers authorized alternatives
Explanation
Correct Answer
B:
Synthetic samples may reproduce omissions and hide rare real-world issues. Performance claims explicitly separate synthetic results from real validation. Synthetic examples are good for safe boundary testing, but generated cases may reproduce the omissions of the prompt. Keep a separately governed representative evaluation set for real-world coverage.
Incorrect Answers
A:
“Users get safe failure messages and no duplicate high-impact actions” validates an outcome in a case where the evidence is: A single dependent service outage stops all customer help. This question instead requires an action that can distinguish synthetic test utility from representative production evidence. Synthetic samples may reproduce omissions and hide rare real-world issues. Performance claims explicitly separate synthetic results from real validation. Synthetic examples are good for safe boundary testing, but generated cases may reproduce the omissions of the prompt. Keep a separately governed representative evaluation set for real-world coverage.
C:
“Chosen approach meets quality targets at lower cost and acceptable latency” validates an outcome in a case where the evidence is: Simple tags may be handled by a lower-cost capable model. This question instead requires an action that can distinguish synthetic test utility from representative production evidence. Synthetic samples may reproduce omissions and hide rare real-world issues. Performance claims explicitly separate synthetic results from real validation. Synthetic examples are good for safe boundary testing, but generated cases may reproduce the omissions of the prompt. Keep a separately governed representative evaluation set for real-world coverage.
D:
“The plan blocks unapproved actions and offers authorized alternatives” validates an outcome in a case where the evidence is: Some requests require separate integrations, authority and audited actions. This question instead requires an action that can distinguish synthetic test utility from representative production evidence. Synthetic samples may reproduce omissions and hide rare real-world issues. Performance claims explicitly separate synthetic results from real validation. Synthetic examples are good for safe boundary testing, but generated cases may reproduce the omissions of the prompt. Keep a separately governed representative evaluation set for real-world coverage.
Question 801
At Clearwater Partners, a sales lead wants generated customer contracts sent without inspection. Evidence: Generated text could contain binding promises beyond approval. The business goal is to insert accountable review before external actions. The proposed solution must respect documented authority and organizational controls. Which approach most directly meets the stated business objective within the approved authority?
- Use safe prompt hierarchy, sanitization and authorized tool controls
- Pilot renewal tiers based on observed task-level benefit and governance
- Use Copilot for drafting while withholding final send until review
- Define workflow instructions independent of unsupported model toggles
Explanation
Correct Answer
C:
Generated text could contain binding promises beyond approval. Use Copilot for drafting while withholding final send until review. Draft generation is not final authorization; a human approval step prevents unsupported contractual commitments from being sent.
Incorrect Answers
A:
Prompt-injection defenses mitigate malicious instructions in retrieved content, but cannot substitute for commercial or legal approval of the promises in a proposed response.
B:
“Pilot renewal tiers based on observed task-level benefit and governance” is a design or implementation response to the condition: License usage is not equivalent to measurable task improvement. This question instead requires an action that can insert accountable review before external actions. Generated text could contain binding promises beyond approval. Use Copilot for drafting while withholding final send until review. Draft generation is not final authorization; a human approval step prevents unsupported contractual commitments from being sent.
D:
“Define workflow instructions independent of unsupported model toggles” is a design or implementation response to the condition: The supported surface and administrative policies govern model availability. This question instead requires an action that can insert accountable review before external actions. Generated text could contain binding promises beyond approval. Use Copilot for drafting while withholding final send until review. Draft generation is not final authorization; a human approval step prevents unsupported contractual commitments from being sent.
Question 802
At BrightPath Media, a booking agent needs read-only calendar availability but requests mailbox write access. Evidence: An unnecessary permission increases risk without improving the task. The business goal is to grant minimum action scopes to agents. 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?
- List required read and write operations separately
- Identify source memo, requested sections and edit permissions
- Analyze failed requests and intended topic categories
- Record token volumes, retries and actual chosen billing meters
Explanation
Correct Answer
A:
An unnecessary permission increases risk without improving the task. List required read and write operations separately.
Incorrect Answers
B:
“Identify source memo, requested sections and edit permissions” tests a proposed change for the condition: The task needs a working document with source-grounded language. This question instead requires an action that can grant minimum action scopes to agents. An unnecessary permission increases risk without improving the task. List required read and write operations separately.
C:
“Analyze failed requests and intended topic categories” is a diagnostic step for the reported condition: Different user intents need separate approved answer paths. This question instead requires an action that can grant minimum action scopes to agents. An unnecessary permission increases risk without improving the task. List required read and write operations separately.
D:
“Record token volumes, retries and actual chosen billing meters” is a diagnostic step for the reported condition: Input and output tokens and deployment type all affect billing. This question instead requires an action that can grant minimum action scopes to agents. An unnecessary permission increases risk without improving the task. List required read and write operations separately.
Question 803
At Marigold Education, an accounts team wants printed invoice text digitized. Evidence: The original files are scans and require text extraction. The business goal is to use supported OCR vision processing and verify data quality. Only non-sensitive synthetic data and approved sandbox systems may be used. Which bounded trial with synthetic records best tests the proposed approach?
- Run fictional queries with deliberately distractor paragraphs
- Process mock invoices with rotated and low-contrast pages
- Replay synthetic old and new product requests
- Prototype a simulated approval step without production writes
Explanation
Correct Answer
B:
The original files are scans and require text extraction. Process mock invoices with rotated and low-contrast pages. Scanned documents must be converted into machine-readable text before reliable extraction. Compare OCR output with challenging sample pages because an error in recognition propagates to downstream answers.
Incorrect Answers
A:
“Run fictional queries with deliberately distractor paragraphs” tests a proposed change for the condition: First-stage keyword search has poor top-result precision. This question instead requires an action that can use supported OCR vision processing and verify data quality. The original files are scans and require text extraction. Process mock invoices with rotated and low-contrast pages. Scanned documents must be converted into machine-readable text before reliable extraction. Compare OCR output with challenging sample pages because an error in recognition propagates to downstream answers.
C:
“Replay synthetic old and new product requests” tests a proposed change for the condition: New request types were absent from evaluation data. This question instead requires an action that can use supported OCR vision processing and verify data quality. The original files are scans and require text extraction. Process mock invoices with rotated and low-contrast pages. Scanned documents must be converted into machine-readable text before reliable extraction. Compare OCR output with challenging sample pages because an error in recognition propagates to downstream answers.
D:
“Prototype a simulated approval step without production writes” tests a proposed change for the condition: Base Copilot chat cannot directly implement every business-specific action. This question instead requires an action that can use supported OCR vision processing and verify data quality. The original files are scans and require text extraction. Process mock invoices with rotated and low-contrast pages. Scanned documents must be converted into machine-readable text before reliable extraction. Compare OCR output with challenging sample pages because an error in recognition propagates to downstream answers.
Question 804
At Pinecrest Services, an internal assistant confuses HR benefits questions with IT support tickets. Evidence: Different user intents need separate approved answer paths. The business goal is to design agent topics and disambiguation for distinct intents. Only non-sensitive synthetic data and approved sandbox systems may be used. Which bounded trial with synthetic records best tests the proposed approach?
- Test fictional overlapping benefits and IT issue phrases
- Test with a synthetic private file visible only to one user
- Add a conflicting mock industry report to a trial
- Compare a no-retrieval baseline with an indexed policy pilot
Explanation
Correct Answer
A:
Different user intents need separate approved answer paths. Test fictional overlapping benefits and IT issue phrases.
Incorrect Answers
B:
“Test with a synthetic private file visible only to one user” tests a proposed change for the condition: The service has not been authorized to access the requested private file. This question instead requires an action that can design agent topics and disambiguation for distinct intents. Different user intents need separate approved answer paths. Test fictional overlapping benefits and IT issue phrases.
C:
“Add a conflicting mock industry report to a trial” tests a proposed change for the condition: The agent draws from a small biased sample of industry articles. This question instead requires an action that can design agent topics and disambiguation for distinct intents. Different user intents need separate approved answer paths. Test fictional overlapping benefits and IT issue phrases.
D:
“Compare a no-retrieval baseline with an indexed policy pilot” tests a proposed change for the condition: Fresh facts and style behavior are separate concerns. This question instead requires an action that can design agent topics and disambiguation for distinct intents. Different user intents need separate approved answer paths. Test fictional overlapping benefits and IT issue phrases.
Question 805
At BrightPath Media, a business analyst wants to connect meeting notes to related documents. Evidence: People, files and collaboration events have permissioned relationships. The business goal is to use available work graph context for authorized cross-app insights. 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 chosen model, quota, region and data-processing terms
- Review purpose, access permissions and sensitive relationship categories
- Determine which connected entities and files the analyst can access
- List which facts need public search versus work file access
Explanation
Correct Answer
C:
People, files and collaboration events have permissioned relationships. Determine which connected entities and files the analyst can access. Work graph relationships can connect authorized people, files and collaboration history, but inherited access boundaries still apply. Test whether the requester can view each cited item.
Incorrect Answers
A:
“Check chosen model, quota, region and data-processing terms” is a diagnostic step for the reported condition: Not every model and deployment option exists in all regions. This question instead requires an action that can use available work graph context for authorized cross-app insights. People, files and collaboration events have permissioned relationships. Determine which connected entities and files the analyst can access. Work graph relationships can connect authorized people, files and collaboration history, but inherited access boundaries still apply. Test whether the requester can view each cited item.
B:
“Review purpose, access permissions and sensitive relationship categories” is a diagnostic step for the reported condition: Graph relationships may expose sensitive work patterns and restricted files. This question instead requires an action that can use available work graph context for authorized cross-app insights. People, files and collaboration events have permissioned relationships. Determine which connected entities and files the analyst can access. Work graph relationships can connect authorized people, files and collaboration history, but inherited access boundaries still apply. Test whether the requester can view each cited item.
D:
“List which facts need public search versus work file access” is a diagnostic step for the reported condition: Public web claims and workplace records have different provenance. This question instead requires an action that can use available work graph context for authorized cross-app insights. People, files and collaboration events have permissioned relationships. Determine which connected entities and files the analyst can access. Work graph relationships can connect authorized people, files and collaboration history, but inherited access boundaries still apply. Test whether the requester can view each cited item.
Question 806
At Silverpine Group, 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 proposed solution must respect documented authority and organizational controls. Which approach most directly meets the stated business objective within the approved authority?
- Apply source and index security filtering before the LLM sees passages
- Add representative new cohorts and trigger governed re-evaluation
- Test a suitable existing model before considering custom training
- Require cited counterevidence and human assessment before recommendations
Explanation
Correct Answer
B:
New request types were absent from evaluation data. Add representative new cohorts and trigger governed re-evaluation. 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:
“Apply source and index security filtering before the LLM sees passages” is a design or implementation response to the condition: Index query results lack caller-specific authorization filters. 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. Add representative new cohorts and trigger governed re-evaluation. 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:
“Test a suitable existing model before considering custom training” is a design or implementation response to the condition: The available budget and data are far smaller than required for foundation training. 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. Add representative new cohorts and trigger governed re-evaluation. 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:
“Require cited counterevidence and human assessment before recommendations” is a design or implementation response to the condition: The agent draws from a small biased sample of industry articles. 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. Add representative new cohorts and trigger governed re-evaluation. 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 807
At Redwood Consulting, a manufacturer wants to predict rare equipment failures from sensor histories. Evidence: Threshold rules miss complex multi-variable patterns. The business goal is to use predictive ML when labeled trends add real signal. The project owner requires a measurable acceptance criterion before signoff. Which evidence would most convincingly demonstrate that the requirement has been met?
- The investment case shows net value under realistic cost assumptions
- A clean holdout provides a credible lower performance estimate
- The model identifies real anomalies without unmanageable false alarms
- The workflow functions with minimal necessary fields and no leaks
Explanation
Correct Answer
C:
Threshold rules miss complex multi-variable patterns. The model identifies real anomalies without unmanageable false alarms. When multiple signals interact, a fixed threshold may lose useful predictive information. Evaluate a labeled model against the current rule baseline using task-relevant false-positive and false-negative costs.
Incorrect Answers
A:
“The investment case shows net value under realistic cost assumptions” validates an outcome in a case where the evidence is: Integration, human review, logging and employee training are omitted. This question instead requires an action that can use predictive ML when labeled trends add real signal. Threshold rules miss complex multi-variable patterns. The model identifies real anomalies without unmanageable false alarms. When multiple signals interact, a fixed threshold may lose useful predictive information. Evaluate a labeled model against the current rule baseline using task-relevant false-positive and false-negative costs.
B:
“A clean holdout provides a credible lower performance estimate” validates an outcome in a case where the evidence is: The same customer records appear in tuning and testing data. This question instead requires an action that can use predictive ML when labeled trends add real signal. Threshold rules miss complex multi-variable patterns. The model identifies real anomalies without unmanageable false alarms. When multiple signals interact, a fixed threshold may lose useful predictive information. Evaluate a labeled model against the current rule baseline using task-relevant false-positive and false-negative costs.
D:
“The workflow functions with minimal necessary fields and no leaks” validates an outcome in a case where the evidence is: Only scheduling constraints and contact details are needed. This question instead requires an action that can use predictive ML when labeled trends add real signal. Threshold rules miss complex multi-variable patterns. The model identifies real anomalies without unmanageable false alarms. When multiple signals interact, a fixed threshold may lose useful predictive information. Evaluate a labeled model against the current rule baseline using task-relevant false-positive and false-negative costs.
Question 808
At Atlas Design, an internal chatbot occasionally produces offensive responses to provocative inputs. Evidence: Unmoderated content could harm employees or customers. The business goal is to apply proportionate output safety checks and escalation. 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?
- Identify quality, privacy, safety and latency measures separately
- Classify relevant content risks and existing moderation coverage
- Identify unnecessary health fields in each request payload
- Inspect the approved labels, class imbalance and routing criteria
Explanation
Correct Answer
B:
Unmoderated content could harm employees or customers. Classify relevant content risks and existing moderation coverage. Potentially harmful outputs need safety screening suited to the audience and use case. Test borderline cases, provide escalation routes and log policy failures without treating a classifier as infallible.
Incorrect Answers
A:
“Identify quality, privacy, safety and latency measures separately” is a diagnostic step for the reported condition: Data protection failures are not reflected in the accuracy metric. This question instead requires an action that can apply proportionate output safety checks and escalation. Unmoderated content could harm employees or customers. Classify relevant content risks and existing moderation coverage. Potentially harmful outputs need safety screening suited to the audience and use case. Test borderline cases, provide escalation routes and log policy failures without treating a classifier as infallible.
C:
“Identify unnecessary health fields in each request payload” is a diagnostic step for the reported condition: Only scheduling constraints and contact details are needed. This question instead requires an action that can apply proportionate output safety checks and escalation. Unmoderated content could harm employees or customers. Classify relevant content risks and existing moderation coverage. Potentially harmful outputs need safety screening suited to the audience and use case. Test borderline cases, provide escalation routes and log policy failures without treating a classifier as infallible.
D:
“Inspect the approved labels, class imbalance and routing criteria” is a diagnostic step for the reported condition: The desired answer is a stable label rather than a free-form narrative. This question instead requires an action that can apply proportionate output safety checks and escalation. Unmoderated content could harm employees or customers. Classify relevant content risks and existing moderation coverage. Potentially harmful outputs need safety screening suited to the audience and use case. Test borderline cases, provide escalation routes and log policy failures without treating a classifier as infallible.
Question 809
At Lakeshore Commerce, a team requests a custom research agent despite an approved existing one. Evidence: A duplicate custom project adds governance and maintenance cost. The business goal is to compare existing tools with customization requirements before building. 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 fictional make-versus-buy scoring workshop
- Test a fictional budget with two known variances
- Compare two models on the same synthetic document cases
- Try a simulated inventory update with denied and allowed mock roles
Explanation
Correct Answer
A:
A duplicate custom project adds governance and maintenance cost. Run a fictional make-versus-buy scoring workshop.
Incorrect Answers
B:
“Test a fictional budget with two known variances” tests a proposed change for the condition: The original workbook contains numbers and formulas requiring verification. This question instead requires an action that can compare existing tools with customization requirements before building. A duplicate custom project adds governance and maintenance cost. Run a fictional make-versus-buy scoring workshop.
C:
“Compare two models on the same synthetic document cases” tests a proposed change for the condition: The published scores do not measure the company's actual document tasks. This question instead requires an action that can compare existing tools with customization requirements before building. A duplicate custom project adds governance and maintenance cost. Run a fictional make-versus-buy scoring workshop.
D:
“Try a simulated inventory update with denied and allowed mock roles” tests a proposed change for the condition: Reading policy is different from invoking a high-impact connector action. This question instead requires an action that can compare existing tools with customization requirements before building. A duplicate custom project adds governance and maintenance cost. Run a fictional make-versus-buy scoring workshop.
Question 810
At Atlas Design, a transformation team receives dozens of unrelated AI requests. Evidence: No process ranks ideas by value, readiness and risk. The business goal is to govern intake and prioritization for sustainable AI investment. 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?
- Segment audiences and identify practical business impacts
- Record incident rates, review timeliness and unresolved risks
- Identify sponsor, business goals and decision rights
- Collect use-case owners, value hypotheses, datasets and risk assessments
Explanation
Correct Answer
D:
No process ranks ideas by value, readiness and risk. Collect use-case owners, value hypotheses, datasets and risk assessments.
Incorrect Answers
A:
“Segment audiences and identify practical business impacts” is a diagnostic step for the reported condition: Teams miss updates and do not understand why workflows change. This question instead requires an action that can govern intake and prioritization for sustainable AI investment. No process ranks ideas by value, readiness and risk. Collect use-case owners, value hypotheses, datasets and risk assessments.
B:
“Record incident rates, review timeliness and unresolved risks” is a diagnostic step for the reported condition: No evidence links governance to fewer incidents or better outcomes. This question instead requires an action that can govern intake and prioritization for sustainable AI investment. No process ranks ideas by value, readiness and risk. Collect use-case owners, value hypotheses, datasets and risk assessments.
C:
“Identify sponsor, business goals and decision rights” is a diagnostic step for the reported condition: Employees receive mixed signals about priority and approved use. This question instead requires an action that can govern intake and prioritization for sustainable AI investment. No process ranks ideas by value, readiness and risk. Collect use-case owners, value hypotheses, datasets and risk assessments.