Microsoft AB-731 Practice Test 77
Question 760
At Monarch Energy, a model seems perfect on its benchmark and fails in production. Evidence: The same customer records appear in tuning and testing data. The business goal is to prevent training-test contamination. 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 benchmark with intentionally overlapping cases
- Compare smaller and larger models on stratified fictional examples
- Replace a fictional policy version and rerun the same question
- Simulate an endpoint timeout during a fictional support request
Explanation
Correct Answer
A:
The same customer records appear in tuning and testing data. Run a mock benchmark with intentionally overlapping cases. If the same customers appear in training and test sets, evaluation can reward memorization. Split by the appropriate entity and time boundary, then verify that leaked records do not cross partitions.
Incorrect Answers
B:
“Compare smaller and larger models on stratified fictional examples” tests a proposed change for the condition: Simple tags may be handled by a lower-cost capable model. This question instead requires an action that can prevent training-test contamination. The same customer records appear in tuning and testing data. Run a mock benchmark with intentionally overlapping cases. If the same customers appear in training and test sets, evaluation can reward memorization. Split by the appropriate entity and time boundary, then verify that leaked records do not cross partitions.
C:
“Replace a fictional policy version and rerun the same question” tests a proposed change for the condition: The knowledge index was last refreshed before the recent policy revision. This question instead requires an action that can prevent training-test contamination. The same customer records appear in tuning and testing data. Run a mock benchmark with intentionally overlapping cases. If the same customers appear in training and test sets, evaluation can reward memorization. Split by the appropriate entity and time boundary, then verify that leaked records do not cross partitions.
D:
“Simulate an endpoint timeout during a fictional support request” tests a proposed change for the condition: A single dependent service outage stops all customer help. This question instead requires an action that can prevent training-test contamination. The same customer records appear in tuning and testing data. Run a mock benchmark with intentionally overlapping cases. If the same customers appear in training and test sets, evaluation can reward memorization. Split by the appropriate entity and time boundary, then verify that leaked records do not cross partitions.
Question 761
At Cedar Analytics, a sponsor compares two models using only their per-million-token prices. Evidence: Indexing, embeddings, storage, evaluation and idle hosting also cost money. The business goal is to evaluate end-to-end workload economics rather than unit rates. The project owner requires a measurable acceptance criterion before signoff. Which evidence would most convincingly demonstrate that the requirement has been met?
- The preferred option wins on measured total value and constraints
- Only synced permitted tickets support the answer
- The chosen option meets the defined internal metric gates
- The study demonstrates capability without exposing real students
Explanation
Correct Answer
A:
Indexing, embeddings, storage, evaluation and idle hosting also cost money. The preferred option wins on measured total value and constraints. Model token charges are only part of a retrieval workload. Include embedding, indexing, storage, idle deployment and monitoring costs at expected volume before comparing options.
Incorrect Answers
B:
“Only synced permitted tickets support the answer” validates an outcome in a case where the evidence is: The data is not necessarily available to Microsoft 365 Copilot without configuration. This question instead requires an action that can evaluate end-to-end workload economics rather than unit rates. Indexing, embeddings, storage, evaluation and idle hosting also cost money. The preferred option wins on measured total value and constraints. Model token charges are only part of a retrieval workload. Include embedding, indexing, storage, idle deployment and monitoring costs at expected volume before comparing options.
C:
“The chosen option meets the defined internal metric gates” validates an outcome in a case where the evidence is: The published scores do not measure the company's actual document tasks. This question instead requires an action that can evaluate end-to-end workload economics rather than unit rates. Indexing, embeddings, storage, evaluation and idle hosting also cost money. The preferred option wins on measured total value and constraints. Model token charges are only part of a retrieval workload. Include embedding, indexing, storage, idle deployment and monitoring costs at expected volume before comparing options.
D:
“The study demonstrates capability without exposing real students” validates an outcome in a case where the evidence is: A less sensitive evaluation could answer the technical question. This question instead requires an action that can evaluate end-to-end workload economics rather than unit rates. Indexing, embeddings, storage, evaluation and idle hosting also cost money. The preferred option wins on measured total value and constraints. Model token charges are only part of a retrieval workload. Include embedding, indexing, storage, idle deployment and monitoring costs at expected volume before comparing options.
Question 762
At Cascade Utilities, a proposed knowledge assistant sends an entire library to the model each time. Evidence: Large retrieved context increases token spending and response time. The business goal is to budget retrieval and generation costs together. The project owner requires a measurable acceptance criterion before signoff. Which evidence would most convincingly demonstrate that the requirement has been met?
- Reported improvement holds across both successful and escalated cases
- Only required services access credentials and audit logs record use
- Inter-annotator agreement and minority-class performance improve
- Cost and latency improve while answer grounding passes quality tests
Explanation
Correct Answer
D:
Large retrieved context increases token spending and response time. Cost and latency improve while answer grounding passes quality tests.
Incorrect Answers
A:
“Reported improvement holds across both successful and escalated cases” validates an outcome in a case where the evidence is: Feedback is self-selected and omits escalated failures. This question instead requires an action that can budget retrieval and generation costs together. Large retrieved context increases token spending and response time. Cost and latency improve while answer grounding passes quality tests.
B:
“Only required services access credentials and audit logs record use” validates an outcome in a case where the evidence is: The credentials enable unauthorized paid inference requests. This question instead requires an action that can budget retrieval and generation costs together. Large retrieved context increases token spending and response time. Cost and latency improve while answer grounding passes quality tests.
C:
“Inter-annotator agreement and minority-class performance improve” validates an outcome in a case where the evidence is: Training labels are inconsistent and rare classes have few examples. This question instead requires an action that can budget retrieval and generation costs together. Large retrieved context increases token spending and response time. Cost and latency improve while answer grounding passes quality tests.
Question 763
At Elmwood Banking, an executive declares an AI pilot successful because weekly users doubled. Evidence: No customer quality, time savings or cost data were collected. The business goal is to define success through business outcomes rather than activity. 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?
- Map jurisdictions, permitted features and consultation requirements
- Identify how users report issues and how severe they are
- Record baseline task cycle time, error and user experience
- Check visual provenance, license and any implied factual claims
Explanation
Correct Answer
C:
No customer quality, time savings or cost data were collected. Record baseline task cycle time, error and user experience.
Incorrect Answers
A:
“Map jurisdictions, permitted features and consultation requirements” is a diagnostic step for the reported condition: Availability, privacy law and employee consultation rules differ by region. This question instead requires an action that can define success through business outcomes rather than activity. No customer quality, time savings or cost data were collected. Record baseline task cycle time, error and user experience.
B:
“Identify how users report issues and how severe they are” is a diagnostic step for the reported condition: Issues are visible only to individual users and are not triaged. This question instead requires an action that can define success through business outcomes rather than activity. No customer quality, time savings or cost data were collected. Record baseline task cycle time, error and user experience.
D:
“Check visual provenance, license and any implied factual claims” is a diagnostic step for the reported condition: The image may falsely imply a real situation occurred. This question instead requires an action that can define success through business outcomes rather than activity. No customer quality, time savings or cost data were collected. Record baseline task cycle time, error and user experience.
Question 764
At Crownridge Travel, a customer service analyst needs data from an external ticket system. Evidence: The data is not necessarily available to Microsoft 365 Copilot without configuration. The business goal is to verify connector availability, authorization and freshness. 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 stale citations and missed owner review cycles
- Monitor connector outages, leakage and stale records
- Monitor broken links and unstated workbook revisions
- Monitor throttling, retry storms and peak cost
Explanation
Correct Answer
B:
The data is not necessarily available to Microsoft 365 Copilot without configuration. Monitor connector outages, leakage and stale records.
Incorrect Answers
A:
“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 verify connector availability, authorization and freshness. The data is not necessarily available to Microsoft 365 Copilot without configuration. Monitor connector outages, leakage and stale records.
C:
“Monitor broken links and unstated workbook revisions” monitors this signal: Some underlying work records were revised after the first draft. This question instead requires an action that can verify connector availability, authorization and freshness. The data is not necessarily available to Microsoft 365 Copilot without configuration. Monitor connector outages, leakage and stale records.
D:
“Monitor throttling, retry storms and peak cost” monitors this signal: Quota limits and token throughput may prevent desired traffic. This question instead requires an action that can verify connector availability, authorization and freshness. The data is not necessarily available to Microsoft 365 Copilot without configuration. Monitor connector outages, leakage and stale records.
Question 765
At Sapphire Mobility, a pilot reports rising user complaints but lacks request identifiers. Evidence: There is no trace from answer to model version and retrieved sources. The business goal is to make AI outcomes diagnosable without exposing secrets. Only non-sensitive synthetic data and approved sandbox systems may be used. Which bounded trial with synthetic records best tests the proposed approach?
- Reconstruct a fictional incorrect answer from recorded test metadata
- Replace a fictional policy version and rerun the same question
- Compare synthetic descriptions with a separately calculated sales forecast
- Simulate a trial model release with a failed safety test
Explanation
Correct Answer
A:
There is no trace from answer to model version and retrieved sources. Reconstruct a fictional incorrect answer from recorded test metadata.
Incorrect Answers
B:
“Replace a fictional policy version and rerun the same question” tests a proposed change for the condition: The knowledge index was last refreshed before the recent policy revision. This question instead requires an action that can make AI outcomes diagnosable without exposing secrets. There is no trace from answer to model version and retrieved sources. Reconstruct a fictional incorrect answer from recorded test metadata.
C:
“Compare synthetic descriptions with a separately calculated sales forecast” tests a proposed change for the condition: The deliverable is new natural-language copy, not a numeric sales prediction. This question instead requires an action that can make AI outcomes diagnosable without exposing secrets. There is no trace from answer to model version and retrieved sources. Reconstruct a fictional incorrect answer from recorded test metadata.
D:
“Simulate a trial model release with a failed safety test” tests a proposed change for the condition: There is no documented validation, deployment or monitoring plan. This question instead requires an action that can make AI outcomes diagnosable without exposing secrets. There is no trace from answer to model version and retrieved sources. Reconstruct a fictional incorrect answer from recorded test metadata.
Question 766
At Juniper Operations, an operations team needs consistent extraction of invoice reference numbers. Evidence: Creative variation adds no value and can introduce mismatches. The business goal is to prefer constrained extraction and validation for deterministic tasks. 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 false visual assertions and camera-domain drift
- Monitor systemic retries and identical widespread message errors
- Monitor schema failures and silent fallback to unstructured text
- Track inconsistent extractions and formatting failures
Explanation
Correct Answer
D:
Creative variation adds no value and can introduce mismatches. Track inconsistent extractions and formatting failures.
Incorrect Answers
A:
“Monitor false visual assertions and camera-domain drift” monitors this signal: Text descriptions alone omit visual indicators needed for triage. This question instead requires an action that can prefer constrained extraction and validation for deterministic tasks. Creative variation adds no value and can introduce mismatches. Track inconsistent extractions and formatting failures.
B:
“Monitor systemic retries and identical widespread message errors” monitors this signal: One incorrect dispatch date could be repeated to thousands of customers. This question instead requires an action that can prefer constrained extraction and validation for deterministic tasks. Creative variation adds no value and can introduce mismatches. Track inconsistent extractions and formatting failures.
C:
“Monitor schema failures and silent fallback to unstructured text” monitors this signal: An inconsistent structure causes automation failures. This question instead requires an action that can prefer constrained extraction and validation for deterministic tasks. Creative variation adds no value and can introduce mismatches. Track inconsistent extractions and formatting failures.
Question 767
At Prairie Health, a benefits team wants AI to approve employee exceptions automatically. Evidence: Decisions have legal consequences and policy ambiguity. The business goal is to limit automation to low-risk repeatable tasks and preserve human approvals. The project owner requires a measurable acceptance criterion before signoff. Which evidence would most convincingly demonstrate that the requirement has been met?
- A clean holdout provides a credible lower performance estimate
- The result reports uncertainty and does not imply unsupported accuracy
- Only routine work resolves automatically and exceptions reach decision owners
- Outputs match the requested structure without copying example facts
Explanation
Correct Answer
C:
Decisions have legal consequences and policy ambiguity. Only routine work resolves automatically and exceptions reach decision owners. Routine policy retrieval can be automated, but discretionary legal or benefits exceptions require an authorized human decision.
Incorrect Answers
A:
“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 limit automation to low-risk repeatable tasks and preserve human approvals. Decisions have legal consequences and policy ambiguity. Only routine work resolves automatically and exceptions reach decision owners. Routine policy retrieval can be automated, but discretionary legal or benefits exceptions require an authorized human decision.
B:
“The result reports uncertainty and does not imply unsupported accuracy” validates an outcome in a case where the evidence is: The model lacks an approved statistical forecast and uncertainty bounds. This question instead requires an action that can limit automation to low-risk repeatable tasks and preserve human approvals. Decisions have legal consequences and policy ambiguity. Only routine work resolves automatically and exceptions reach decision owners. Routine policy retrieval can be automated, but discretionary legal or benefits exceptions require an authorized human decision.
D:
“Outputs match the requested structure without copying example facts” validates an outcome in a case where the evidence is: Written instructions alone yield inconsistent section order. This question instead requires an action that can limit automation to low-risk repeatable tasks and preserve human approvals. Decisions have legal consequences and policy ambiguity. Only routine work resolves automatically and exceptions reach decision owners. Routine policy retrieval can be automated, but discretionary legal or benefits exceptions require an authorized human decision.
Question 768
At Atlas Design, a team deploys a service that declines customer requests without reason. Evidence: Customers need a comprehensible explanation and correction mechanism. The business goal is to provide appropriate transparency for automated recommendations. Only non-sensitive synthetic data and approved sandbox systems may be used. Which bounded trial with synthetic records best tests the proposed approach?
- Test a fictional data inventory for restricted reuse licenses
- Test a fictional rejection with a correctable data error
- Run a tabletop exercise on a fictional incorrect-policy incident
- Test two fictional regions with differing data rules
Explanation
Correct Answer
B:
Customers need a comprehensible explanation and correction mechanism. Test a fictional rejection with a correctable data error.
Incorrect Answers
A:
“Test a fictional data inventory for restricted reuse licenses” tests a proposed change for the condition: Rights to read documents do not automatically permit model training. This question instead requires an action that can provide appropriate transparency for automated recommendations. Customers need a comprehensible explanation and correction mechanism. Test a fictional rejection with a correctable data error.
C:
“Run a tabletop exercise on a fictional incorrect-policy incident” tests a proposed change for the condition: There is no owner or process for AI-specific incidents. This question instead requires an action that can provide appropriate transparency for automated recommendations. Customers need a comprehensible explanation and correction mechanism. Test a fictional rejection with a correctable data error.
D:
“Test two fictional regions with differing data rules” tests a proposed change for the condition: Availability, privacy law and employee consultation rules differ by region. This question instead requires an action that can provide appropriate transparency for automated recommendations. Customers need a comprehensible explanation and correction mechanism. Test a fictional rejection with a correctable data error.
Question 769
At Clearwater Partners, a retailer asks whether a generative chatbot should predict inventory demand. Evidence: Historical seasonality and promotions drive numerical forecasts. The business goal is to prefer evaluated forecasting models for quantitative predictions. The project owner requires a measurable acceptance criterion before signoff. Which evidence would most convincingly demonstrate that the requirement has been met?
- Forecast error beats a defined naive baseline across key items
- The returned wording matches the approved source and its citation
- Only required services access credentials and audit logs record use
- No unauthorized action follows the retrieved directive
Explanation
Correct Answer
A:
Historical seasonality and promotions drive numerical forecasts. Forecast error beats a defined naive baseline across key items. Numerical forecasts require measured error against historical demand and known seasonal drivers. A fluent language-model narrative alone does not establish predictive accuracy or calibrated uncertainty.
Incorrect Answers
B:
“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 prefer evaluated forecasting models for quantitative predictions. Historical seasonality and promotions drive numerical forecasts. Forecast error beats a defined naive baseline across key items. Numerical forecasts require measured error against historical demand and known seasonal drivers. A fluent language-model narrative alone does not establish predictive accuracy or calibrated uncertainty.
C:
“Only required services access credentials and audit logs record use” validates an outcome in a case where the evidence is: The credentials enable unauthorized paid inference requests. This question instead requires an action that can prefer evaluated forecasting models for quantitative predictions. Historical seasonality and promotions drive numerical forecasts. Forecast error beats a defined naive baseline across key items. Numerical forecasts require measured error against historical demand and known seasonal drivers. A fluent language-model narrative alone does not establish predictive accuracy or calibrated uncertainty.
D:
“No unauthorized action follows the retrieved directive” validates an outcome in a case where the evidence is: The embedded directive came from lower-trust document content. This question instead requires an action that can prefer evaluated forecasting models for quantitative predictions. Historical seasonality and promotions drive numerical forecasts. Forecast error beats a defined naive baseline across key items. Numerical forecasts require measured error against historical demand and known seasonal drivers. A fluent language-model narrative alone does not establish predictive accuracy or calibrated uncertainty.
Question 770
At Ashwood Public Services, an AI pilot claims 100 percent success after five easy tests. Evidence: The sample cannot represent real volume and edge cases. The business goal is to design statistically meaningful evaluations for the use case. Only non-sensitive synthetic data and approved sandbox systems may be used. Which bounded trial with synthetic records best tests the proposed approach?
- Run synthetic easy and hard cases rather than five cherry-picked prompts
- Evaluate mock suppliers with a clear scoring breakdown
- Test unseen fictional cases after adding two diverse exemplars
- Compare retrieval on test questions before and after chunk changes
Explanation
Correct Answer
A:
Build a larger, stratified synthetic evaluation set with easy, difficult, ambiguous, and edge-case prompts, then measure error rates by case type and record how failures change the decision. Five hand-picked successes cannot establish a reliable 100% outcome or estimate real-world performance; the experiment needs representative challenges and explicit stopping criteria.
Incorrect Answers
B:
“Evaluate mock suppliers with a clear scoring breakdown” tests a proposed change for the condition: The decision-maker must justify awards to an audit committee. This question instead requires an action that can design statistically meaningful evaluations for the use case. Build a larger, stratified synthetic evaluation set with easy, difficult, ambiguous, and edge-case prompts, then measure error rates by case type and record how failures change the decision. Five hand-picked successes cannot establish a reliable 100% outcome or estimate real-world performance; the experiment needs representative challenges and explicit stopping criteria.
C:
“Test unseen fictional cases after adding two diverse exemplars” tests a proposed change for the condition: Written instructions alone yield inconsistent section order. This question instead requires an action that can design statistically meaningful evaluations for the use case. Build a larger, stratified synthetic evaluation set with easy, difficult, ambiguous, and edge-case prompts, then measure error rates by case type and record how failures change the decision. Five hand-picked successes cannot establish a reliable 100% outcome or estimate real-world performance; the experiment needs representative challenges and explicit stopping criteria.
D:
“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 design statistically meaningful evaluations for the use case. Build a larger, stratified synthetic evaluation set with easy, difficult, ambiguous, and edge-case prompts, then measure error rates by case type and record how failures change the decision. Five hand-picked successes cannot establish a reliable 100% outcome or estimate real-world performance; the experiment needs representative challenges and explicit stopping criteria.