Microsoft AB-731 Practice Test 23
Question 220
At Crownridge Travel, a multinational enterprise launches AI pilots in several departments without common oversight. Evidence: Different owners approve conflicting risk exceptions. The business goal is to create cross-functional governance accountability. Only non-sensitive synthetic data and approved sandbox systems may be used. Which bounded trial with synthetic records best tests the proposed approach?
- Model fictional employee cohorts under several license mixes
- Test a synthetic unauthorized purchase request
- Run a fictional policy exception through the proposed council workflow
- Run a tabletop exercise on a fictional incorrect-policy incident
Explanation
Correct Answer
C:
Different owners approve conflicting risk exceptions. Run a fictional policy exception through the proposed council workflow. Named cross-functional decision rights and escalation paths prevent departments from independently granting incompatible AI-risk exceptions.
Incorrect Answers
A:
Licensing determines which users can access Copilot capabilities, but it cannot resolve conflicting ownership of AI-risk exception decisions. Named sponsors and cross-functional approval roles are required.
B:
“Test a synthetic unauthorized purchase request” tests a proposed change for the condition: The agent's action scope could exceed its intended advisory role. This question instead requires an action that can create cross-functional governance accountability. Different owners approve conflicting risk exceptions. Run a fictional policy exception through the proposed council workflow. Named cross-functional decision rights and escalation paths prevent departments from independently granting incompatible AI-risk exceptions.
D:
“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 create cross-functional governance accountability. Different owners approve conflicting risk exceptions. Run a fictional policy exception through the proposed council workflow. Named cross-functional decision rights and escalation paths prevent departments from independently granting incompatible AI-risk exceptions.
Question 221
At Orion Distribution, a manager requests actionable tasks from a long team chat. Evidence: Messages contain proposals, approvals and timestamps. The business goal is to summarize only confirmed actions from authorized discussion. 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 connector outages, leakage and stale records
- Monitor precision, re-ranking cost and query language drift
- Monitor wrong speakers and omitted chat segments
- Monitor overbroad connector scopes and unexpected writes
Explanation
Correct Answer
C:
Messages contain proposals, approvals and timestamps. Monitor wrong speakers and omitted chat segments.
Incorrect Answers
A:
“Monitor connector outages, leakage and stale records” monitors this signal: The data is not necessarily available to Microsoft 365 Copilot without configuration. This question instead requires an action that can summarize only confirmed actions from authorized discussion. Messages contain proposals, approvals and timestamps. Monitor wrong speakers and omitted chat segments.
B:
“Monitor precision, re-ranking cost and query language drift” monitors this signal: First-stage keyword search has poor top-result precision. This question instead requires an action that can summarize only confirmed actions from authorized discussion. Messages contain proposals, approvals and timestamps. Monitor wrong speakers and omitted chat segments.
D:
“Monitor overbroad connector scopes and unexpected writes” monitors this signal: Reading policy is different from invoking a high-impact connector action. This question instead requires an action that can summarize only confirmed actions from authorized discussion. Messages contain proposals, approvals and timestamps. Monitor wrong speakers and omitted chat segments.
Question 222
At Willow Creek Finance, a project manager pastes 300 pages into a single assistant prompt. Evidence: Critical terms from the oldest section may be omitted or truncated. The business goal is to design context selection rather than assuming unlimited model memory. The proposed solution must respect documented authority and organizational controls. Which approach most directly meets the stated business objective within the approved authority?
- Require staged design, evaluation, pilot and governed deployment gates
- Chunk and retrieve relevant passages with source labels
- Compare eligible models using the same approved benchmarks and terms
- Require clear citations, limits and a route to human policy owners
Explanation
Correct Answer
B:
Critical terms from the oldest section may be omitted or truncated. Chunk and retrieve relevant passages with source labels.
Incorrect Answers
A:
“Require staged design, evaluation, pilot and governed deployment gates” is a design or implementation response to the condition: There is no documented validation, deployment or monitoring plan. This question instead requires an action that can design context selection rather than assuming unlimited model memory. Critical terms from the oldest section may be omitted or truncated. Chunk and retrieve relevant passages with source labels.
C:
“Compare eligible models using the same approved benchmarks and terms” is a design or implementation response to the condition: Contractual hosting, customization and governance requirements differ. This question instead requires an action that can design context selection rather than assuming unlimited model memory. Critical terms from the oldest section may be omitted or truncated. Chunk and retrieve relevant passages with source labels.
D:
“Require clear citations, limits and a route to human policy owners” is a design or implementation response to the condition: Confidence language may conceal missing or stale evidence. This question instead requires an action that can design context selection rather than assuming unlimited model memory. Critical terms from the oldest section may be omitted or truncated. Chunk and retrieve relevant passages with source labels.
Question 223
At Elmwood Banking, an AI team proposes tuning a model with licensed reports and confidential contracts. Evidence: Rights to read documents do not automatically permit model training. The business goal is to confirm consent and rights for training data reuse. The proposed solution must respect documented authority and organizational controls. Which approach most directly meets the stated business objective within the approved authority?
- Define stop/rollback, notification and correction procedures
- Label synthetic imagery and obtain editorial approval
- Build transparent ROI reporting with benefit attribution
- Use only approved datasets with documented rights and minimization
Explanation
Correct Answer
D:
Rights to read documents do not automatically permit model training. Use only approved datasets with documented rights and minimization.
Incorrect Answers
A:
“Define stop/rollback, notification and correction procedures” is a design or implementation response to the condition: There is no owner or process for AI-specific incidents. This question instead requires an action that can confirm consent and rights for training data reuse. Rights to read documents do not automatically permit model training. Use only approved datasets with documented rights and minimization.
B:
“Label synthetic imagery and obtain editorial approval” is a design or implementation response to the condition: The image may falsely imply a real situation occurred. This question instead requires an action that can confirm consent and rights for training data reuse. Rights to read documents do not automatically permit model training. Use only approved datasets with documented rights and minimization.
C:
“Build transparent ROI reporting with benefit attribution” is a design or implementation response to the condition: Real costs include licenses, training, evaluation and human review. This question instead requires an action that can confirm consent and rights for training data reuse. Rights to read documents do not automatically permit model training. Use only approved datasets with documented rights and minimization.
Question 224
At Willow Creek Finance, an HR unit selects a candidate-screening model using overall accuracy only. Evidence: Performance differs sharply across legally relevant applicant groups. The business goal is to evaluate fairness by appropriate cohorts before consequential use. The project owner requires a measurable acceptance criterion before signoff. Which evidence would most convincingly demonstrate that the requirement has been met?
- Users understand they are speaking to AI and when humans review
- The customer can understand and challenge the decision
- Champions recommend authorized workflows and escalate uncertainty
- Similar qualified applicants receive comparable screening treatment
Explanation
Correct Answer
D:
Performance differs sharply across legally relevant applicant groups. Similar qualified applicants receive comparable screening treatment. A single aggregate score can hide disparate error rates among legally relevant applicant groups. Examine subgroup sampling, impact and decision criteria before using output in consequential screening.
Incorrect Answers
A:
“Users understand they are speaking to AI and when humans review” validates an outcome in a case where the evidence is: Users cannot understand the limits or source of the interaction. This question instead requires an action that can evaluate fairness by appropriate cohorts before consequential use. Performance differs sharply across legally relevant applicant groups. Similar qualified applicants receive comparable screening treatment. A single aggregate score can hide disparate error rates among legally relevant applicant groups. Examine subgroup sampling, impact and decision criteria before using output in consequential screening.
B:
“The customer can understand and challenge the decision” validates an outcome in a case where the evidence is: Customers need a comprehensible explanation and correction mechanism. This question instead requires an action that can evaluate fairness by appropriate cohorts before consequential use. Performance differs sharply across legally relevant applicant groups. Similar qualified applicants receive comparable screening treatment. A single aggregate score can hide disparate error rates among legally relevant applicant groups. Examine subgroup sampling, impact and decision criteria before using output in consequential screening.
C:
“Champions recommend authorized workflows and escalate uncertainty” validates an outcome in a case where the evidence is: Some champions share risky shortcuts and outdated feature claims. This question instead requires an action that can evaluate fairness by appropriate cohorts before consequential use. Performance differs sharply across legally relevant applicant groups. Similar qualified applicants receive comparable screening treatment. A single aggregate score can hide disparate error rates among legally relevant applicant groups. Examine subgroup sampling, impact and decision criteria before using output in consequential screening.
Question 225
At Harborline Products, a finance team cannot use Copilot after a license migration. Evidence: Users do not know whether the issue is access, training or configuration. The business goal is to provide clear support channels for onboarding problems. The proposed solution must respect documented authority and organizational controls. Which approach most directly meets the stated business objective within the approved authority?
- Choose an approved in-app experience and reduce unnecessary handoffs
- Publish plain-language approved usage and sensitive-data rules
- Create structured support intake and common fix guidance
- Evaluate and mitigate proxy-driven disparities with approved methods
Explanation
Correct Answer
C:
Users do not know whether the issue is access, training or configuration. Create structured support intake and common fix guidance.
Incorrect Answers
A:
“Choose an approved in-app experience and reduce unnecessary handoffs” is a design or implementation response to the condition: The assistant is not integrated into their ordinary work process. This question instead requires an action that can provide clear support channels for onboarding problems. Users do not know whether the issue is access, training or configuration. Create structured support intake and common fix guidance.
B:
“Publish plain-language approved usage and sensitive-data rules” is a design or implementation response to the condition: Conflicting messages describe data use and access controls. This question instead requires an action that can provide clear support channels for onboarding problems. Users do not know whether the issue is access, training or configuration. Create structured support intake and common fix guidance.
D:
“Evaluate and mitigate proxy-driven disparities with approved methods” is a design or implementation response to the condition: Proxies may create discriminatory patterns even without explicit protected fields. This question instead requires an action that can provide clear support channels for onboarding problems. Users do not know whether the issue is access, training or configuration. Create structured support intake and common fix guidance.
Question 226
At Juniper Operations, a compliance team needs both matching clauses and exact regulation numbers. Evidence: Vector similarity alone is not enough to preserve identifiers. The business goal is to choose retrieval matching both meaning and exact references. 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 model catalog, evaluation, deployment and security requirements
- Inspect the prebuilt catalog and documented functional gaps
- Estimate dataset scale, compute cost and target performance
- Test sample queries by exact number and paraphrased policy terms
Explanation
Correct Answer
D:
Vector similarity alone is not enough to preserve identifiers. Test sample queries by exact number and paraphrased policy terms.
Incorrect Answers
A:
“Identify model catalog, evaluation, deployment and security requirements” is a diagnostic step for the reported condition: The team must compare models, tools and safety controls as one lifecycle. This question instead requires an action that can choose retrieval matching both meaning and exact references. Vector similarity alone is not enough to preserve identifiers. Test sample queries by exact number and paraphrased policy terms.
B:
“Inspect the prebuilt catalog and documented functional gaps” is a diagnostic step for the reported condition: A prebuilt agent may already support search without bespoke code. This question instead requires an action that can choose retrieval matching both meaning and exact references. Vector similarity alone is not enough to preserve identifiers. Test sample queries by exact number and paraphrased policy terms.
C:
“Estimate dataset scale, compute cost and target performance” is a diagnostic step for the reported condition: The available budget and data are far smaller than required for foundation training. This question instead requires an action that can choose retrieval matching both meaning and exact references. Vector similarity alone is not enough to preserve identifiers. Test sample queries by exact number and paraphrased policy terms.
Question 227
At Cedar Analytics, a team pastes confidential contracts into a public AI service. Evidence: The vendor retention and training terms were not evaluated. The business goal is to choose an approved data processing path before AI use. Only non-sensitive synthetic data and approved sandbox systems may be used. Which bounded trial with synthetic records best tests the proposed approach?
- Try a synthetic contract in an approved test environment
- Compare retrieval on test questions before and after chunk changes
- Test a harmless fake exfiltration instruction in a sandbox document
- Load-test a simulated peak while monitoring error and safety rates
Explanation
Correct Answer
A:
The vendor retention and training terms were not evaluated. Try a synthetic contract in an approved test environment.
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 choose an approved data processing path before AI use. The vendor retention and training terms were not evaluated. Try a synthetic contract in an approved test environment.
C:
“Test a harmless fake exfiltration instruction in a sandbox document” tests a proposed change for the condition: The embedded directive came from lower-trust document content. This question instead requires an action that can choose an approved data processing path before AI use. The vendor retention and training terms were not evaluated. Try a synthetic contract in an approved test environment.
D:
“Load-test a simulated peak while monitoring error and safety rates” tests a proposed change for the condition: The current manual process cannot expand rapidly without hiring. This question instead requires an action that can choose an approved data processing path before AI use. The vendor retention and training terms were not evaluated. Try a synthetic contract in an approved test environment.
Question 228
At Harborline Products, a department moves from a prototype directly to paid production. Evidence: There is no documented validation, deployment or monitoring plan. The business goal is to apply a lifecycle from data preparation through ongoing evaluation. 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 vendor policy changes and data upload exceptions
- Monitor performance, incident reports and retraining triggers
- Monitor drift, maintenance feedback and false-positive workload
- Monitor per-class accuracy and drift after new complaint types appear
Explanation
Correct Answer
B:
There is no documented validation, deployment or monitoring plan. Monitor performance, incident reports and retraining triggers.
Incorrect Answers
A:
“Monitor vendor policy changes and data upload exceptions” monitors this signal: The vendor retention and training terms were not evaluated. This question instead requires an action that can apply a lifecycle from data preparation through ongoing evaluation. There is no documented validation, deployment or monitoring plan. Monitor performance, incident reports and retraining triggers.
C:
“Monitor drift, maintenance feedback and false-positive workload” monitors this signal: Threshold rules miss complex multi-variable patterns. This question instead requires an action that can apply a lifecycle from data preparation through ongoing evaluation. There is no documented validation, deployment or monitoring plan. Monitor performance, incident reports and retraining triggers.
D:
“Monitor per-class accuracy and drift after new complaint types appear” monitors this signal: The desired answer is a stable label rather than a free-form narrative. This question instead requires an action that can apply a lifecycle from data preparation through ongoing evaluation. There is no documented validation, deployment or monitoring plan. Monitor performance, incident reports and retraining triggers.
Question 229
At Meridian Imports, a public chatbot prototype occasionally outputs unsafe or abusive language. Evidence: The release needs safety checks beyond normal accuracy tests. The business goal is to integrate suitable content safety screening into deployment criteria. 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 bounded fictional adversarial input set
- Compare old and new models on an approved regression set
- Hold out fictional reports to compare prompt and tuned outputs
- Research a fictional sector using several test source summaries
Explanation
Correct Answer
A:
The release needs safety checks beyond normal accuracy tests. Test a bounded fictional adversarial input set. Conventional task accuracy does not measure unsafe content, tool misuse or privacy leakage. Add appropriate adversarial and policy tests as a separate release gate.
Incorrect Answers
B:
“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 integrate suitable content safety screening into deployment criteria. The release needs safety checks beyond normal accuracy tests. Test a bounded fictional adversarial input set. Conventional task accuracy does not measure unsafe content, tool misuse or privacy leakage. Add appropriate adversarial and policy tests as a separate release gate.
C:
“Hold out fictional reports to compare prompt and tuned outputs” tests a proposed change for the condition: An existing prompt baseline still violates approved template constraints. This question instead requires an action that can integrate suitable content safety screening into deployment criteria. The release needs safety checks beyond normal accuracy tests. Test a bounded fictional adversarial input set. Conventional task accuracy does not measure unsafe content, tool misuse or privacy leakage. Add appropriate adversarial and policy tests as a separate release gate.
D:
“Research a fictional sector using several test source summaries” tests a proposed change for the condition: The question requires multi-step evidence gathering rather than quick lookup. This question instead requires an action that can integrate suitable content safety screening into deployment criteria. The release needs safety checks beyond normal accuracy tests. Test a bounded fictional adversarial input set. Conventional task accuracy does not measure unsafe content, tool misuse or privacy leakage. Add appropriate adversarial and policy tests as a separate release gate.
Question 230
At Tamarack Logistics, an employee forwards a vendor page containing instructions to ignore system policies. Evidence: The page is untrusted external content rather than a user-approved command. The business goal is to treat lower-trust input as data and defend against instruction injection. 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?
- Locate the untrusted directive and the authorized business question
- Determine whether representative examples already meet acceptance criteria
- Identify prompts that should have equivalent outcomes
- Open cited passages and compare each claim to supporting text
Explanation
Correct Answer
A:
The page is untrusted external content rather than a user-approved command. Locate the untrusted directive and the authorized business question.
Incorrect Answers
B:
“Determine whether representative examples already meet acceptance criteria” is a diagnostic step for the reported condition: No task-specific labeled training data has yet been collected. This question instead requires an action that can treat lower-trust input as data and defend against instruction injection. The page is untrusted external content rather than a user-approved command. Locate the untrusted directive and the authorized business question.
C:
“Identify prompts that should have equivalent outcomes” is a diagnostic step for the reported condition: Small phrasing changes cause inconsistent answers in a business-critical step. This question instead requires an action that can treat lower-trust input as data and defend against instruction injection. The page is untrusted external content rather than a user-approved command. Locate the untrusted directive and the authorized business question.
D:
“Open cited passages and compare each claim to supporting text” is a diagnostic step for the reported condition: Retrieval produced a plausible paragraph but not supporting evidence. This question instead requires an action that can treat lower-trust input as data and defend against instruction injection. The page is untrusted external content rather than a user-approved command. Locate the untrusted directive and the authorized business question.