Microsoft AB-731 Practice Test 2
Question 10
At Tamarack Logistics, 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 project owner requires a measurable acceptance criterion before signoff. Which evidence would most convincingly demonstrate that the requirement has been met?
- The external message is sent only after approved corrections
- Answers stay grounded and preserve source version context
- Only accepted decisions are reported as final
- No unauthorized tool call or sensitive disclosure occurs
Explanation
Correct Answer
A:
Generated text could contain binding promises beyond approval. The external message is sent only after approved corrections. Draft generation is not final authorization; a human approval step prevents unsupported contractual commitments from being sent.
Incorrect Answers
B:
A notebook helps maintain background source context; it does not authorize AI-drafted contractual commitments. Have an accountable approver validate external promises.
C:
A meeting recap summarizes approved discussion content. It does not make every AI-generated contractual promise approved for release.
D:
Prompt-injection defenses mitigate malicious instructions in retrieved content, but cannot substitute for commercial or legal approval of the promises in a proposed response.
Question 11
At Cedar Analytics, a warehouse agent is asked to update stock levels in an ERP. Evidence: Reading policy is different from invoking a high-impact connector action. The business goal is to separate knowledge responses from authorized transactional actions. The proposed solution must respect documented authority and organizational controls. Which approach most directly meets the stated business objective within the approved authority?
- Review Copilot recap and flag unconfirmed commitments
- Enable an action only with appropriate identity and permissions
- Pilot an approved vision model with supervised damage labels
- Draft in Outlook with verified language and require manual sending
Explanation
Correct Answer
B:
Reading policy is different from invoking a high-impact connector action. Enable an action only with appropriate identity and permissions. A knowledge answer is read-only, whereas an ERP inventory update changes business state. The connector must enforce action-level identity, authorization and an auditable approval path before any write.
Incorrect Answers
A:
Meeting summarization is a read-oriented assistance task. Executing a high-impact connector action requires separately authorized permissions, guardrails and explicit approval.
C:
Visual content extraction cannot establish that an agent may invoke a connector that modifies sensitive enterprise data. Tool execution has an independent authorization boundary.
D:
Drafting an Outlook message and performing a high-impact external action have different risk levels. The latter needs user permissions, confirmation and an appropriate action policy.
Question 12
At Redwood Consulting, a language model begins answering poorly after a source system changes schema. Evidence: Retrieved field meanings changed even though model version is unchanged. The business goal is to diagnose data pipeline changes before blaming model weights. Only non-sensitive synthetic data and approved sandbox systems may be used. Which bounded trial with synthetic records best tests the proposed approach?
- Generate synthetic icons without competitor logos for a mock campaign
- Run a synthetic migration with renamed metadata fields
- Run repeat extraction on invoices with the same hidden reference
- Compare top-three and top-ten retrieval on synthetic questions
Explanation
Correct Answer
B:
Retrieved field meanings changed even though model version is unchanged. Run a synthetic migration with renamed metadata fields. Changing an upstream field definition changes the data the model sees even if the model weights are fixed. Check mapping lineage, schema and retrieval output before attributing the regression to model behavior.
Incorrect Answers
A:
Third-party intellectual-property controls govern data and rights, not semantic drift in source-field definitions. Revalidate the retrieval schema and field interpretation.
C:
Temperature affects generation sampling, but cannot correct source field definitions that changed without a model-version update.
D:
Precision and recall measure retrieval effectiveness; they do not establish that an unchanged field name still has the same business meaning. Validate the new schema semantics.
Question 13
At Crownridge Travel, a finance director sees model spending rise despite stable request counts. Evidence: Longer retrieved passages and responses increase processed input and output tokens. The business goal is to estimate consumption using token counts and actual price meters. 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?
- Determine whether representative examples already meet acceptance criteria
- Inspect input, output and retrieval tokens per request type
- Locate the untrusted directive and the authorized business question
- Check consent, recording access and transcript retention requirements
Explanation
Correct Answer
B:
Longer retrieved passages and responses increase processed input and output tokens. Inspect input, output and retrieval tokens per request type. Retrieval length and generated response length both contribute billable tokens. Estimate workload cost from measured input/output token distributions and the actual deployed model tariff, not a single average prompt.
Incorrect Answers
A:
A pretrained model choice may alter performance or price, but it does not explain extra usage from longer prompts and generated responses under the same model.
C:
Prompt-injection risk concerns untrusted instructions. The cost increase here follows additional input and output tokens, so token accounting and context reduction address it.
D:
Audio transcription has its own privacy requirements; it is not the cause of higher text-token billing from expanded retrieved passages.
Question 14
At Coral Bay Systems, a team selects a model after reading its marketing benchmark. Evidence: The published scores do not measure the company's actual document tasks. The business goal is to build task-specific evaluation before deployment. 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 leaked tokens and abnormal endpoint traffic
- Monitor underused capacity and capacity-constrained requests
- Monitor feature drift and duplicate custom expenses
- Monitor evaluation drift as source documents change
Explanation
Correct Answer
D:
The published scores do not measure the company's actual document tasks. Monitor evaluation drift as source documents change. A public benchmark cannot establish reliability on the organization’s documents, permission boundaries or failure modes. A representative task set with expected outcomes exposes those deployment-specific gaps.
Incorrect Answers
A:
Managed identity and endpoint authentication protect who can call a model. They do not show how accurately it performs the company’s actual document tasks.
B:
Dedicated hosting changes capacity and cost tradeoffs, not the relevance of an external benchmark to the organization’s own document workload.
C:
Choosing a custom or prebuilt agent determines the implementation route; performance still needs evaluation against representative internal tasks and criteria.
Question 15
At Silverpine Group, a loan assessment model omits race but uses highly correlated location variables. Evidence: Proxies may create discriminatory patterns even without explicit protected fields. The business goal is to test fairness effects beyond removal of direct identifiers. 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 stakeholders, skills and handoffs for rollout stages
- Identify data categories, purposes and potential intrusive inference
- Check visual provenance, license and any implied factual claims
- Analyze feature correlations and group outcome differences
Explanation
Correct Answer
D:
Proxies may create discriminatory patterns even without explicit protected fields. Analyze feature correlations and group outcome differences. Variables correlated with protected characteristics can preserve unequal outcomes after explicit identifiers are removed. Cohort-level testing and investigation of the influencing features are needed to identify proxy effects.
Incorrect Answers
A:
Organizing an adoption team is a rollout decision. Proxy variables can still create disparate outcomes, so evaluate the model’s impact across relevant groups and use cases.
B:
A privacy impact assessment considers data-protection risks; discrimination can also arise from proxy features even when explicit sensitive fields are absent. Test outcomes for disparate impact.
C:
Image-generation provenance and safety govern generated visuals, not bias from proxy variables in a decision model.
Question 16
At Blue Ridge Technology, a provider announces deprecation of an existing deployed model. Evidence: The successor may change behavior and cost. The business goal is to plan version migration using a controlled compatibility evaluation. 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?
- Define representative input examples and expected outcomes
- Review templates for names, passwords, private links and assumptions
- Review retirement dates and dependent application behavior
- Review human relevance labels and current ranking
Explanation
Correct Answer
C:
The successor may change behavior and cost. Review retirement dates and dependent application behavior. A replacement model can alter prompts, output shape, latency, safety behavior and unit pricing. Compare the new version against fixed acceptance cases before redirecting production traffic.
Incorrect Answers
A:
“Define representative input examples and expected outcomes” is a diagnostic step for the reported condition: The published scores do not measure the company's actual document tasks. This question instead requires an action that can plan version migration using a controlled compatibility evaluation. The successor may change behavior and cost. Review retirement dates and dependent application behavior. A replacement model can alter prompts, output shape, latency, safety behavior and unit pricing. Compare the new version against fixed acceptance cases before redirecting production traffic.
B:
“Review templates for names, passwords, private links and assumptions” is a diagnostic step for the reported condition: A prompt may be reusable but embedded personal details are not. This question instead requires an action that can plan version migration using a controlled compatibility evaluation. The successor may change behavior and cost. Review retirement dates and dependent application behavior. A replacement model can alter prompts, output shape, latency, safety behavior and unit pricing. Compare the new version against fixed acceptance cases before redirecting production traffic.
D:
“Review human relevance labels and current ranking” is a diagnostic step for the reported condition: First-stage keyword search has poor top-result precision. This question instead requires an action that can plan version migration using a controlled compatibility evaluation. The successor may change behavior and cost. Review retirement dates and dependent application behavior. A replacement model can alter prompts, output shape, latency, safety behavior and unit pricing. Compare the new version against fixed acceptance cases before redirecting production traffic.
Question 17
At Elmwood Banking, a board secretary wants a concise deck from a 30-page approved strategy. Evidence: Slides need reviewable editing in the target presentation. The business goal is to select a presentation-aware workflow rather than chat-only prose. The project owner requires a measurable acceptance criterion before signoff. Which evidence would most convincingly demonstrate that the requirement has been met?
- No unauthorized tool call or sensitive disclosure occurs
- The public-only experience never claims unseen private details
- The agent cites the latest authorized edition and dates
- Slides preserve decision asks, limitations and supporting figures
Explanation
Correct Answer
D:
Slides need reviewable editing in the target presentation. Slides preserve decision asks, limitations and supporting figures. Producing slide text in chat is not equivalent to editing an actual presentation. The chosen workflow must retain editable slide structure, source attribution and review of visual layout in the target file.
Incorrect Answers
A:
“No unauthorized tool call or sensitive disclosure occurs” validates an outcome in a case where the evidence is: The document includes text telling the assistant to export hidden records. This question instead requires an action that can select a presentation-aware workflow rather than chat-only prose. Slides need reviewable editing in the target presentation. Slides preserve decision asks, limitations and supporting figures. Producing slide text in chat is not equivalent to editing an actual presentation. The chosen workflow must retain editable slide structure, source attribution and review of visual layout in the target file.
B:
“The public-only experience never claims unseen private details” validates an outcome in a case where the evidence is: The service has not been authorized to access the requested private file. This question instead requires an action that can select a presentation-aware workflow rather than chat-only prose. Slides need reviewable editing in the target presentation. Slides preserve decision asks, limitations and supporting figures. Producing slide text in chat is not equivalent to editing an actual presentation. The chosen workflow must retain editable slide structure, source attribution and review of visual layout in the target file.
C:
“The agent cites the latest authorized edition and dates” validates an outcome in a case where the evidence is: The source library was not refreshed after recent rule changes. This question instead requires an action that can select a presentation-aware workflow rather than chat-only prose. Slides need reviewable editing in the target presentation. Slides preserve decision asks, limitations and supporting figures. Producing slide text in chat is not equivalent to editing an actual presentation. The chosen workflow must retain editable slide structure, source attribution and review of visual layout in the target file.
Question 18
At Riverbend Publishing, a customer-support pilot reports five-star ratings from only satisfied users. Evidence: Feedback is self-selected and omits escalated failures. The business goal is to collect representative outcome evidence rather than vanity scores. The project owner requires a measurable acceptance criterion before signoff. Which evidence would most convincingly demonstrate that the requirement has been met?
- New specification is retrieved while retired content is excluded
- Similar eligible customers receive comparable recommendations
- Baseline quality and cost are measured before fine-tuning is proposed
- Reported improvement holds across both successful and escalated cases
Explanation
Correct Answer
D:
Feedback is self-selected and omits escalated failures. Reported improvement holds across both successful and escalated cases. Voluntary feedback overrepresents people who chose to respond and can omit the most serious escalations. Link evaluation to representative task outcomes, failure logs and independent acceptance criteria.
Incorrect Answers
A:
“New specification is retrieved while retired content is excluded” validates an outcome in a case where the evidence is: Its vector index does not reflect new catalog documents. This question instead requires an action that can collect representative outcome evidence rather than vanity scores. Feedback is self-selected and omits escalated failures. Reported improvement holds across both successful and escalated cases. Voluntary feedback overrepresents people who chose to respond and can omit the most serious escalations. Link evaluation to representative task outcomes, failure logs and independent acceptance criteria.
B:
“Similar eligible customers receive comparable recommendations” validates an outcome in a case where the evidence is: Training examples omit important customer contexts. This question instead requires an action that can collect representative outcome evidence rather than vanity scores. Feedback is self-selected and omits escalated failures. Reported improvement holds across both successful and escalated cases. Voluntary feedback overrepresents people who chose to respond and can omit the most serious escalations. Link evaluation to representative task outcomes, failure logs and independent acceptance criteria.
C:
“Baseline quality and cost are measured before fine-tuning is proposed” validates an outcome in a case where the evidence is: No task-specific labeled training data has yet been collected. This question instead requires an action that can collect representative outcome evidence rather than vanity scores. Feedback is self-selected and omits escalated failures. Reported improvement holds across both successful and escalated cases. Voluntary feedback overrepresents people who chose to respond and can omit the most serious escalations. Link evaluation to representative task outcomes, failure logs and independent acceptance criteria.
Question 19
At Willow Creek Finance, a project leader asks for an automatic weekly progress summary. Evidence: An ordinary chat response does not itself establish future recurring delivery. The business goal is to use an approved scheduling workflow where supported. 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 disabled schedules and unexpected recipients
- Monitor overbroad connector scopes and unexpected writes
- Monitor conflicting edits and unreviewed content changes
- Monitor evaluation drift as source documents change
Explanation
Correct Answer
A:
An ordinary chat response does not itself establish future recurring delivery. Monitor disabled schedules and unexpected recipients. A one-time response cannot deliver future updates by itself. Recurring execution requires an enabled scheduling mechanism, an explicit cadence, permitted data access and a way to observe failures.
Incorrect Answers
B:
“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 use an approved scheduling workflow where supported. An ordinary chat response does not itself establish future recurring delivery. Monitor disabled schedules and unexpected recipients. A one-time response cannot deliver future updates by itself. Recurring execution requires an enabled scheduling mechanism, an explicit cadence, permitted data access and a way to observe failures.
C:
“Monitor conflicting edits and unreviewed content changes” monitors this signal: Parallel copies cause conflicting versions of one plan. This question instead requires an action that can use an approved scheduling workflow where supported. An ordinary chat response does not itself establish future recurring delivery. Monitor disabled schedules and unexpected recipients. A one-time response cannot deliver future updates by itself. Recurring execution requires an enabled scheduling mechanism, an explicit cadence, permitted data access and a way to observe failures.
D:
“Monitor evaluation drift as source documents change” monitors this signal: The published scores do not measure the company's actual document tasks. This question instead requires an action that can use an approved scheduling workflow where supported. An ordinary chat response does not itself establish future recurring delivery. Monitor disabled schedules and unexpected recipients. A one-time response cannot deliver future updates by itself. Recurring execution requires an enabled scheduling mechanism, an explicit cadence, permitted data access and a way to observe failures.
Question 20
At Marigold Education, a knowledge assistant returns an executive file to an unauthorized clerk. Evidence: Retrieval happens before checking the caller's permissions. The business goal is to enforce access at retrieval, not just by asking the model to refuse. 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 dummy transfer instruction against a read-only sandbox
- Compare a prompt-only baseline to a constrained fine-tuned pilot on held-out cases
- Search for a fictional policy using both phrase and paraphrase
- Test permitted and denied identities against mock confidential files
Explanation
Correct Answer
D:
Use synthetic documents with distinct access grants for two test identities, then query the same assistant as both. Confirm retrieval applies the caller's permissions before passages reach the model and that the unauthorized clerk cannot receive or infer the executive file. A refusal-style prompt alone is not an authorization control.
Incorrect Answers
A:
“Run a dummy transfer instruction against a read-only sandbox” tests a proposed change for the condition: Retrieving an account statement does not authorize transactions. This question instead requires an action that can enforce access at retrieval, not just by asking the model to refuse. Use synthetic documents with distinct access grants for two test identities, then query the same assistant as both. Confirm retrieval applies the caller's permissions before passages reach the model and that the unauthorized clerk cannot receive or infer the executive file. A refusal-style prompt alone is not an authorization control.
B:
“Compare a prompt-only baseline to a constrained fine-tuned pilot on held-out cases” tests a proposed change for the condition: Grounding documents are current but outputs repeatedly violate a defined writing format. This question instead requires an action that can enforce access at retrieval, not just by asking the model to refuse. Use synthetic documents with distinct access grants for two test identities, then query the same assistant as both. Confirm retrieval applies the caller's permissions before passages reach the model and that the unauthorized clerk cannot receive or infer the executive file. A refusal-style prompt alone is not an authorization control.
C:
“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 enforce access at retrieval, not just by asking the model to refuse. Use synthetic documents with distinct access grants for two test identities, then query the same assistant as both. Confirm retrieval applies the caller's permissions before passages reach the model and that the unauthorized clerk cannot receive or infer the executive file. A refusal-style prompt alone is not an authorization control.