Microsoft AB-100 (Agentic AI Business Solutions Architect) Practice Test 115

Question 1140

A staging test has to detect the specific defect that it autonomously contacts a supplier without approval. In the workflow where a bot only needs to check warranty and prepare a draft, which exercise verifies the corrective behavior?

  1. Change a test case status and inspect audit events
  2. Generate stressful and adversarial user examples
  3. Ask unrelated questions and verify no supplier action is triggered
  4. Ask for a variance explanation with incomplete data

Explanation

Correct Answer

C:

The historical failure is that it autonomously contacts a supplier without approval; for the setting where a bot only needs to check warranty and prepare a draft, asking unrelated questions and verifying no supplier action is triggered verifies the dependency where limited intent and tool scope prevent unnecessary actions.

Incorrect Answers

A:

Although the exercise might catch its knowledge source can explain but cannot commit a database change, it would not directly validate the mechanism that limited intent and tool scope prevent unnecessary actions in the target implementation.

B:

For compliance letters must preserve approved language and disclaimers, the alternate test is about whether templating and approved clauses limit generated variation; it cannot prove the outcome needed after the reported regression where it autonomously contacts a supplier without approval.

D:

The tested behavior is that grounded explanatory responses distinguish evidence from hypothesis; the necessary evidence here concerns limited intent and tool scope prevent unnecessary actions and the previously observed it autonomously contacts a supplier without approval.

Question 1141

The pilot needs a durable correction: A full agent workflow adds cost to simple text tagging. This affects the workflow in which the system must detect customer sentiment and named entities at scale. Which design choice should the architect approve?

  1. Choose speech processing for transcribed contact-center audio
  2. Version models and datasets together for controlled adaptation
  3. Choose language analytics for focused text classification
  4. Map sales-stage language to actual opportunity fields

Explanation

Correct Answer

C:

The system must detect customer sentiment and named entities at scale; its failing behavior is that a full agent workflow adds cost to simple text tagging; applying the underlying design principle—purpose-built text analysis handles bounded language-analysis tasks—connects the requirement to an implementable control.

Incorrect Answers

A:

Speech transcription transforms audio to text for subsequent analysis; its main purpose is quality teams need searchable insights from recorded calls, not correction of the reported defect where a full agent workflow adds cost to simple text tagging in the target process.

B:

For retraining uses revised labels and a different source snapshot, this other approach relies on the fact that dataset lineage and model versioning support repeatable assessment; the issue here is a full agent workflow adds cost to simple text tagging; that mechanism does not establish the required purpose-built text analysis handles bounded language-analysis tasks.

D:

An agent updates the wrong pipeline field from an informal phrase; the alternate design can address that situation because canonical business terminology maps language to underlying Dynamics records, whereas this question needs a solution to a full agent workflow adds cost to simple text tagging.

Question 1142

The architecture board has evidence that logs show deletion but not the requesting principal or approval. The affected situation is: A tool deletes a record following an ambiguous conversation. What change most directly establishes the missing behavior?

  1. Test Finance-to-Supply-Chain exception handling after partial failure
  2. Record immutable model deployment events with approver and version
  3. Register each trained model with dataset and evaluation provenance
  4. Audit privileged agent actions with correlation to user intent

Explanation

Correct Answer

D:

The architecture must reconcile two facts: a tool deletes a record following an ambiguous conversation, yet logs show deletion but not the requesting principal or approval; the selected pattern does so through the mechanism that end-to-end audit binds user, tool, affected record and authorization event.

Incorrect Answers

A:

When a blocked invoice triggers a hold on related procurement actions, failure-injection tests verify retries and compensation between systems; in the stated workflow, however, logs show deletion but not the requesting principal or approval; the relevant architectural behavior is that end-to-end audit binds user, tool, affected record and authorization event.

B:

When a model is replaced without a traceable change ticket, deployment audit links actor, approved artifact and release time; in the stated workflow, however, logs show deletion but not the requesting principal or approval; the relevant architectural behavior is that end-to-end audit binds user, tool, affected record and authorization event.

C:

The alternative addresses models move to production without an audit of training inputs by a design in which a model registry ties versions to lineage, evaluations and approval; it does not supply the missing relationship for a tool deletes a record following an ambiguous conversation, where logs show deletion but not the requesting principal or approval.

Question 1143

A product owner must measure whether an AI-enabled change actually improves this process: An answer contradicts the handbook despite high model fluency. The earlier problem was: Engineers tune prompts without checking whether evidence was retrieved. Which outcome indicator is most relevant?

  1. Median time to correctly triage an agent incident
  2. Root-cause clusters converted to verified corrections
  3. Improvement in evidence-supported answer correctness
  4. Concessions with full recommendation-to-posting audit chain
  5. Correct source passage retrieved within top K

Explanation

Correct Answer

C:

For the use case where an answer contradicts the handbook despite high model fluency, the indicator improvement in evidence-supported answer correctness is worth baselining against the earlier issue where engineers tune prompts without checking whether evidence was retrieved.

Incorrect Answers

A:

The alternative tracks median time to correctly triage an agent incident, arising from the mechanism where structured defect taxonomy assigns the right remediation owner; its movement alone says little about the failure that engineers tune prompts without checking whether evidence was retrieved.

B:

Although root-cause clusters converted to verified corrections is observable, the affected process involves an answer contradicts the handbook despite high model fluency; its acceptance outcome is improvement in evidence-supported answer correctness, not the alternative measurement.

D:

Although concessions with full recommendation-to-posting audit chain is observable, the affected process involves an answer contradicts the handbook despite high model fluency; its acceptance outcome is improvement in evidence-supported answer correctness, not the alternative measurement.

E:

Correct source passage retrieved within top K; that indicator describes a distinct workflow where an assistant responds fluently but cites unrelated support articles; it does not quantify improvement in evidence-supported answer correctness, which is needed for the stated case.

Question 1144

The operating team must monitor whether an architectural correction works. Business context: Prompt revisions cause intermittent regressions. Historical symptom: Makers validate only a few manual chats. Which outcome measure is the best match?

  1. Preproduction regressions caught before agent publication
  2. User-impacting incidents detected within response target
  3. Multi-turn tasks resolved without entity confusion
  4. Candidate releases satisfying every critical acceptance metric

Explanation

Correct Answer

A:

This scorecard needs an outcome related to prompt revisions cause intermittent regressions; preproduction regressions caught before agent publication. That indicator can be checked against the earlier symptom that makers validate only a few manual chats.

Incorrect Answers

B:

Although user-impacting incidents detected within response target is observable, the affected process involves prompt revisions cause intermittent regressions; its acceptance outcome is preproduction regressions caught before agent publication, not the alternative measurement.

C:

For the situation where agent must remember order context while asking for missing details, multi-turn tasks resolved without entity confusion is a sensible indicator; the problem that makers validate only a few manual chats requires measurement of preproduction regressions caught before agent publication.

D:

For the workflow where teams promote a custom model because it feels better, candidate releases satisfying every critical acceptance metric is an actionable measure; it does not establish whether makers validate only a few manual chats in the situation under review.

Question 1145

A solution architecture is being reviewed for this business situation: Many teams copy slightly different instructions for the same task. The observed defect is that a successful prompt change cannot be reproduced after rollout. Which architectural decision addresses the underlying requirement?

  1. Build a bespoke agent when differentiated orchestration is essential
  2. Maintain a reusable approved-agent and prompt catalogue
  3. Give prompts explicit role, task, constraints and response schema
  4. Version reusable prompts with owners and evaluation examples

Explanation

Correct Answer

D:

Without the intended behavior, a successful prompt change cannot be reproduced after rollout; the proposed architecture establishes the specific design principle that governed templates provide versioned contracts and test fixtures for the process where many teams copy slightly different instructions for the same task.

Incorrect Answers

A:

This design would make sense if competitive requirements depend on unique multi-model control, particularly where a boxed product cannot express the execution and ownership rules; the requested architecture instead needs the dependency that governed templates provide versioned contracts and test fixtures because a successful prompt change cannot be reproduced after rollout.

B:

In the case where teams rebuild similar assistants without seeing existing assets, the alternate design depends on discovery and ownership metadata support reuse and controlled deprecation; it does not correct the target failure, a successful prompt change cannot be reproduced after rollout.

C:

The technical behavior of this alternative is that clear instructions and structured outputs reduce avoidable ambiguity; the user case being assessed is many teams copy slightly different instructions for the same task with a different failure: a successful prompt change cannot be reproduced after rollout.

Question 1146

The rollout supports this workflow: A retrieved email contains instructions to bypass approval. Leaders want a metric that responds to the reported shortcoming that the agent follows lower-trust text as an order. Which measure best reflects success?

  1. Privileged mutations without verified principal authorization
  2. Successful indirect instruction override attempts
  3. Unauthorized action rate in adversarial evaluations
  4. Unsupported claims emitted in low-evidence tests

Explanation

Correct Answer

B:

Successful indirect instruction override attempts; this outcome measure belongs to the workflow where a retrieved email contains instructions to bypass approval; a change in it can help detect whether the defect that the agent follows lower-trust text as an order is resolved.

Incorrect Answers

A:

Although privileged mutations without verified principal authorization is observable, the affected process involves a retrieved email contains instructions to bypass approval; its acceptance outcome is successful indirect instruction override attempts, not the alternative measurement.

C:

For the workflow where benchmarks cover only cooperative user wording, unauthorized action rate in adversarial evaluations is an actionable measure; it does not establish whether the agent follows lower-trust text as an order in the situation under review.

D:

For the situation where policy answers must be tied to a current source, unsupported claims emitted in low-evidence tests is a sensible indicator; the problem that the agent follows lower-trust text as an order requires measurement of successful indirect instruction override attempts.

Question 1147

A staging test has to detect the specific defect that a broad action agent can modify records unnecessarily. In the workflow where marketing specialists need tailored descriptions from existing briefings, which exercise verifies the corrective behavior?

  1. Request drafts using multiple brief formats
  2. Request the same account help by voice and chat
  3. Change a test case status and inspect audit events
  4. Ask unrelated questions and verify no supplier action is triggered

Explanation

Correct Answer

A:

A broad action agent can modify records unnecessarily; a release test must establish that this no longer occurs; requesting drafts using multiple brief formats measures the specific behavior for which a response-focused agent composes controlled text from supplied context.

Incorrect Answers

B:

Although the exercise might catch voice responses contain unreadable links and lengthy tables, it would not directly validate the mechanism that a response-focused agent composes controlled text from supplied context in the target implementation.

C:

The suggested test serves an agent needs to update a governed ticket status, in which its knowledge source can explain but cannot commit a database change; the present scenario instead concerns the reported problem where a broad action agent can modify records unnecessarily.

D:

In the alternate setting where a bot only needs to check warranty and prepare a draft, the test exercises limited intent and tool scope prevent unnecessary actions; even a pass would not show that a broad action agent can modify records unnecessarily has been prevented.

Question 1148

Troubleshooting starts from a business complaint: Final answer text conceals the erroneous tool choice. The affected process is: The planner calls an order-cancel tool for status requests. Which investigation is most directly tied to the reported defect?

  1. Compare complaint time with retrieved source and model versions
  2. Compare intended task with executed tool names and inputs
  3. Inspect storage retention settings and log contents
  4. Inspect pending runs, deadlines and compensation events

Explanation

Correct Answer

B:

For the situation where the planner calls an order-cancel tool for status requests, the defect is that final answer text conceals the erroneous tool choice; comparing intended task with executed tool names and inputs helps verify the cause-and-effect path: step traces expose planning, selected tools and arguments.

Incorrect Answers

A:

That probe evaluates trace and content-version linkage makes failures testable in a workflow where users report stale answers after a policy refresh; the reported failure concerns the path for which step traces expose planning, selected tools and arguments instead.

C:

In the separate scenario where debugging logs include prompts containing customer data, the same investigation could explain historical traces remain available longer than approved retention; it does not isolate the reported cause in a system where the planner calls an order-cancel tool for status requests.

D:

Timeout creates an abandoned workflow with no owner; examining that other path does not determine whether step traces expose planning, selected tools and arguments, which is what distinguishes the issue in the stated use case.

Question 1149

In a business situation where prompt wording persuades the agent to submit a privileged operation, a previous build showed this defect: API trusts the model's text assertion of permission. Which reproducible test belongs in the release gate?

  1. Disable knowledge access and verify safe no-evidence behavior
  2. Inject a malicious export instruction into a source document
  3. Insert override strings into a test retrieved document
  4. Try a forbidden refund despite a permissive model instruction

Explanation

Correct Answer

D:

The check is targeted rather than generic: server-side access checks enforce real actor entitlements; its pass/fail evidence applies to the workflow where prompt wording persuades the agent to submit a privileged operation and to the prior defect that API trusts the model's text assertion of permission.

Incorrect Answers

A:

A pass on that verification would speak to policy answers must be tied to a current source; it would not show that server-side access checks enforce real actor entitlements, which is needed to avoid the observed defect API trusts the model's text assertion of permission.

B:

In the alternate setting where a prompt injection requests data export to an outside address, the test exercises execution-time destination and permission checks block exfiltration; even a pass would not show that API trusts the model's text assertion of permission has been prevented.

C:

Inserting override strings into a test retrieved document verifies an adjacent objective where a retrieved email contains instructions to bypass approval; it leaves untested the causal path in which server-side access checks enforce real actor entitlements for this case.

Question 1150

Risk reviewers need a control, not merely a monitoring dashboard. The use case is: An answer contradicts the handbook despite high model fluency. Potential harm: The team might publish a model that never retrieves the evidence it requires. Which protection fits?

  1. Set cost budgets and investigate anomalous context expansion
  2. Do not share private cached outputs across principals
  3. Block publishing when required evidence cannot be recovered
  4. Fail closed for unsupported sensitive answers

Explanation

Correct Answer

C:

When the team might publish a model that never retrieves the evidence it requires, the use case an answer contradicts the handbook despite high model fluency requires a guardrail aligned with the behavior that separate relevance and groundedness tests isolate the failing layer.

Incorrect Answers

A:

That protection is tied to the mechanism that token telemetry isolates prompt, retrieval and generation cost changes; it does not prevent the distinct risk where the team might publish a model that never retrieves the evidence it requires in the system under review.

B:

Another process could need this guardrail because a user might receive another principal's cached protected answer; the design being assessed supports an answer contradicts the handbook despite high model fluency and has the explicit exposure that the team might publish a model that never retrieves the evidence it requires.

D:

The safeguard for an assistant responds fluently but cites unrelated support articles deals with a different threat: missing retrieval evidence might still generate definitive sensitive advice; even after it is enforced, the exposure that the team might publish a model that never retrieves the evidence it requires remains possible in the target workflow where separate relevance and groundedness tests isolate the failing layer.

Agentic AI Business Solutions Architect (AB-100) Practice Tests