Microsoft AB-100 (Agentic AI Business Solutions Architect) Practice Test 15

Question 140

An incident review has one confirmed symptom: Small-model outputs omit critical contractual exceptions. For the use case where some cases need cross-document reasoning and detailed justification, which targeted diagnostic check would reveal the most pertinent evidence?

  1. Review escalated case outcomes and failure patterns
  2. Inspect join keys and collision rates in the ingestion pipeline
  3. Audit whether failures concern style rather than missing current facts
  4. Compare host limitations against the requirements matrix

Explanation

Correct Answer

A:

The reported fault is that small-model outputs omit critical contractual exceptions in a system where some cases need cross-document reasoning and detailed justification; evidence from reviewing escalated case outcomes and failure patterns can show whether complexity classification selects a capable model for difficult tasks.

Incorrect Answers

B:

That probe evaluates canonical keys and deduplication prevent incorrect cross-record joins in a workflow where multiple systems use inconsistent customer identifiers and duplicate cases; the reported failure concerns the path for which complexity classification selects a capable model for difficult tasks instead.

C:

The proposed probe concerns a stable specialized classification style persists despite prompt examples, where fine-tuning can adapt recurring behavior using vetted examples; this incident instead requires evidence about complexity classification selects a capable model for difficult tasks and the failure that small-model outputs omit critical contractual exceptions.

D:

That probe evaluates owning the engine provides control over models, routing and execution in a workflow where a workflow needs proprietary scheduling, model choice and runtime control; the reported failure concerns the path for which complexity classification selects a capable model for difficult tasks instead.

Question 141

An operational risk could arise because unreviewed training examples might poison a promoted model. The AI system supports the business need that business analysts can submit examples but not change production behavior. Which guardrail is designed for this threat?

  1. Require reviewed datasets and independent promotion sign-off
  2. Prevent sensitive data from entering prompts unless required
  3. Rotate compromised keys and enforce least-privilege callers
  4. Reject unregistered or unapproved models from production

Explanation

Correct Answer

A:

In the scenario where business analysts can submit examples but not change production behavior, the explicit risk is that unreviewed training examples might poison a promoted model; enforcement must follow the mechanism by which separated data-curation and release permissions preserve integrity.

Incorrect Answers

B:

An internal model sometimes returns confidential text verbatim; its guardrail addresses the possibility that a model might expose sensitive user data in uncontrolled prompt output; the present architecture must constrain the different threat that unreviewed training examples might poison a promoted model.

C:

The alternate guard addresses a leaked endpoint credential might permit inference from an untrusted client for a sensitive model endpoint is reachable using a broadly shared key; the present threat, unreviewed training examples might poison a promoted model, needs a different enforcement boundary.

D:

For the use case where data scientists cannot explain why a newer classifier behaves differently, the control addresses the threat that a model outside the approved registry might reach production; the specified agent must be protected against unreviewed training examples might poison a promoted model.

Question 142

A business stakeholder describes this operating situation: An assistant responds fluently but cites unrelated support articles. Testing instead shows that surface quality scores appear healthy despite wrong documents. What design approach is most appropriate?

  1. Use cost-versus-quality dashboards to evaluate model changes
  2. Link negative feedback to grounded traces and source versions
  3. Inspect retrieval spans when cited answers omit evidence
  4. Inspect model-call timing separately from integration latency

Explanation

Correct Answer

C:

An assistant responds fluently but cites unrelated support articles; its failing behavior is that surface quality scores appear healthy despite wrong documents; applying the underlying design principle—retrieval telemetry distinguishes evidence search from answer generation—connects the requirement to an implementable control.

Incorrect Answers

A:

For a new model reduces price but also degrades exact extractions, this other approach relies on the fact that joint quality and consumption measures capture real efficiency; the issue here is surface quality scores appear healthy despite wrong documents; that mechanism does not establish the required retrieval telemetry distinguishes evidence search from answer generation.

B:

In the case where users report stale answers after a policy refresh, the alternate design depends on trace and content-version linkage makes failures testable; it does not correct the target failure, surface quality scores appear healthy despite wrong documents.

D:

Per-span latency isolates model work, network waits and tool execution; its main purpose is agents regularly exceed an interactive response-time objective, not correction of the reported defect where surface quality scores appear healthy despite wrong documents in the target process.

Question 143

After deployment, stakeholders need an interpretable measure for this use case: Real task mix differs from prelaunch performance tests. The pilot showed that router saves tokens but sends difficult tasks to weak models. Which metric captures the needed result?

  1. High-complexity case pass rate by model route
  2. Unsupported-host requirements met by the custom design
  3. Wrong-entity retrieval rate in curated evaluation cases
  4. Accuracy-adjusted cost per completed business task

Explanation

Correct Answer

D:

Because feedback-driven thresholds rebalance accuracy, latency and spend, the outcome worth following is accuracy-adjusted cost per completed business task; the metric is tied to the reported issue—router saves tokens but sends difficult tasks to weak models—for the scenario where real task mix differs from prelaunch performance tests.

Incorrect Answers

A:

For the workflow where some cases need cross-document reasoning and detailed justification, high-complexity case pass rate by model route is an actionable measure; it does not establish whether router saves tokens but sends difficult tasks to weak models in the situation under review.

B:

For the workflow where a workflow needs proprietary scheduling, model choice and runtime control, unsupported-host requirements met by the custom design is an actionable measure; it does not establish whether router saves tokens but sends difficult tasks to weak models in the situation under review.

C:

The alternative tracks wrong-entity retrieval rate in curated evaluation cases, arising from the mechanism where canonical keys and deduplication prevent incorrect cross-record joins; its movement alone says little about the failure that router saves tokens but sends difficult tasks to weak models.

Question 144

Security and business owners want to keep this capability: Customers ask for current delivery or payment status. They must also prevent the following: An incorrect customer identifier might retrieve or change another customer's order. Which guardrail should the design incorporate?

  1. Scope connector permissions to specific business operations
  2. Limit each specialist to its own permissions and tools
  3. Validate identifiers and forbid cross-customer lookups
  4. Narrow allowed tool capabilities and verify server-side authorization
  5. Suppress assertions that require evidence when no source was retrieved

Explanation

Correct Answer

C:

In the scenario where customers ask for current delivery or payment status, the explicit risk is that an incorrect customer identifier might retrieve or change another customer's order; enforcement must follow the mechanism by which authorized actions query the system of record at request time.

Incorrect Answers

A:

Another process could need this guardrail because an extension could receive broad write permissions beyond its single new action; the design being assessed supports customers ask for current delivery or payment status and has the explicit exposure that an incorrect customer identifier might retrieve or change another customer's order.

B:

Another process could need this guardrail because a specialized agent could receive broader tools than its assigned business role; the design being assessed supports customers ask for current delivery or payment status and has the explicit exposure that an incorrect customer identifier might retrieve or change another customer's order.

D:

The alternate guard addresses a shared MCP server might honor an unauthorized tool invocation for a Copilot Studio agent needs a reusable interface to governed operations; the present threat, an incorrect customer identifier might retrieve or change another customer's order, needs a different enforcement boundary.

E:

The alternate guard addresses a retrieval outage could make the model invent unsupported factual claims for required knowledge stores are intermittently unreachable; the present threat, an incorrect customer identifier might retrieve or change another customer's order, needs a different enforcement boundary.

Question 145

One enterprise rollout has 2 separately testable requirements. (1) Business use: Agent instructions mention many APIs but only two are permitted. Existing shortcoming: The agent can invoke an unintended high-privilege operation. (2) Business use: The agent may refund large amounts or deactivate accounts. Existing shortcoming: The conversation alone authorizes a permanent business change. Which 2 success metrics directly satisfy these stated cases? Choose TWO.

  1. Error-rate delta between canary and control cohorts
  2. Irreversible changes with verified approver attribution
  3. Schema-valid response rate
  4. Unauthorized tool invocation success rate

Explanation

Correct Answers

B:

The outcome irreversible changes with verified approver attribution measures whether the workflow where the agent may refund large amounts or deactivate accounts is improving; the scorecard separately needs unauthorized tool invocation success rate.

D:

The outcome unauthorized tool invocation success rate measures whether the workflow where agent instructions mention many APIs but only two are permitted is improving; the scorecard separately needs irreversible changes with verified approver attribution.

Incorrect Answers

A:

The alternative measures error-rate delta between canary and control cohorts for the situation where new behavior must be evaluated without exposing the full workforce; it cannot substitute for performance outcomes in the separate named cases: agent instructions mention many APIs but only two are permitted; the agent may refund large amounts or deactivate accounts.

C:

The alternative measures schema-valid response rate for the situation where agents return inconsistent fields for downstream integrations; it cannot substitute for performance outcomes in the separate named cases: agent instructions mention many APIs but only two are permitted; the agent may refund large amounts or deactivate accounts.

Question 146

A business stakeholder describes this operating situation: A workflow needs proprietary scheduling, model choice and runtime control. Testing instead shows that a predefined host cannot meet the required orchestration contract. What design approach is most appropriate?

  1. Use MCP for an agent's access to external tools and data
  2. Route simple classification requests to a validated inexpensive model
  3. Replace an unsuitable prebuilt design with a custom-engine agent
  4. Adopt a supported prebuilt agent for standard productivity workflows

Explanation

Correct Answer

C:

A workflow needs proprietary scheduling, model choice and runtime control; its failing behavior is that a predefined host cannot meet the required orchestration contract; applying the underlying design principle—owning the engine provides control over models, routing and execution—connects the requirement to an implementable control.

Incorrect Answers

A:

For another type of architecture, MCP presents supported tools and resources through a common integration contract is useful; this design must instead support a workflow needs proprietary scheduling, model choice and runtime control despite the problem where a predefined host cannot meet the required orchestration contract.

B:

There is a valid place for the mechanism that task-aware routing chooses economical models after quality gates, namely many calls require only a small fixed-label output; it does not remove the specific defect that a predefined host cannot meet the required orchestration contract from this case.

D:

For another type of architecture, a prebuilt Microsoft 365 capability minimizes custom maintenance for commodity tasks is useful; this design must instead support a workflow needs proprietary scheduling, model choice and runtime control despite the problem where a predefined host cannot meet the required orchestration contract.

Question 147

Troubleshooting starts from a business complaint: The agent is exposed to repeated jailbreak patterns. The affected process is: Hostile instructions arrive both as user input and retrieved documents. Which investigation is most directly tied to the reported defect?

  1. Compare affected cohorts and failure types before and after release
  2. Inspect guardrail detection and filtering annotations
  3. Analyze failures by paraphrase group and boundary type
  4. Review blocked-output annotations and access attempts

Explanation

Correct Answer

B:

The investigation directly tests the causal connection: input and document-attack detection supplies layered protection; it collects evidence for a workflow where hostile instructions arrive both as user input and retrieved documents but the outcome is instead that the agent is exposed to repeated jailbreak patterns.

Incorrect Answers

A:

The evidence this method produces applies to a workflow where fixes are marked complete when a prompt is deployed; it cannot by itself establish why the agent is exposed to repeated jailbreak patterns in the current scenario.

C:

Slightly different wording breaks correct entity selection; examining that other path does not determine whether input and document-attack detection supplies layered protection, which is what distinguishes the issue in the stated use case.

D:

For the failure that a client can request arbitrary portions of training examples, reviewing blocked-output annotations and access attempts is relevant; this question instead reports that the agent is exposed to repeated jailbreak patterns, so different evidence is needed.

Question 148

An agent rollout is intended for this situation: A new action version changes how cases are updated. The known problem is that rolling back the agent leaves an incompatible backend action. Which success measure connects to that specific outcome?

  1. Unauthorized cross-environment requests blocked
  2. Recovery time for a failed connector-action deployment
  3. Unauthorized production customization events
  4. Production agent components using approved workload identity

Explanation

Correct Answer

B:

Recovery time for a failed connector-action deployment; this outcome measure belongs to the workflow where a new action version changes how cases are updated; a change in it can help detect whether the defect that rolling back the agent leaves an incompatible backend action is resolved.

Incorrect Answers

A:

Unauthorized cross-environment requests blocked; that indicator describes a distinct workflow where test automation must not reach live customer records; it does not quantify recovery time for a failed connector-action deployment, which is needed for the stated case.

C:

For the situation where production components should not be edited casually, unauthorized production customization events is a sensible indicator; the problem that rolling back the agent leaves an incompatible backend action requires measurement of recovery time for a failed connector-action deployment.

D:

Although production agent components using approved workload identity is observable, the affected process involves a new action version changes how cases are updated; its acceptance outcome is recovery time for a failed connector-action deployment, not the alternative measurement.

Question 149

A release must prove that read-only knowledge alone cannot create a service request will not recur. The application is used in this situation: A Teams user should query and update a supported external service. Which acceptance test most directly checks that failure mode?

  1. Create a test request and verify backend audit
  2. Create an escalation from a test case and inspect audit records
  3. Change a permitted credit field and verify updated authorized retrieval
  4. Query synonyms and near-neighbor phrases and check intended classification

Explanation

Correct Answer

A:

The check is targeted rather than generic: a plugin provides authorized callable tools to an agent; its pass/fail evidence applies to the workflow where a Teams user should query and update a supported external service and to the prior defect that read-only knowledge alone cannot create a service request.

Incorrect Answers

B:

This check exercises a different system behavior: an authorized action creates a governed workflow item from the current case; a pass there cannot establish that read-only knowledge alone cannot create a service request no longer occurs in the required workflow.

C:

Although the exercise might catch Copilot relies on last quarter's manually imported account report, it would not directly validate the mechanism that a plugin provides authorized callable tools to an agent in the target implementation.

D:

The tested behavior is that business-term definitions align user vocabulary with governed service entities; the necessary evidence here concerns a plugin provides authorized callable tools to an agent and the previously observed read-only knowledge alone cannot create a service request.

Question 150

A design adds autonomous behavior to a process where two customers store similarly named support incidents. This creates a documented risk: Search might mix tenant records because isolation relies only on prompt wording. Which enforced control is the best fit?

  1. Never rely on prompt text alone for tenancy isolation
  2. Revoke shared credentials and forbid cross-stage access
  3. Retain audit records of source replacements and withdrawals
  4. Keep enough prior approved artifacts for recovery

Explanation

Correct Answer

A:

The business setting is two customers store similarly named support incidents, with the exposure that search might mix tenant records because isolation relies only on prompt wording; the enforceable protection must correspond to the path where partitioning and authorized filters constrain relevant candidates.

Incorrect Answers

B:

This restriction is associated with test automation must not reach live customer records and addresses the exposure that staging credentials might allow writes to production business systems; it does not specifically protect the target workflow against search might mix tenant records because isolation relies only on prompt wording.

C:

Another process could need this guardrail because a disputed answer might be untraceable to its source and index revision; the design being assessed supports two customers store similarly named support incidents and has the explicit exposure that search might mix tenant records because isolation relies only on prompt wording.

D:

A preprocessing release changes category labels incorrectly; its guardrail addresses the possibility that a rollback might restore a model but not its required preprocessing artifacts; the present architecture must constrain the different threat that search might mix tenant records because isolation relies only on prompt wording.

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