Microsoft AB-100 (Agentic AI Business Solutions Architect) Practice Test 64

Question 630

A proposed deployment supports this business case: Users request a status check and an authorized follow-up action. Reviewers also warn that a generative planner might chain tools that exceed the approved workflow scope. Which restriction mitigates that concrete risk?

  1. Stop and compensate when a required approval fails
  2. Limit tool availability and confirmations for risky steps
  3. Ensure the backend rejects repeat writes for the same intent
  4. Require explicit consent for consequential create or update operations
  5. Give the agent only the required return operation

Explanation

Correct Answer

B:

When a generative planner might chain tools that exceed the approved workflow scope, the use case users request a status check and an authorized follow-up action requires a guardrail aligned with the behavior that the planner can select supported knowledge and tools for compound intents.

Incorrect Answers

A:

The alternate guard addresses a flow might notify the customer even though a required approval failed for approval processing requires sequential verification and notification; the present threat, a generative planner might chain tools that exceed the approved workflow scope, needs a different enforcement boundary.

C:

The proposed control targets an API retry might create a second purchase for the same user instruction, while the given risk scenario is a generative planner might chain tools that exceed the approved workflow scope; these exposures arise in different operational contexts.

D:

The proposed control targets a plugin might create business records without user consent or approval, while the given risk scenario is a generative planner might chain tools that exceed the approved workflow scope; these exposures arise in different operational contexts.

E:

That protection is tied to the mechanism that single-purpose tools and typed parameters bound the task; it does not prevent the distinct risk where a generative planner might chain tools that exceed the approved workflow scope in the system under review.

Question 631

A policy review covers an agent for which multiple AI teams consume a changing source API. Its most relevant exposure is that an API schema change could break consuming agents without migration notice. What safeguard should be required before release?

  1. Require review for breaking changes and publish migration windows
  2. Dead-letter failures and prevent duplicate financial postings
  3. Separate recommendation privileges from posting permissions
  4. Restrict production prompt edits to approved maintainers

Explanation

Correct Answer

A:

In the context where multiple AI teams consume a changing source API, the exposure is that an API schema change could break consuming agents without migration notice; containing it requires a guardrail in the operation where versioned schemas and compatibility policy make interfaces predictable.

Incorrect Answers

B:

Another process could need this guardrail because retried finance events might post the same invoice adjustment twice; the design being assessed supports multiple AI teams consume a changing source API and has the explicit exposure that an API schema change could break consuming agents without migration notice.

C:

The alternate guard addresses an AI recommendation could become an actual posting without an authorized reviewer for service, sales and finance jointly decide high-value concessions; the present threat, an API schema change could break consuming agents without migration notice, needs a different enforcement boundary.

D:

The alternate guard addresses a prompt editor could change production decisions without tracked approval for many teams copy slightly different instructions for the same task; the present threat, an API schema change could break consuming agents without migration notice, needs a different enforcement boundary.

Question 632

An investment review asks what measurable result would justify continuing an AI solution where model bills increase although request volume is stable. The original issue is: A large context prompt quietly grows every release. Which indicator should be tracked?

  1. Tokens consumed per verified completed task
  2. Ninety-fifth percentile end-to-end response latency
  3. Contract-valid predictions after custom-model updates
  4. Quality-adjusted cost per approved business result

Explanation

Correct Answer

A:

Tokens consumed per verified completed task; this outcome measure belongs to the workflow where model bills increase although request volume is stable; a change in it can help detect whether the defect that a large context prompt quietly grows every release is resolved.

Incorrect Answers

B:

For the workflow where a small portion of calls time out during peak demand, ninety-fifth percentile end-to-end response latency is an actionable measure; it does not establish whether a large context prompt quietly grows every release in the situation under review.

C:

Although contract-valid predictions after custom-model updates is observable, the affected process involves model bills increase although request volume is stable; its acceptance outcome is tokens consumed per verified completed task, not the alternative measurement.

D:

That scorecard concerns the activity in which average accuracy hides failures on complex finance questions and its result of quality-adjusted cost per approved business result; the reported shortcoming here is a large context prompt quietly grows every release.

Question 633

Before introducing a fix, the team must explain why one user's cached confidential answer appears for another user. The affected setting is: Shared result caching makes answers fast across roles. What should be inspected?

  1. Audit tokens and connection bindings for each stage
  2. Review retrieved snippets alongside final claims
  3. Compare answers with retrieved and deliberately missing documents
  4. Inspect cache keys and authorization on reused entries

Explanation

Correct Answer

D:

One user's cached confidential answer appears for another user; to attribute the defect rather than guess at a fix, inspecting cache keys and authorization on reused entries examines the dependency where cache keys and access validation prevent security-context leakage.

Incorrect Answers

A:

Test automation must not reach live customer records; its matching diagnostic check is reasonable for a shared connection identity grants sandbox agents production write access; the present investigation needs evidence for the different failure where one user's cached confidential answer appears for another user.

B:

For the failure that engineers tune prompts without checking whether evidence was retrieved, reviewing retrieved snippets alongside final claims is relevant; this question instead reports that one user's cached confidential answer appears for another user, so different evidence is needed.

C:

In the separate scenario where policy answers must be tied to a current source, the same investigation could explain the agent confidently answers when retrieval fails; it does not isolate the reported cause in a system where shared result caching makes answers fast across roles.

Question 634

A planned AI solution must support this situation: The classifier confuses complaints with refund requests. Its prototype fails because a generic instruction cannot express subtle category boundaries. Which design best fits the constraint?

  1. Implement an abstain-or-escalate response policy for uncertain outputs
  2. Use examples and counterexamples to clarify edge cases
  3. Version reusable prompts with owners and evaluation examples
  4. Use retrieval augmentation before tuning for changing facts

Explanation

Correct Answer

B:

Because representative examples teach decision boundaries without changing underlying facts, this approach corrects the reported breakdown where a generic instruction cannot express subtle category boundaries; the affected situation is that the classifier confuses complaints with refund requests.

Incorrect Answers

A:

When the system cannot always find evidence for regulated guidance, confidence and evidence checks route insufficiently supported cases to humans; in the stated workflow, however, a generic instruction cannot express subtle category boundaries; the relevant architectural behavior is that representative examples teach decision boundaries without changing underlying facts.

C:

The technical behavior of this alternative is that governed templates provide versioned contracts and test fixtures; the user case being assessed is the classifier confuses complaints with refund requests with a different failure: a generic instruction cannot express subtle category boundaries.

D:

The technical behavior of this alternative is that retrieval separates volatile knowledge from model parameters; the user case being assessed is the classifier confuses complaints with refund requests with a different failure: a generic instruction cannot express subtle category boundaries.

Question 635

Before introducing a fix, the team must explain why restoring code alone does not restore previous predictions. The affected setting is: A preprocessing release changes category labels incorrectly. What should be inspected?

  1. Compare deployed agent configuration to its approved manifest
  2. Match release manifest to transformation and model checksums
  3. Compare input distributions and labeled error trends
  4. Inspect artifacts for training data hashes and evaluation reports

Explanation

Correct Answer

B:

For the situation where a preprocessing release changes category labels incorrectly, the defect is that restoring code alone does not restore previous predictions; matching release manifest to transformation and model checksums helps verify the cause-and-effect path: versioned data transformations and model artifacts enable consistent rollback.

Incorrect Answers

A:

That probe evaluates a complete agent release manifest pins related runtime artifacts in a workflow where model and tool changes are promoted independently; the reported failure concerns the path for which versioned data transformations and model artifacts enable consistent rollback instead.

C:

Team retrains monthly despite unchanged model behavior; examining that other path does not determine whether versioned data transformations and model artifacts enable consistent rollback, which is what distinguishes the issue in the stated use case.

D:

In the process where data scientists cannot explain why a newer classifier behaves differently, the diagnostic procedure could explain models move to production without an audit of training inputs; this particular incident instead turns on whether versioned data transformations and model artifacts enable consistent rollback and why restoring code alone does not restore previous predictions.

Question 636

An incident review has one confirmed symptom: Engineers blame generation although the backend is slow. For the use case where agents regularly exceed an interactive response-time objective, which targeted diagnostic check would reveal the most pertinent evidence?

  1. Audit token use and secret inventory for agent services
  2. Inspect model and tool spans under one operation identifier
  3. Review critical-path durations for slow requests
  4. Compare requested export destinations against policy allowlists

Explanation

Correct Answer

C:

Where engineers blame generation although the backend is slow, the useful evidence comes from reviewing critical-path durations for slow requests; its focus matches the architectural mechanism that per-span latency isolates model work, network waits and tool execution in the process where agents regularly exceed an interactive response-time objective.

Incorrect Answers

A:

In the separate scenario where a backend agent must call an approved Azure service without embedded secrets, the same investigation could explain long-lived API keys leak from configuration and diagnostics; it does not isolate the reported cause in a system where agents regularly exceed an interactive response-time objective.

B:

The proposed probe concerns operations need to correlate prompts, model calls and tools in a Foundry workflow, where agent trace spans provide component-level execution evidence through connected telemetry; this incident instead requires evidence about per-span latency isolates model work, network waits and tool execution and the failure that engineers blame generation although the backend is slow.

D:

In the process where a prompt injection requests data export to an outside address, the diagnostic procedure could explain model-level refusal alone may not prevent a tool call; this particular incident instead turns on whether per-span latency isolates model work, network waits and tool execution and why engineers blame generation although the backend is slow.

Question 637

Security and business owners want to keep this capability: Two systems share company names but not primary keys. They must also prevent the following: A guessed cross-system account match might attach credit risk to the wrong customer. Which guardrail should the design incorporate?

  1. Honor source access when syncing or federating content
  2. Let authorized incident commanders override recommendations
  3. Reject or escalate uncertain cross-system matches
  4. Assign glossary approval to service process owners

Explanation

Correct Answer

C:

Two systems share company names but not primary keys; its risk scenario is that a guessed cross-system account match might attach credit risk to the wrong customer; the safeguard fits because a mapping service resolves stable identities before enrichment and the restriction addresses that specific exposure.

Incorrect Answers

A:

The alternate guard addresses an external connector might index documents without honoring original ACLs for workers need internal wiki articles from outside Microsoft 365; the present threat, a guessed cross-system account match might attach credit risk to the wrong customer, needs a different enforcement boundary.

B:

This restriction is associated with a specialized support team uses a nonstandard severity matrix and addresses the exposure that automated severity suggestions might override an incident commander's decision; it does not specifically protect the target workflow against a guessed cross-system account match might attach credit risk to the wrong customer.

D:

The proposed control targets unreviewed synonyms could redefine legally significant customer-service classifications, while the given risk scenario is a guessed cross-system account match might attach credit risk to the wrong customer; these exposures arise in different operational contexts.

Question 638

One enterprise rollout has 2 separately testable requirements. (1) Business use: Users need to distinguish verified fields from AI suggestions. Known defect: Staff mistake speculative recommendations for recorded facts. (2) Business use: Staff need a contextual workspace assembled around one customer case. Known defect: Generic chat loses visibility into related records and actions. Which 2 architectural choices directly satisfy these stated cases? Choose TWO.

  1. Build a code-first generative page with explicitly bound data
  2. Choose speech processing for transcribed contact-center audio
  3. Design clear provenance and confidence cues in generated interfaces
  4. Use finance and supply-chain in-app agent knowledge through Dataverse virtual entities

Explanation

Correct Answers

A:

For the independent requirement that staff need a contextual workspace assembled around one customer case, generated interface components can be grounded in governed app data and defined actions; it addresses the defect generic chat loses visibility into related records and actions, while users need to distinguish verified fields from AI suggestions needs its own architecture.

C:

For the independent requirement that users need to distinguish verified fields from AI suggestions, visual provenance separates authoritative records from generated advice; it addresses the defect staff mistake speculative recommendations for recorded facts, while staff need a contextual workspace assembled around one customer case needs its own architecture.

Incorrect Answers

B:

This design is associated with the different failure where manual notes omit important customer statements in a workflow where quality teams need searchable insights from recorded calls; it does not directly solve the listed conditions: staff mistake speculative recommendations for recorded facts; generic chat loses visibility into related records and actions.

D:

This design is associated with the different failure where a static knowledge upload gives stale purchase-order states in a workflow where finance users need current operational data in contextual help; it does not directly solve the listed conditions: staff mistake speculative recommendations for recorded facts; generic chat loses visibility into related records and actions.

Question 639

Before production promotion, the team needs evidence for this setting: A candidate appears accurate on its training examples. A prior regression caused the following: Overfitting hides failures on new business records. Which validation is most persuasive?

  1. Evaluate difficult queries against annotated source passages
  2. Install the agent to a pilot Microsoft 365 audience
  3. Score unseen cases representing each critical segment
  4. Change page layout in a test environment and observe recovery
  5. Test calls with background noise and varied accents

Explanation

Correct Answer

C:

Scoring unseen cases representing each critical segment is aligned with the failure that overfitting hides failures on new business records; its causal relevance is that separation of training and evaluation yields realistic generalization checks for the situation where a candidate appears accurate on its training examples.

Incorrect Answers

A:

The verification concerning stakeholders report hallucinations but cannot isolate search errors checks labeled queries and expected evidence distinguish retrieval from synthesis against generator tuning fails because the required documents were never retrieved; its result says little about the observed issue where overfitting hides failures on new business records.

B:

For employees want curated assistance without owning a separate LLM runtime, the alternate test is about whether declarative agents define instructions, knowledge and actions in the supported host; it cannot prove the outcome needed after the reported regression where overfitting hides failures on new business records.

D:

The tested behavior is that uI automation can operate application interfaces but requires stability checks; the necessary evidence here concerns separation of training and evaluation yields realistic generalization checks and the previously observed overfitting hides failures on new business records.

E:

In the alternate setting where quality teams need searchable insights from recorded calls, the test exercises speech transcription transforms audio to text for subsequent analysis; even a pass would not show that overfitting hides failures on new business records has been prevented.

Question 640

A test plan focuses on a previously missed failure: The agent follows lower-trust text as an order. It occurred when a retrieved email contains instructions to bypass approval. Which check directly validates that behavior?

  1. Try injection, invalid inputs and out-of-scope actions
  2. Block the search endpoint during a policy question
  3. Try a forbidden refund despite a permissive model instruction
  4. Insert override strings into a test retrieved document

Explanation

Correct Answer

D:

Inserting override strings into a test retrieved document is aligned with the failure that the agent follows lower-trust text as an order; its causal relevance is that adversarial fixtures verify separation of sources and instructions for the situation where a retrieved email contains instructions to bypass approval.

Incorrect Answers

A:

The verification concerning benchmarks cover only cooperative user wording checks negative tests expose policy and validation boundaries against a malicious instruction triggers a privileged tool; its result says little about the observed issue where the agent follows lower-trust text as an order.

B:

Blocking the search endpoint during a policy question verifies an adjacent objective where knowledge retrieval is temporarily unavailable; it leaves untested the causal path in which adversarial fixtures verify separation of sources and instructions for this case.

C:

Although the exercise might catch API trusts the model's text assertion of permission, it would not directly validate the mechanism that adversarial fixtures verify separation of sources and instructions in the target implementation.

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