Microsoft AB-100 (Agentic AI Business Solutions Architect) Practice Test 47

Question 460

An operational risk could arise because a background agent might exceed spending authority by editing its own prompt. The AI system supports the business need that agent can research options but cannot approve a vendor payment. Which guardrail is designed for this threat?

  1. Validate types and authorization at the called service
  2. Store policy thresholds outside model-editable instructions
  3. Require least-privilege service permissions and action confirmation
  4. Do not auto-save an uncertain prediction as authoritative data

Explanation

Correct Answer

B:

In the context where agent can research options but cannot approve a vendor payment, the exposure is that a background agent might exceed spending authority by editing its own prompt; containing it requires a guardrail in the operation where approval checkpoints split independent analysis from high-impact execution.

Incorrect Answers

A:

Another process could need this guardrail because an API might accept a valid-looking but unauthorized generated identifier; the design being assessed supports agent can research options but cannot approve a vendor payment and has the explicit exposure that a background agent might exceed spending authority by editing its own prompt.

C:

The alternate boundary mitigates a tool might update tickets using a shared identity with too many permissions, which is relevant to an agent needs to update a governed ticket status; it leaves the current workflow vulnerable to a background agent might exceed spending authority by editing its own prompt.

D:

Staff reject suggestions because approval controls are unclear; its guardrail addresses the possibility that an uncertain generated field might silently become authoritative data; the present architecture must constrain the different threat that a background agent might exceed spending authority by editing its own prompt.

Question 461

The architect wants evidence before altering the deployment. In a workflow where the system cannot always find evidence for regulated guidance, an agent guesses a policy instead of escalating. What should the team examine first?

  1. Inspect parse failures against missing schema instructions
  2. Compare feature gaps and vendor operating obligations
  3. Compare low-evidence answers with escalation logs
  4. Examine tool calls by transaction impact and approval requirement

Explanation

Correct Answer

C:

Where an agent guesses a policy instead of escalating, the useful evidence comes from comparing low-evidence answers with escalation logs; its focus matches the architectural mechanism that confidence and evidence checks route insufficiently supported cases to humans in the process where the system cannot always find evidence for regulated guidance.

Incorrect Answers

A:

Narrative answers cannot be parsed by consuming workflows; examining that other path does not determine whether confidence and evidence checks route insufficiently supported cases to humans, which is what distinguishes the issue in the stated use case.

B:

Comparing feature gaps and vendor operating obligations explores a separate connection, where managed functionality shifts routine platform operations to the service; the diagnosed requirement is the system cannot always find evidence for regulated guidance, and the actual defect is an agent guesses a policy instead of escalating.

D:

For the failure that the same agent can read an invoice and commit a payment unchecked, examining tool calls by transaction impact and approval requirement is relevant; this question instead reports that an agent guesses a policy instead of escalating, so different evidence is needed.

Question 462

Implementation alternatives must be compared for a workflow where an updated agent performs differently months after release. The pilot shows this gap: Old scores cannot be reproduced because prompts and tests changed. Which alternative should lead the solution?

  1. Track grounding-source updates and index rebuild provenance
  2. Version agent instructions, tool definitions and model references together
  3. Treat index schema and embeddings as versioned artifacts
  4. Preserve reproducible evaluation evidence for major model changes

Explanation

Correct Answer

D:

For the need that an updated agent performs differently months after release, the relevant mechanism is that versioned evaluation sets, model IDs and thresholds document release claims; that architectural dependency addresses the observed failure where old scores cannot be reproduced because prompts and tests changed.

Incorrect Answers

A:

No log identifies which source version was indexed; the alternate design can address that situation because content lineage records update, transformation and index revision, whereas this question needs a solution to old scores cannot be reproduced because prompts and tests changed.

B:

There is a valid place for the mechanism that a complete agent release manifest pins related runtime artifacts, namely model and tool changes are promoted independently; it does not remove the specific defect that old scores cannot be reproduced because prompts and tests changed from this case.

C:

When a search index changes embedding model and chunking approach, explicit index versions support compatible retrieval and rollback; in the stated workflow, however, old scores cannot be reproduced because prompts and tests changed; the relevant architectural behavior is that versioned evaluation sets, model IDs and thresholds document release claims.

Question 463

Before production promotion, the team needs evidence for this setting: The model submits malformed identifiers to an operational API. A prior regression caused the following: Invalid record IDs produce updates to the wrong business object. Which validation is most persuasive?

  1. Attempt concurrent updates through both app interfaces
  2. Send invalid IDs, oversized values and out-of-range quantities
  3. Prototype the distinguishing behavior under production constraints
  4. Map each proposed use case to a baseline metric and named owner

Explanation

Correct Answer

B:

The scenario is that the model submits malformed identifiers to an operational API; the defect to rule out is that invalid record IDs produce updates to the wrong business object; testing via sending invalid IDs, oversized values and out-of-range quantities challenges the design dependency where typed inputs and server validation reject malformed or out-of-scope calls.

Incorrect Answers

A:

For two applications try to overwrite the same customer credit field, the alternate test is about whether field ownership and integration contracts prevent competing writes; it cannot prove the outcome needed after the reported regression where invalid record IDs produce updates to the wrong business object.

C:

The suggested test serves competitive requirements depend on unique multi-model control, in which a boxed product cannot express the execution and ownership rules; the present scenario instead concerns the reported problem where invalid record IDs produce updates to the wrong business object.

D:

Mapping each proposed use case to a baseline metric and named owner verifies an adjacent objective where executives request AI before defining an operational improvement target; it leaves untested the causal path in which typed inputs and server validation reject malformed or out-of-scope calls for this case.

Question 464

One enterprise rollout has 2 separately testable requirements. (1) Business use: A Teams user should query and update a supported external service. Exposure: A plugin might create business records without user consent or approval. (2) Business use: Employees want curated assistance without owning a separate LLM runtime. Exposure: A declarative agent might be published beyond permitted tenant groups. Which 2 safeguards directly satisfy these stated cases? Choose TWO.

  1. Require explicit consent for consequential create or update operations
  2. Apply tenant policies and scoped content permissions
  3. Validate types and authorization at the called service
  4. Reject generated fields outside the approved schema

Explanation

Correct Answers

A:

For the exposure that a plugin might create business records without user consent or approval, a plugin provides authorized callable tools to an agent explains the protection required when a Teams user should query and update a supported external service; this does not replace control of a declarative agent might be published beyond permitted tenant groups.

B:

For the exposure that a declarative agent might be published beyond permitted tenant groups, declarative agents define instructions, knowledge and actions in the supported host explains the protection required when employees want curated assistance without owning a separate LLM runtime; this does not replace control of a plugin might create business records without user consent or approval.

Incorrect Answers

C:

This security safeguard addresses the unrelated exposure that an API might accept a valid-looking but unauthorized generated identifier in a workflow where the orchestrator chooses between similar invoice and credit tools; the question instead names the specific risks a plugin might create business records without user consent or approval; a declarative agent might be published beyond permitted tenant groups; this choice does not directly mitigate those specific exposures.

D:

This security safeguard addresses the unrelated exposure that generated response keys might be trusted even when they violate the schema in a workflow where an agent must turn service notes into a fixed JSON summary; the question instead names the specific risks a plugin might create business records without user consent or approval; a declarative agent might be published beyond permitted tenant groups; this choice does not directly mitigate those specific exposures.

Question 465

During integration testing, the agent invents answers rather than saying evidence is missing. The associated business context is: Knowledge retrieval is temporarily unavailable. Which architectural remedy is the strongest match?

  1. Test tool-permission escalation attempts at the execution layer
  2. Test resilience when a grounding dependency fails
  3. Test negative and adversarial cases alongside happy paths
  4. Build risk-based test coverage for critical user journeys

Explanation

Correct Answer

B:

Without the intended behavior, the agent invents answers rather than saying evidence is missing; the proposed architecture establishes the specific design principle that failure injection checks controlled degradation and abstention for the process where knowledge retrieval is temporarily unavailable.

Incorrect Answers

A:

The model may try a tool outside the user's rights; the alternate design can address that situation because authorization evaluations verify server-side deny behavior, whereas this question needs a solution to the agent invents answers rather than saying evidence is missing.

C:

A malicious instruction triggers a privileged tool; the alternate design can address that situation because negative tests expose policy and validation boundaries, whereas this question needs a solution to the agent invents answers rather than saying evidence is missing.

D:

The test suite spends most cases on low-risk FAQs; the alternate design can address that situation because risk-weighting prioritizes high-impact actions and failure paths, whereas this question needs a solution to the agent invents answers rather than saying evidence is missing.

Question 466

The architect wants evidence before altering the deployment. In a workflow where benchmarks cover only cooperative user wording, a malicious instruction triggers a privileged tool. What should the team examine first?

  1. Inspect pending runs, deadlines and compensation events
  2. Compare denial logs and executed actions for attack samples
  3. Map test IDs back to resolved incident categories
  4. Inspect tool invocations following injected content

Explanation

Correct Answer

B:

Where a malicious instruction triggers a privileged tool, the useful evidence comes from comparing denial logs and executed actions for attack samples; its focus matches the architectural mechanism that negative tests expose policy and validation boundaries in the process where benchmarks cover only cooperative user wording.

Incorrect Answers

A:

In the process where an autonomous agent relies on slow approvals and delayed tool results, the diagnostic procedure could explain timeout creates an abandoned workflow with no owner; this particular incident instead turns on whether negative tests expose policy and validation boundaries and why a malicious instruction triggers a privileged tool.

C:

In the process where support tickets reveal an unusual multistep conversation failure, the diagnostic procedure could explain the fixed scenario is not added to future test suites; this particular incident instead turns on whether negative tests expose policy and validation boundaries and why a malicious instruction triggers a privileged tool.

D:

In the process where a retrieved email contains instructions to bypass approval, the diagnostic procedure could explain the agent follows lower-trust text as an order; this particular incident instead turns on whether negative tests expose policy and validation boundaries and why a malicious instruction triggers a privileged tool.

Question 467

An enterprise team is making a platform design decision. Business context: A specialized support team uses a nonstandard severity matrix. Present failure: Default summaries treat critical outages as low priority. Which approach resolves the mismatch?

  1. Extend service assistance with an approved escalation action
  2. Design timeout and degraded-mode handling for Sales enrichments
  3. Customize service triage around business-specific severity criteria
  4. Version models and datasets together for controlled adaptation

Explanation

Correct Answer

C:

The defect is that default summaries treat critical outages as low priority; the expected business case is that a specialized support team uses a nonstandard severity matrix; the design fits because custom business rules align AI-assisted triage with actual impact definitions.

Incorrect Answers

A:

The alternative addresses AI suggestions are never converted into traceable escalation records by a design in which an authorized action creates a governed workflow item from the current case; it does not supply the missing relationship for a specialized support team uses a nonstandard severity matrix, where default summaries treat critical outages as low priority.

B:

There is a valid place for the mechanism that bounded timeouts and graceful fallback preserve essential user work, namely external credit service responds slowly during regional outages; it does not remove the specific defect that default summaries treat critical outages as low priority from this case.

D:

There is a valid place for the mechanism that dataset lineage and model versioning support repeatable assessment, namely retraining uses revised labels and a different source snapshot; it does not remove the specific defect that default summaries treat critical outages as low priority from this case.

Question 468

Security and business owners want to keep this capability: Service representatives draft responses from authorized knowledge. They must also prevent the following: AI-drafted customer responses might be sent before the responsible reviewer approves. Which guardrail should the design incorporate?

  1. Reject or escalate uncertain cross-system matches
  2. Require explicit consent for consequential create or update operations
  3. Require reviewer approval before outbound communication
  4. Enforce source-system entitlements in the integration
  5. Prevent writes or disclosure outside user permissions

Explanation

Correct Answer

C:

The business setting is service representatives draft responses from authorized knowledge, with the exposure that aI-drafted customer responses might be sent before the responsible reviewer approves; the enforceable protection must correspond to the path where contextual grounding provides relevant policy and customer facts.

Incorrect Answers

A:

The alternate guard addresses a guessed cross-system account match might attach credit risk to the wrong customer for two systems share company names but not primary keys; the present threat, aI-drafted customer responses might be sent before the responsible reviewer approves, needs a different enforcement boundary.

B:

A Teams user should query and update a supported external service; its guardrail addresses the possibility that a plugin might create business records without user consent or approval; the present architecture must constrain the different threat that aI-drafted customer responses might be sent before the responsible reviewer approves.

D:

Another process could need this guardrail because sales enrichment might query external records beyond the seller's entitled scope; the design being assessed supports service representatives draft responses from authorized knowledge and has the explicit exposure that aI-drafted customer responses might be sent before the responsible reviewer approves.

E:

The safeguard for finance users need current operational data in contextual help deals with a different threat: live finance records might become editable or visible to unauthorized user roles; even after it is enforced, the exposure that aI-drafted customer responses might be sent before the responsible reviewer approves remains possible in the target workflow where contextual grounding provides relevant policy and customer facts.

Question 469

An investment review asks what measurable result would justify continuing an AI solution where a model is replaced without a traceable change ticket. The original issue is: Investigators cannot explain when behavior changed. Which indicator should be tracked?

  1. Model changes with complete actor and approval records
  2. High-risk tool actions with full decision-to-effect trace
  3. True performance drift detected before critical SLO breach
  4. Ingested documents passing quality and access gates

Explanation

Correct Answer

A:

The observed shortcoming is that investigators cannot explain when behavior changed; measuring model changes with complete actor and approval records gives business owners a directly relevant indicator for the process where a model is replaced without a traceable change ticket.

Incorrect Answers

B:

High-risk tool actions with full decision-to-effect trace; that indicator describes a distinct workflow where a tool deletes a record following an ambiguous conversation; it does not quantify model changes with complete actor and approval records, which is needed for the stated case.

C:

A change in true performance drift detected before critical SLO breach is informative for customer terminology changes over time; it does not establish whether the current process achieves model changes with complete actor and approval records.

D:

For the workflow where new documents enter a production source automatically, ingested documents passing quality and access gates is an actionable measure; it does not establish whether investigators cannot explain when behavior changed in the situation under review.

Question 470

During integration testing, a trained model remembers obsolete guidance after source changes. The associated business context is: Answers depend on policies revised every week. Which architectural remedy is the strongest match?

  1. Extend an existing Dynamics or Copilot capability for a small gap
  2. Use a custom model only when general models fail a defined task
  3. Use Dataverse knowledge for structured business-record questions
  4. Use retrieval augmentation before tuning for changing facts

Explanation

Correct Answer

D:

Because retrieval separates volatile knowledge from model parameters, this approach corrects the reported breakdown where a trained model remembers obsolete guidance after source changes; the affected situation is that answers depend on policies revised every week.

Incorrect Answers

A:

For current users need one governed custom action inside their workflow, this other approach relies on the fact that extension retains familiar user experience and adds bounded custom behavior; the issue here is a trained model remembers obsolete guidance after source changes; that mechanism does not establish the required retrieval separates volatile knowledge from model parameters.

B:

The technical behavior of this alternative is that a measured capability gap, appropriate dataset and MLOps plan justify customization; the user case being assessed is answers depend on policies revised every week with a different failure: a trained model remembers obsolete guidance after source changes.

C:

Dataverse grounding exposes appropriate structured records to the agent; its main purpose is users ask about governed CRM records rather than prose manuals, not correction of the reported defect where a trained model remembers obsolete guidance after source changes in the target process.

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