Microsoft AB-100 Practice Test 81

Question 800

A team builds a Foundry agent that can answer manuals and initiate work orders. The model test passes, but no tool-authorization test exists. What is the appropriate release action?

Disable trace collection so test failures do not block launch.
Hold production promotion until end-to-end agent tool permission and workflow tests pass.
Promote because a Foundry model deployment already succeeded.
Grant broad work-order permissions to remove the testing gap.

Explanation

Correct Answer

Model language quality is only one control; business action safety is a distinct release requirement.

Incorrect Answers

Hiding diagnostics cannot turn an untested workflow into an acceptable release.

Model deployment does not establish correct tool use or authorized work orders.

Overprivilege increases risk and does not validate caller scope.

Question 801

A Foundry agent regression appears after a new prompt, tool definition and model deployment are released together. Which artifact best supports diagnosis?

Only the last developer who opened the agent workspace.
An average token-cost report without version identifiers.
An immutable manifest linking prompt, model deployment, tool schema, sources and evaluation results.
The deployment’s friendly display name only.

Explanation

Correct Answer

Agent behavior depends on the combination of code, instructions, tools and grounding data.

Incorrect Answers

Editor activity alone cannot reproduce tool and model dependencies.

Cost metrics are not a sufficient record of the configuration that caused a failure.

A display name does not identify the exact components served in that release.

Question 802

A new Foundry agent release changes how supplier discounts are proposed. Leadership fears an unexpected pricing error. What rollout strategy minimizes blast radius?

Switch every sales user at once to maximize evaluation volume.
Observe only whether the chat UI loads.
Change pricing policy whenever the new model output differs.
Canary the version to an approved limited cohort with monitored quote correctness and rollback criteria.

Explanation

Correct Answer

A controlled exposure detects material pricing regressions before broad production use.

Incorrect Answers

A simultaneous full release exposes all customers to the same potential defect.

A working interface does not prove discount amounts or approvals are correct.

Model suggestions must not redefine approved business policy to fit themselves.

Question 803

An agent uses a model alias whose backing deployment changes during a vendor update. Support cannot identify which model served an incorrect answer. What lifecycle control is needed?

Log the resolved model deployment/version per invocation and pin evaluated release routing where required.
Use the alias name as proof all responses came from the original build.
Ask the user to remember the exact model version from the answer style.
Clear model-routing traces after each rollout to simplify storage.

Explanation

Correct Answer

An alias alone is insufficient provenance when its underlying model can change.

Incorrect Answers

An alias can be rebound without changing its textual name.

Writing style cannot establish a trusted deployment identifier.

Removing provenance prevents regression comparison and audit reconstruction.

Question 804

A new Foundry agent tool handler fails after some orders enter pending state. What rollback behavior preserves customer transactions?

Treat the agent’s prior acceptance message as proof all orders finished.
Restore the previous compatible agent release while reconciling pending order IDs with the external service.
Retry every order under a new identifier regardless of actual status.
Delete pending records without consulting the order service.

Explanation

Correct Answer

Agent version rollback must not double-submit or abandon durable tasks already begun.

Incorrect Answers

Generated acknowledgement is not authoritative external commit evidence.

New identifiers defeat deduplication and may create extra supplier purchases.

Blind deletion can conceal actual committed transactions.

Question 805

An agent update adds an inventory lookup tool hosted under a broad service account. It returns stock for all regional warehouses to any user. Which release test is necessary?

Remove region names from the model output while returning all rows.
Grant every agent user the service account’s broad role.
Negative authorization tests confirming regional users cannot retrieve warehouse records outside their assigned scope.
Check only that the tool responds to an administrator.

Explanation

Correct Answer

Adding a tool creates a new backend data-access boundary that model tests do not automatically secure.

Incorrect Answers

Output hiding is not a substitute for server-side scope filtering.

Expanding access defeats least privilege and multiplies leakage risk.

Admin success says nothing about restrictions for normal employees.

Question 806

A Foundry research agent performs five parallel queries to a supplier API with a strict per-second quota. Production sees throttling. What release capability needs validation?

Increase parallel requests until the rate limit is ignored.
Generate missing supplier responses from the agent’s prior memory.
Mark all throttled tool calls successful because requests were transmitted.
Concurrency control, bounded exponential backoff and truthful partial-result handling under API quota limits.

Explanation

Correct Answer

Tool orchestration must respect the downstream service’s capacity and report unavailable information accurately.

Incorrect Answers

More simultaneous calls intensify 429 throttling and may lead to bans.

Predicted values cannot substitute for current authoritative supplier data.

An attempted HTTP call is not a completed data retrieval.

Question 807

An agent passes a model benchmark while using an old knowledge index. A release replaces both model and index. What must the final acceptance test include?

End-to-end grounded questions against the new indexed revisions and current model/tool combination.
Use the previous index’s citation results as proof the new one works.
Skip retrieval tests because model weights were evaluated.
Promote if the agent generates a friendly greeting.

Explanation

Correct Answer

Model-only evaluation does not confirm source authority or retrieval behavior after index changes.

Incorrect Answers

New chunk identifiers or filters can invalidate earlier references.

Grounding is a separate dependency and can fail despite a strong model.

Greeting text does not validate knowledge retrieval or source effectivity.

Question 808

A customer-classification model is tuned monthly, but the training team does not record which labeled records were used. What must ALM preserve?

Allow each team to overwrite labels after the release without history.
Dataset version, label provenance, preprocessing parameters, training configuration and evaluation artifact.
Keep only the final model’s friendly name.
Use a screenshot of model accuracy as the dataset manifest.

Explanation

Correct Answer

Model reproducibility and authorized data use depend on traceable inputs and transformation decisions.

Incorrect Answers

Untracked corrections make model performance and audit reproduction impossible.

Names cannot reconstruct the training examples or rules that produced the artifact.

A chart does not identify source record versions or data permissions.

Question 809

A tuned model improves average classification but loses 12 points of recall on high-risk fraud cases. What release policy should decide?

Exclude fraud cases from evaluation to prevent a red metric.
Change fraud labels to match the new predictions.
Block or remediate release under the fraud-class acceptance threshold despite the improved average.
Promote because the overall percentage increased.

Explanation

Correct Answer

Risk-specific errors can outweigh gains in frequent low-impact classes.

Incorrect Answers

Removing critical cases invalidates rather than improves the evidence.

Ground truth should not be rewritten to justify a model artifact.

Aggregate performance can hide dangerous false negatives.

Question 810

A model registry alias `current` points to one artifact in test and another in production, but the promotion report lists only the alias. How should provenance be recorded?

Assume matching alias names imply matching model bytes.
Use only the latest deployment timestamp.
Compare output tone to infer that both environments use identical artifacts.
Record immutable model artifact hashes/versions and environment-to-artifact bindings.

Explanation

Correct Answer

Mutable aliases cannot prove that the tested artifact is the one serving real customers.

Incorrect Answers

Environment-local alias pointers can resolve to different models.

Time alone does not establish which artifact was deployed.

Generated wording is not cryptographic or registry evidence of model identity.

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

Microsoft AB-100 Practice Test 79 - 10 Questions (Last Reviewed 23-Oct-2026)
Microsoft AB-100 Practice Test 80 - 10 Questions (Last Reviewed 22-Oct-2026)
Microsoft AB-100 Practice Test 81 - 10 Questions (Last Reviewed 21-Oct-2026)
Microsoft AB-100 Practice Test 82 - 10 Questions (Last Reviewed 20-Oct-2026)
Microsoft AB-100 Practice Test 83 - 10 Questions (Last Reviewed 19-Oct-2026)

Pass Your Microsoft AB-100 Exam Easy!

Question 800

Explanation

Correct Answer

Incorrect Answers

Question 801

Explanation

Correct Answer

Incorrect Answers

Question 802

Explanation

Correct Answer

Incorrect Answers

Question 803

Explanation

Correct Answer

Incorrect Answers

Question 804

Explanation

Correct Answer

Incorrect Answers

Question 805

Explanation

Correct Answer

Incorrect Answers

Question 806

Explanation

Correct Answer

Incorrect Answers

Question 807

Explanation

Correct Answer

Incorrect Answers

Question 808

Explanation

Correct Answer

Incorrect Answers

Question 809

Explanation

Correct Answer

Incorrect Answers

Question 810

Explanation

Correct Answer

Incorrect Answers

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