Microsoft AB-100 Practice Test 54

Question 530

A finance team wants to add a supplier-risk summary. It can extend a supported in-app agent or replace the entire process with a custom portal. What must the economic evaluation include?

Only the custom model’s per-token price because all other application functions are already implemented.
The number of prompt templates, without comparing user outcomes or operating cost.
Assume a product replacement automatically inherits every role and audit configuration from the prior application.
Training, identity integration, data migration, downtime, ongoing support and the value of any unsupported requirement.

Explanation

Correct Answer

A replacement has organizational and operational transition costs beyond its API or model price.

Incorrect Answers

A new portal still needs identity, audit, maintenance and business process integration.

Template counts do not show the real costs of either maintaining or migrating the process.

Access and evidence controls must be deliberately migrated and revalidated; identical names do not prove parity.

Question 531

A human-resources assistant answers accurately about last year’s policy but repeatedly cites a superseded version. The model’s reasoning on the retrieved text is correct. What is the economically sensible first fix?

Repair the approved-source and effective-date retrieval filters, then reevaluate before custom training.
Train a custom model on all historical policy PDFs including the superseded versions.
Increase generation temperature so the assistant is more willing to select newer wording.
Remove citations from answers to make conflicts between policy versions less visible.

Explanation

Correct Answer

The failure comes from stale knowledge selection, so fine-tuning may add expense without correcting the governing source.

Incorrect Answers

More obsolete training content could increase reliance on policy wording that is no longer binding.

Sampling randomness cannot establish publication authority or effective dates.

Suppressing provenance hides the underlying correctness defect and defeats the requested auditability.

Question 532

A warehouse assistant frequently gives wrong stock counts but its replies use the correct product language. Inventory changes every minute, and the model reads a nightly vector index. What should be prioritized?

Treat the model’s confidence about quantity as sufficient to release inventory reservations.
Integrate authorized live inventory lookup with a freshness policy rather than training model weights on changing quantities.
Fine-tune on yesterday’s stock tables so the model memorizes the quantities more strongly.
Expand the prompt with older nightly exports and average their stock figures.

Explanation

Correct Answer

High-frequency operational facts belong in a current transactional source; tuning cannot keep weights synchronized by the minute.

Incorrect Answers

Probabilistic language does not create or verify current physical stock allocation.

The learned values will already be obsolete as soon as new allocations and receipts post.

Averaging stale states does not equal live available inventory for a particular warehouse.

Question 533

An insurance document agent uses correct current source files but still confuses two visually similar regulated form classes. Prompt and OCR changes have been tested. When is custom tuning defensible?

When the model supplier offers free GPU capacity without evaluation labels.
When a better model produces longer descriptions even though class assignments remain wrong.
When a labeled independent evaluation shows a persistent material classification gap and the tuning data is governed.
Whenever one executive dislikes the assistant’s preferred wording.

Explanation

Correct Answer

A verified task-specific failure after source repairs supports comparing a custom model’s benefit and maintenance cost.

Incorrect Answers

Compute availability cannot demonstrate improved categorization or correct legal-form identification.

Output verbosity does not establish the actual document class needed by the processing workflow.

Style preference is not a measured technical classification failure or a model-training investment case.

Question 534

A compliance assistant changes its policy knowledge every week. A team proposes monthly retraining of a custom model to keep up, while an approved searchable policy library already exists. What should a TCO comparison include?

Assume training the custom model once makes weekly policy changes free.
Ignore ongoing index and access-control costs when modeling the retrieval alternative.
Judge the two designs solely by the length of a typical generated response.
Content refresh, search/index maintenance and citations versus labeling, retraining, evaluation and model redeployment.

Explanation

Correct Answer

Rapidly changing policies impose recurring lifecycle costs that may favor governed retrieval over repeated weight updates.

Incorrect Answers

A fixed model does not automatically acquire new published policy revisions.

Retrieval still needs source governance, refresh and monitoring; its costs should be compared fairly.

Response length does not describe the model-refresh, evidence or policy-maintenance work that drives ownership cost.

Question 535

An AI router sends routine order-status questions to a small model. A customer asks for an interpretation of an ambiguous indemnity clause. What routing policy is appropriate?

Use risk- and complexity-based routing to an approved capable model or qualified reviewer, preserving source and cost evidence.
Force the cheap model to answer every case so token cost stays identical.
Choose the expensive model based only on the length of the customer’s email.
Send clause text directly to an unrestricted external endpoint without validating permitted data transfer.

Explanation

Correct Answer

The exceptional legal interpretation has different accuracy and authorization needs than a routine status lookup.

Incorrect Answers

Uniform low cost can increase expensive errors when a task exceeds the selected model’s validated capability.

Long messages are not necessarily high-risk, while short questions can require complex authorized decisions.

Model routing must still enforce confidentiality and approved provider boundaries for legal content.

Question 536

An agent router starts serving simple product FAQs with a smaller model. Cost falls but the small model’s wrong-answer rate increases on a subset of technical manuals. What should be measured?

A generic model-provider benchmark unrelated to the technical manuals.
Per-route factual error, fallback rate, total cost per correctly answered request and the affected topic slices.
Only the lower per-token expense of the small model.
Only how often the router selected the small model, excluding answer correctness.

Explanation

Correct Answer

A cost-saving route is successful only when the resulting quality stays inside the accepted error budget.

Incorrect Answers

External averages do not test the company’s actual hard topics and retrieval conditions.

Cheap tokens can become expensive if customers receive wrong technical advice and require follow-up.

Route frequency is a usage statistic and cannot establish the actual quality of the decisions served.

Question 537

An HR workflow needs to display a fixed remaining-leave number from an authenticated API. The agent router always invokes a costly reasoning model before making the read-only call. What optimization is justified?

Ask the expensive model to invent the leave balance from past chat messages.
Route requests randomly between expensive and cheap models without measuring workload complexity.
Use the deterministic typed lookup for straightforward authorized values and reserve model reasoning for genuinely ambiguous questions.
Remove the permission check so the value can be retrieved faster.

Explanation

Correct Answer

A stable structured field does not require general-purpose inference, so bypassing unnecessary calls can lower cost safely.

Incorrect Answers

A guessed value cannot replace the current HR record returned by an authorized API.

Random model choice is not a defensible cost/quality policy for different task classes.

Authorization is a mandatory service boundary even when the lookup itself is computationally simple.

Question 538

A contract assistant processes 2,500 input tokens and 600 output tokens per request, but finance budgeted for 600 tokens total. What needs to change?

Multiply only the final output characters by one token price.
Ignore prompt and retrieved-document tokens because they were not shown to the user.
Assume every request costs the same as the shortest contract in testing.
Model input and output usage separately at their applicable rates, including retrieval overhead and expected retries.

Explanation

Correct Answer

The budget undercounts the actual billed context and may also omit retries and larger-than-average documents.

Incorrect Answers

Input context and output have distinct token quantities and potentially different unit prices.

Hidden input still consumes model inference resources and can be billable.

Real document-length variation can materially change token consumption and cost per case.

Question 539

An enterprise equips 500 employees with licensed Copilot access but expects only 120 monthly active workers to use a new agent tool. How should the budget represent costs?

Model committed per-user license charges separately from metered agent calls and actual usage expectations.
Charge all recurring expenses only to 120 active users as though the other licenses are free.
Count every license as if it produces the maximum number of model calls daily.
Ignore employee licenses because the new agent is a software feature.

Explanation

Correct Answer

Seat-based access costs can be owed even when users are inactive, while metered service cost depends on workload.

Incorrect Answers

The license commitment can exceed the number of monthly agent users and must remain in ownership cost.

Licensing volume does not establish actual metered invocation demand, creating an unsupported usage estimate.

Existing access entitlements and any incremental user or agent licensing remain financial assumptions to verify.

Question 540

A finance agent handles 10,000 purchase-order requests monthly. Seven percent require human review and three percent need retries. The proposed budget includes only successful model tokens. What must be added?

Treat any retry as a second business success to raise the productivity metric.
Expected reviewer time, retry and tool-call usage, exception overhead and the cost of correcting failed outcomes.
Exclude every escalated request from the cost model because it is handled by a human.
Multiply the number of attempted requests by the average token cost and assume every transaction completes.

Explanation

Correct Answer

Manual and retried cases consume capacity and affect cost per correctly completed purchase request.

Incorrect Answers

An additional attempt is not an independently completed purchase and may reflect a prior ambiguous failure.

Escalation remains part of the end-to-end agent-enabled workflow and its labor cost belongs in TCO.

A raw attempt estimate ignores the extra retries, manual processing and incomplete operations specified in the case.

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

Microsoft AB-100 Practice Test 52 - 10 Questions (Last Reviewed 23-Sep-2026)
Microsoft AB-100 Practice Test 53 - 10 Questions (Last Reviewed 22-Sep-2026)
Microsoft AB-100 Practice Test 54 - 10 Questions (Last Reviewed 21-Sep-2026)
Microsoft AB-100 Practice Test 55 - 10 Questions (Last Reviewed 20-Sep-2026)
Microsoft AB-100 Practice Test 56 - 10 Questions (Last Reviewed 19-Sep-2026)

Pass Your Microsoft AB-100 Exam Easy!

Question 530

Explanation

Correct Answer

Incorrect Answers

Question 531

Explanation

Correct Answer

Incorrect Answers

Question 532

Explanation

Correct Answer

Incorrect Answers

Question 533

Explanation

Correct Answer

Incorrect Answers

Question 534

Explanation

Correct Answer

Incorrect Answers

Question 535

Explanation

Correct Answer

Incorrect Answers

Question 536

Explanation

Correct Answer

Incorrect Answers

Question 537

Explanation

Correct Answer

Incorrect Answers

Question 538

Explanation

Correct Answer

Incorrect Answers

Question 539

Explanation

Correct Answer

Incorrect Answers

Question 540

Explanation

Correct Answer

Incorrect Answers

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