Microsoft AI-901 Practice Test 13

Question 120

An Azure Foundry model endpoint begins returning throttling responses after a large concurrency spike. What should the team inspect first?

Increase the client concurrency target before checking the deployment’s admitted token and request consumption.
Inspect average response latency only, without looking at 429 frequency, request/token limits or rejected admission.
Change prompt length and response-token limits without measuring whether token use or the deployment quota is the actual throttling constraint.
Quota, capacity, request rates and documented backoff behavior for that deployment.

Explanation

Correct Answer

Deployment quotas and measured RPM/TPM usage distinguish saturation or capacity bottlenecks from unrelated errors. Retry-After guidance and bounded backoff then prevent amplified load.

Incorrect Answers

Increasing client concurrency sends more simultaneous demand to the same quota or capacity limit. Without usage data this can worsen the throttling instead of explaining it.

Latency describes completed or timed requests; a service may reject requests promptly when its quota is reached. The analysis must include admission and limit metrics.

Token tuning can help in some workloads, but is premature without knowing whether request rate, token rate or backend capacity caused the 429 responses.

Question 121

A regulated Foundry project already uses a geography-bound model deployment. A later infrastructure-as-code change may add a Global Standard deployment, which could process prompts outside that geography. Which control can block that policy regression at deployment time?

Rely on a deployment-review checklist that records region preferences but permits any new SKU after the review is signed.
Apply an Azure Policy denial for prohibited globally processed deployment SKUs at create/update time, with controlled exceptions.
Use resource RBAC alone while allowing authorized engineers to create any deployment SKU without checking processing restrictions.
Encrypt prompts at rest inside the approved geography and assume the storage setting also restricts inference routing.

Explanation

Correct Answer

A deny policy on the actual deployment SKU prevents an unauthorized Global Standard resource configuration from being introduced through IaC or manual updates. This enforces the processing boundary at the infrastructure control plane, complementing human model selection.

Incorrect Answers

A checklist is an initial governance artifact, not an enforcement boundary. It cannot stop a subsequent template from creating a globally routed deployment that violates the approved inference geography.

RBAC governs who may deploy but not which deployment type an authorized engineer chooses. Even a properly authenticated owner could accidentally introduce a noncompliant Global Standard SKU without a policy guardrail.

Encryption and in-geography storage protect persisted data, but globally routed inference can still process prompts elsewhere. The deployment SKU requires its own policy enforcement to prevent that violation.

Question 122

A developer chooses a Foundry model from the catalog, creates a deployment named service-agent-prod, and then calls the catalog name directly. The client returns a deployment-not-found error. What should be corrected?

Rename the user's system prompt to match the model family.
Pass the actual configured deployment name expected by the inference API.
Use the catalog model version as the inference identifier because it resembles the deployment’s underlying model, without confirming the assigned alias.
Disable client authentication to work around a missing deployment.

Explanation

Correct Answer

In the deployed-model API, the client normally must target the actual deployment name assigned in the resource, not a generic catalog model family label. Confirm the correct endpoint and deployed alias.

Incorrect Answers

The system prompt name is conversational content, not the routing identifier used by an inference client. Renaming instructions cannot repair a missing deployment alias.

A model version may identify a catalog entry but not the deployment alias registered in the targeted resource. Calls must use the identifier required by that API and endpoint.

An authentication change cannot create a missing deployment. Disabling credentials would add a security failure instead of fixing the endpoint or model parameter.

Question 123

A Foundry chat client has a valid model deployment but receives an authorization failure. What should be checked?

Reuse a valid bearer token for a different Foundry project and assume it grants equivalent inference roles on the current resource.
The correct project endpoint, supported credential and effective resource role permissions.
Infer that viewing the Foundry project in a browser grants the same permissions as invoking model inference from the client.
A chat prompt claiming that the caller is an Azure administrator.

Explanation

Correct Answer

Confirm that the client uses the intended project endpoint and a supported credential with the resource roles required for inference. Browser sign-in and runtime token permissions may differ.

Incorrect Answers

A token can be syntactically valid yet lack the needed audience, tenant scope or role assignment for the targeted project. Verify resource-specific effective permissions.

Viewing a project in the portal does not necessarily grant a client identity permission to invoke a model. The effective principal and its assigned roles must be evaluated.

A chat prompt claiming administrator status is unauthenticated text. It cannot add Azure RBAC privileges or change an invalid runtime credential.

Question 124

A team wants a newly released model in a specific Azure region and deployment type. Which action is necessary before design approval?

Check current model, region and deployment-type support for that exact combination.
Select the longest context model and ignore local residency requirements.
Check that a model family exists in the Foundry catalog and infer that its newest revision supports the required regional deployment SKU.
Assume every new model is available immediately in all regions.

Explanation

Correct Answer

Model rollouts, SKU eligibility and regional capacity differ, so the exact model-version, deployment-type and region combination must be checked before design approval. Catalog popularity or release date is not sufficient.

Incorrect Answers

A longer context window does not satisfy restricted processing location or guarantee service availability in that region. Architecture approval must include the real compliance boundary.

Catalog presence does not imply that the exact revision is available under the requested SKU and Azure region. New versions and deployment types roll out independently, so the exact combination needs verification.

New models can launch in selected geographies or tiers before others. Assuming universal immediate availability can produce an unimplementable production plan.

Question 125

Two model replies to the same creative-writing prompt differ in wording after the application changes temperature. Which effect is most relevant?

Raise temperature to increase response diversity, then assume more diverse wording proves every statement is grounded in approved sources.
Lower temperature to reduce variability and treat that setting as a guarantee of byte-identical responses across every model revision.
Adjust temperature and top-p together until answers sound consistent, without checking whether cited facts are present in the trusted source.
Sampling variability can change, but temperature does not verify factual grounding.

Explanation

Correct Answer

Temperature affects the distribution used to select outputs and may change wording or creativity. The same application must separately validate facts and source evidence regardless of decoding settings.

Incorrect Answers

Higher temperature can change sampling and phrasing, but it does not connect generated claims to current source evidence. Creativity and factual grounding are separate quality dimensions that need their own evaluation.

Lower sampling randomness may improve reproducibility under controlled conditions, but provider revisions and runtime behavior can still change output. It cannot guarantee byte identity or factual validity across deployments.

Both parameters influence decoding choices, not the truth or currency of a quoted fact. Apparent wording consistency cannot substitute for source retrieval, citation resolution, and factual testing.

Question 126

A Foundry client sends long system instructions, ten attached document passages and asks for a lengthy answer. What budget must it validate?

Only the number of visible words in the final answer.
Estimate the prompt budget from the number and file size of attached PDFs without tokenizing the extracted text, system instructions or reserved output.
The total input context and output token allowance against the selected model's supported limits.
The file creation timestamp as a substitute for model context capacity.

Explanation

Correct Answer

The combined input tokens plus reserved output allowance must fit the selected deployment’s limits. Preflight checks can prevent rejection or unintended truncation of critical evidence.

Incorrect Answers

Counting only the final visible words ignores the system instructions, retrieved passages and hidden tokenization overhead that can already consume the model’s context budget.

PDF bytes and attachment counts do not equal model tokens. The service must account for extracted passages, chat history and output allowance within its context capacity.

A file timestamp may help select current source material but says nothing about whether the model can fit its text in context.

Question 127

An AI customer-support app must return a JSON object with an integer priority, one allowed status and a string summary. Select THREE necessary design controls.

Accept every response as long as it begins with an opening brace.
Use a structured output contract supported by the chosen model or client path.
Enable JSON mode and assume that syntactically valid JSON automatically guarantees an allowed `status` value and every required field.
Handle rejected or incomplete outputs safely without executing a downstream transaction.
Skip schema version checks when a new model is deployed.
Validate data types, enumerations and required fields in the application before use.

Explanation

Correct Answers

A structured output contract helps the client request the expected keys and field types when supported by the model and API. It reduces format variation but still needs application-side validation.

The client must safely reject, repair or escalate incomplete or invalid output without executing a transaction. Otherwise a truncated priority or missing status can trigger the wrong workflow.

Application validation must confirm integer priority, allowed status enumeration, required string summary and other business constraints before using the object. Parseable JSON alone is insufficient.

Incorrect Answers

An opening brace indicates only that the text might begin a JSON object. It does not prove valid syntax, the required fields or an allowed `status` value.

JSON mode can produce valid JSON without enforcing the application's strict enum and required-field contract. Supported structured outputs reduce shape errors, and the receiving service must still validate business constraints.

Changing model versions can alter output shape or supported structured-output behavior. Schema compatibility should be checked during releases rather than assumed.

Question 128

A developer sets a maximum output-token value beyond the deployment's permitted limit. What is the correct approach?

Copy the maximum-output limit from another model version without checking the actual deployed revision’s supported cap.
Use a supported maximum and test that essential content fits without unsupported truncation.
Choose a syntactically accepted token limit but omit testing whether the required result is cut off before the business-critical conclusion.
Validate only that the response parses as JSON and ignore the generation finish reason if required fields or conclusions were truncated.

Explanation

Correct Answer

Select a supported token allowance and test whether essential information fits. If the response is incomplete, adjust the task or request design rather than silently accepting truncation.

Incorrect Answers

The permitted output-token maximum is a capability of the deployed model and API, not a generic value transferable between families or revisions. An unsupported value must be corrected against the actual deployment contract.

Staying below the maximum avoids an invalid-parameter error but does not prove the response is complete. The app must still inspect completion/truncation signals and verify that essential content fits the allocated tokens.

A parseable fragment does not establish that the requested answer and evidence were fully generated. The completion status and required output contract must be checked before downstream use.

Question 129

A contact-center API already uses a Provisioned deployment but still misses its 95th-percentile response-time target during peak concurrent calls. Before reserving more PTUs, which operational evidence is needed?

Run a representative sustained-load benchmark measuring p95/p99 latency, token/request mix, PTU utilization and throttling before changing capacity.
Increase the reserved PTU count immediately without measuring whether the latency tail is caused by client queueing or large output requests.
Use the average latency of one low-concurrency request as the acceptance evidence for peak traffic.
Compare only the published model PTU minimum and assume it establishes the application’s actual p95 under its own prompt sizes.

Explanation

Correct Answer

The service is already Provisioned, so choosing that SKU again does not diagnose the missed latency target. Measured tail latency, utilization, concurrency and token traffic identify whether PTU capacity or workload shape is the limiting factor.

Incorrect Answers

Adding capacity can help if the deployment is saturated, but it does not prove the bottleneck. The requirement is to identify the cause of missed p95 under actual load, including client pacing and response length.

A single quiet-path average excludes concurrent queueing and long requests that dominate p95/p99 under peak load. The deployment’s steady-state distribution must be measured.

PTU minimums allow deployment but do not specify the latency distribution for this app’s prompt length, generation size or concurrency. A representative workload benchmark is needed.

Question 130

An app team changes its Foundry deployment to a new model revision while keeping the client code. Select TWO release checks that should precede the cutover.

Run representative answer-quality, safety and interface-contract regression tests against the new revision.
Test only the few development prompts that originally selected the old model and omit held-out failure cases for the new revision.
Verify the client targets the correct deployed version and retain a supported rollback plan.
Raise retry limits for the new model to mask tool-response schema failures without validating argument compatibility.
Assume compatibility because the vendor chose a similar model family name.

Explanation

Correct Answers

A revised model can change answers, refusal patterns and structured output behavior even if the client code is unchanged. Representative quality, safety and contract tests detect these regressions before release.

Verifying the actual deployment/version alias prevents accidental calls to an obsolete or wrong model. A supported rollback path lets the team restore known behavior if the new revision violates acceptance gates.

Incorrect Answers

Reusing favorable development examples can conceal regressions in source grounding, safety or structured tool outputs. The release needs representative holdout tests as well as actual deployed-version validation.

More retries may repeat an incompatible tool invocation and can duplicate side effects. The release gate must check contracts and typed arguments rather than treating compatibility defects as transient network errors.

Family-name similarity does not prove identical context limits, output formats or safety behavior. The new deployed revision must be validated rather than assumed compatible.

Microsoft Azure AI Fundamentals (AI-901) Practice Tests

Microsoft AI-901 Practice Test 11 - 10 Questions (Last Reviewed 03-Nov-2026)
Microsoft AI-901 Practice Test 12 - 10 Questions (Last Reviewed 02-Nov-2026)
Microsoft AI-901 Practice Test 13 - 10 Questions (Last Reviewed 01-Nov-2026)
Microsoft AI-901 Practice Test 14 - 10 Questions (Last Reviewed 31-Oct-2026)
Microsoft AI-901 Practice Test 15 - 10 Questions (Last Reviewed 30-Oct-2026)

Pass Your Microsoft AI-901 Exam Easy!

Question 120

Explanation

Correct Answer

Incorrect Answers

Question 121

Explanation

Correct Answer

Incorrect Answers

Question 122

Explanation

Correct Answer

Incorrect Answers

Question 123

Explanation

Correct Answer

Incorrect Answers

Question 124

Explanation

Correct Answer

Incorrect Answers

Question 125

Explanation

Correct Answer

Incorrect Answers

Question 126

Explanation

Correct Answer

Incorrect Answers

Question 127

Explanation

Correct Answers

Incorrect Answers

Question 128

Explanation

Correct Answer

Incorrect Answers

Question 129

Explanation

Correct Answer

Incorrect Answers

Question 130

Explanation

Correct Answers

Incorrect Answers

Microsoft Azure AI Fundamentals (AI-901) Practice Tests