Microsoft AI-901 Practice Test 26

Question 250

A Foundry project contains several deployments with different context windows, modality support and prices. A client must answer a long text question economically. What governs deployment selection?

Call the base model catalog name and assume the client will resolve it automatically to the deployment configured for that project.
Use the configured deployment alias whose tested capability, limits, costs and policy satisfy this workload.
Pick the deployment with the largest stated context window without checking whether its latency, data-location policy or unit price fits this request.
Pick the lowest-priced deployment and truncate the long user request whenever it exceeds that model’s supported token budget.

Explanation

Correct Answer

Match the long-text workload to a tested deployment whose context capacity, text support, cost, latency and policy fit the request. Use its actual callable alias and enforce a token budget.

Incorrect Answers

The deployment alias is an application resource identifier, not necessarily the catalog model identifier. Automatic name substitution can invoke the wrong target or fail resolution.

A larger window can admit long inputs, but it does not prove acceptable residency, cost or task quality. All relevant constraints must be evaluated for the actual deployment.

Truncating important requirements can make a cheaper deployment unusable or incorrect. Economy is evaluated only after context capacity and accuracy are demonstrated.

Question 251

A Foundry chat client already uses retrieval to find authorized policy chunks. For a complex query it retrieves twelve long passages; their combined tokens plus chat history exceed the deployed model’s context limit, leaving no space for a response. What should the client implement?

Increase response temperature and resend the same oversized retrieved context.
Drop the current policy excerpts at random and claim the answer has reviewed the full set of citations.
Preflight the full input and reserved output token budget, prioritize relevant current passages, and truncate or paginate evidence transparently when the prompt would exceed capacity.
Retry the over-limit call using the same deployment and the same prompt until the model eventually accepts it.

Explanation

Correct Answer

Calculate prompt instructions, chat history, retrieved chunk tokens and reserved answer budget before calling inference. Rank current authorized passages, chunk or narrow evidence as needed and disclose when coverage is partial.

Incorrect Answers

Temperature controls output variability, not the configured context limit. Resubmitting the same token-over-budget prompt leaves the capacity violation unresolved.

Random omission can remove controlling exceptions or effective-date clauses. Claiming complete source coverage after deleting passages is inaccurate and may cause a policy answer to be wrong.

An over-context-length error is deterministic for the same model and payload. Retrying without changing evidence selection or deployment capacity does not produce a valid request.

Question 252

A new invoice assistant must produce a fixed set of allowed status values, but some deployments lack the required structured-output feature. What should the architect verify?

Choose a model/API path supporting the desired response contract and validate enumerations and required fields in code.
Accept any status-looking phrase and infer the nearest enum during a financial write.
Assume every Foundry model implements identical JSON schema modes.
Remove output validation because a system prompt requests the field names.

Explanation

Correct Answer

Check that the chosen model and API version support the desired structured mode, then validate the complete parsed object against fixed allowed status values and required fields in application code before a financial write.

Incorrect Answers

Guessing the closest permitted enum from arbitrary prose can convert a rejection or uncertainty into a successful transaction status. Failed validation should be handled explicitly.

Structured-output features vary by model revision, SDK and API route. Catalog presence alone does not guarantee identical JSON schema enforcement across deployments.

A system prompt is guidance, not an API-level parser or schema gate. A fluent response can still omit required fields or invent unsupported statuses.

Question 253

A chat prototype responds quickly for ten internal testers. Product wants to use that single average to guarantee latency for 20,000 concurrent customers. What should be measured?

One successful response from a quiet period used as proof of maximum throughput.
Run a single high-volume synthetic prompt and use its lowest observed latency as the guaranteed 20,000-user tail latency.
Representative concurrency, response-time distribution and error/throttle rates at the intended deployment capacity.
Only the length of the first ten answers.

Explanation

Correct Answer

Load test realistic input/output token sizes and representative concurrency, then measure p50/p95/p99 latency, response errors and throttling against deployment quotas. Internal-ten-user mean time cannot justify a production SLA.

Incorrect Answers

One quiet-period completion shows only that a request could run; it says nothing about queueing, 20,000 concurrent users or high-percentile latency under load.

Best-case timing for one load pattern cannot represent concurrent real requests, varying token lengths, throttling or p95/p99 behavior. Test the actual deployment distribution.

Answer length from ten tests does not capture server queues, token generation rate, network failures or scaling behavior. Latency must be measured under the intended workload.

Question 254

A team member is listed as a tenant administrator but receives a forbidden response when trying to deploy a model in a particular Foundry project. What should be checked?

Share the resource owner's key with every developer as the default workaround.
Retry the same denied operation indefinitely without changing the role.
Use a chat prompt saying "ignore authorization" to convince the deployment API.
The user's effective Azure resource/project role assignment and whether the selected deployment operation is permitted.

Explanation

Correct Answer

Inspect the effective role assignment at the relevant subscription/resource/project scope and any deployment-specific permission. Tenant-level administration does not automatically imply rights to manage each resource.

Incorrect Answers

Distributing an owner’s credential defeats identity accountability and may grant excessive privileges. The correct fix is to authorize the developer for the precise deployment operation.

Retrying a persistent 403 without changing roles or scope cannot grant rights. It can create noisy logs while preserving the same authorization defect.

Prompt text is evaluated by the model after the deployment API has authenticated a caller. It cannot persuade the Azure control plane to bypass RBAC.

Question 255

A Foundry playground demo passes only clean FAQ requests, but the model may be asked for restricted documents and unsupported policy claims. What testing evidence is missing?

Representative negative, ambiguous and source-unavailable cases with measured factual and access-control outcomes.
Validate that the demo responses contain citation-looking references, but do not test whether those references resolve to authorized sources.
One aggregate token count without any scenario verdicts.
More repeats of the same three successful FAQ prompts.

Explanation

Correct Answer

Test negative and ambiguous queries, restricted documents and missing-source cases against expected authorization and factual outcomes. Passing clean FAQ demonstrations does not establish reliability under failure or attack conditions.

Incorrect Answers

A well-formatted citation is not proof of a current permitted source. The restricted-document and missing-evidence scenarios require verifying actual retrieved records, user permissions and truthful abstention.

Aggregate token use measures consumption, not whether a policy answer was supported or whether the assistant leaked restricted content.

Repeating three favorable examples does not expand coverage to missing evidence, prompt injection or unauthorized retrieval. The release gate needs representative failure-mode cases.

Question 256

A Foundry catalog lists a model family, but a deployment request for a specific model version with the geography-bound `Standard` SKU is rejected in a selected Azure region. What should the engineer verify first?

Assume a model listed anywhere in the catalog is available for every SKU in all Azure regions.
Check the model version, deployment type and region combination against the provider’s availability matrix before creating or replacing the deployment.
Change the SKU string to `Standard` in a client configuration after a failed deployment operation, without confirming backend support.
Provision a Global Standard deployment automatically and describe its inference processing as confined to the original geography.

Explanation

Correct Answer

The actual deployability depends on model version, supported deployment type and selected region. The engineer needs that concrete compatibility evidence before considering compliant alternatives and approval.

Incorrect Answers

Catalog visibility is not proof of deployment support for a specific version, region and capacity SKU. Availability must be verified at the combination level.

Changing a client label cannot make an unsupported deployment configuration available. A failed provider request must be diagnosed against documented SKU and region support.

Global processing can route inference outside the requested Azure geography. Relabeling it as geography-bound would hide a material compliance difference rather than fixing the unsupported combination.

Question 257

A Python developer is creating a lightweight Foundry chat client using the current Azure AI Projects 2.x SDK. Which initialization matches Microsoft's documented project-authentication pattern?

Use the Azure Speech region string alone as the Foundry project endpoint.
Paste a permanent administrator key into the system prompt and skip SDK authentication.
Create `AIProjectClient(endpoint=PROJECT_ENDPOINT, credential=DefaultAzureCredential())` and use its supported model client.
Create the project client using an arbitrary model display name in place of the project endpoint.

Explanation

Correct Answer

Initialize the Azure AI Projects 2.x AIProjectClient with the full project endpoint and DefaultAzureCredential, then obtain the supported model client for inference. Verify the running identity has effective project permissions.

Incorrect Answers

A Speech region is not a complete Foundry project endpoint and does not identify the authorized project URL. A project client requires the documented Azure AI Projects endpoint shape.

Placing administrator secrets in a model message discloses them as untrusted input and bypasses no SDK identity requirements. Credentials belong in supported identity management.

The catalog label names a model offering; it cannot substitute for a network project endpoint. The SDK must connect to the resource/project URL independently of model deployment selection.

Question 258

A Foundry sample accidentally committed a real production API key to Git. The team has removed the key from the latest source file, but earlier commits still contain it. Which incident response is necessary?

Add the file to .gitignore and assume earlier public or team-visible commits no longer contain the secret.
Ask the model to promise that it will never repeat the secret from source history.
Base64-encode the credential in the original historical commit so repository readers can no longer use it.
Immediately revoke or rotate the exposed credential, perform approved history/secret-scan remediation, and verify active deployments use protected replacements.

Explanation

Correct Answer

Credential revocation stops continued use of the leaked key. Repository history remediation and scanning reduce redisclosure; managed identity or protected configuration secures the replacement without assuming source edits erase past exposure.

Incorrect Answers

Ignore rules affect future tracked changes; they do not remove data already published in Git history. An active exposed key remains usable until revoked.

A prompt has no authority over repository copies, cached clones or an external attacker who already possesses the credential.

Encoding an already exposed secret does not revoke it and can be reversed. Historical copies and leaked credential value remain accessible.

Question 259

A copied Foundry quickstart still sets `PROJECT_ENDPOINT` to the literal text `your_project_endpoint`. The next request fails before inference. Which correction is necessary?

Put the project name into the user prompt and leave the client endpoint unchanged.
Change the model's maximum token setting while retaining the placeholder URL.
Use the actual project's complete endpoint from the authorized Foundry resource and validate its service path.
Assume the model's catalog name automatically resolves the project's network endpoint.

Explanation

Correct Answer

Copy the project’s real complete endpoint from the authorized Foundry project settings, validate its hostname and project path, and use it in client initialization with approved credentials.

Incorrect Answers

Putting a project name in the user prompt does not change where the SDK sends its HTTPS request. The project endpoint remains an unresolved placeholder.

Maximum output tokens is an inference parameter and cannot repair DNS, endpoint path or project binding before the request reaches a model.

A model catalog display name is not a project network address. Deployment naming and project endpoint configuration are separate required inputs.

Question 260

A sample chat client calls a Foundry model with an empty input list and receives no useful answer. Select TWO fixes that address the actual input contract.

Use the project subscription ID as if it were a user-message body.
Supply a valid user input/message containing the request the model should answer.
Delete the model deployment and assume a new deployment makes an empty message meaningful.
Increase temperature and continue sending an empty request.
Check message roles and structure against the API's supported input format before sending.

Explanation

Correct Answers

Include a supported user message or input item containing the actual question. The model cannot infer the task from an empty input list, even if a deployment exists.

Follow the installed API’s role/content or input-item schema and validate required user content before dispatch. Correct message structure is separate from the model’s deployment name.

Incorrect Answers

A subscription ID identifies an Azure billing/management scope, not the human question the model should answer. Sending it as content does not provide the missing chat request.

Creating another deployment does not add missing request content. This is an input-validation defect at the client, not evidence of bad provisioning.

Temperature only affects decoding choices after the model receives a valid prompt. Empty input remains empty regardless of sampling configuration.

Microsoft Azure AI Fundamentals (AI-901) Practice Tests

Microsoft AI-901 Practice Test 24 - 10 Questions (Last Reviewed 21-Oct-2026)
Microsoft AI-901 Practice Test 25 - 10 Questions (Last Reviewed 20-Oct-2026)
Microsoft AI-901 Practice Test 26 - 10 Questions (Last Reviewed 19-Oct-2026)
Microsoft AI-901 Practice Test 27 - 10 Questions (Last Reviewed 18-Oct-2026)
Microsoft AI-901 Practice Test 28 - 10 Questions (Last Reviewed 17-Oct-2026)

Pass Your Microsoft AI-901 Exam Easy!

Question 250

Explanation

Correct Answer

Incorrect Answers

Question 251

Explanation

Correct Answer

Incorrect Answers

Question 252

Explanation

Correct Answer

Incorrect Answers

Question 253

Explanation

Correct Answer

Incorrect Answers

Question 254

Explanation

Correct Answer

Incorrect Answers

Question 255

Explanation

Correct Answer

Incorrect Answers

Question 256

Explanation

Correct Answer

Incorrect Answers

Question 257

Explanation

Correct Answer

Incorrect Answers

Question 258

Explanation

Correct Answer

Incorrect Answers

Question 259

Explanation

Correct Answer

Incorrect Answers

Question 260

Explanation

Correct Answers

Incorrect Answers

Microsoft Azure AI Fundamentals (AI-901) Practice Tests