Free AI-901 Practice Exam: Microsoft Azure AI Fundamentals

These original IT Mastery questions are independent practice, not official Microsoft questions, copied live-exam content, or exam dumps.

Start Question 1 · Questions 26–50 · Review your attempt

How to use this free exam

1. Set your own 45-minute practice timer if you want a timed attempt. This page does not run a timer, save responses, or calculate a score.
2. Write down your choice before opening the answer and explanation. This set has 47 single-answer questions and three Select TWO questions.
3. For this practice tally, award one point per correct question, for a total out of 50. Count a Select TWO item as correct only when both choices match, with no extra choice. Mark correct guesses for review too.
4. Use each explanation to identify the requirement that separates the strongest answer from a plausible alternative.

The set has 50 questions across both domains. Its length and timing are practice choices; Microsoft does not guarantee a fixed AI-901 question count or item-type mix. The blueprint explains the current official scope.

Practice-set coverage

Practice questions

Questions 1-25

Question 1

Topic: AI Concepts

A hiring model is evaluated before deployment. Overall accuracy is 89%, but recall among applicants independently confirmed as qualified differs by demographic group:

Which action best addresses this evidence?

Options:

• A. Reweight all groups equally during training, then retest recall across all groups.

• B. Remove demographic fields from model inputs, then retest recall across all groups.

• C. Lower one common decision threshold, then retest recall and precision across all groups.

• D. Analyze Group B false negatives and data representation, then retest group-level recall after targeted changes.

Best answer: D

Explanation: Recall measures how many actually qualified applicants the model identifies. Group B’s 68% recall, compared with about 90% for the other similarly sized groups, indicates that qualified Group B applicants are missed more often. Overall accuracy can conceal this disparity because it combines different outcomes and groups into one value. A proportionate response is to examine Group B’s false negatives, relevant data quality and representation, and model behavior before choosing a mitigation. Any resulting change should then be evaluated with group-level metrics.

The evidence supports investigation, but it does not yet establish that thresholds, demographic fields, or training weights caused the disparity.

• Common threshold change is premature because it changes performance tradeoffs without identifying the source of Group B’s lower recall.
• Removing demographic fields does not eliminate proxy effects and can make group-level fairness evaluation harder.
• Equal training weights assume representation imbalance even though the training distribution has not been identified as the cause.

Question 2

Topic: AI Concepts

An invoice pipeline has completed OCR and identified these candidate values:

Issued: 03/04/2026
Amount due: 1.250,00

Its supplied contract selects date-locale and number-conversion rules from the mapped target field. Final validation checks the converted ERP record, including required fields and totals. Which remaining sequence respects these dependencies?

Options:

• A. Normalize candidate values, map them to target fields, validate the record, then insert it.

• B. Map candidates to target fields, normalize their values, insert the record, then validate it.

• C. Map candidates to target fields, normalize their values, validate the record, then insert it.

• D. Map candidates to target fields, validate their raw values, normalize them, then insert the record.

Best answer: C

Explanation: A workflow should follow its actual data dependencies. OCR and candidate identification have already occurred, so the next task is to associate the candidates with target business fields. In this supplied pipeline, those fields select the rules for interpreting the date and decimal notation. Normalization can then apply the appropriate conversions. Final validation checks the resulting ERP record before insertion, including its required fields and numeric relationships. Other systems can normalize or validate at multiple stages; the stated rule-selection and final-validation contracts determine the order here.

• Normalizing first attempts conversion before this pipeline has selected the required field-specific rules.
• Checking only raw values does not validate the converted record that will enter the ERP.
• Validation after insertion permits an invalid converted record to reach the downstream system.

Question 3

Topic: Foundry Implementation

A team needs one deployed multimodal model that accepts either typed questions or recorded spoken questions directly and returns text. Client-side speech recognition is not allowed.

Evidence: deployed model-card summary

Which deployment should the team select?

Options:

• A. Select the voice-dialog deployment.

• B. Select the narrated-vision deployment.

• C. Select the audio-scribe deployment.

• D. Select the vision-chat deployment.

Best answer: A

Explanation: Input and output modalities are separate model capabilities. The application requires a multimodal deployment whose input modalities include both text and audio, while its output modalities include text. voice-dialog satisfies all three conditions and lets the application submit recorded speech without first converting it to text. A model that produces audio does not necessarily accept audio, and a speech-recognition model that accepts only audio does not support the required typed-question path. Model-card input capabilities therefore determine whether spoken prompts can be sent directly.

• vision-chat supports text and images, but recorded speech is not a supported input modality.
• narrated-vision can produce audio, but its input modalities do not include audio.
• audio-scribe accepts recordings, but it lacks text input and is not the required multimodal model.

Question 4

Topic: Foundry Implementation

A lightweight app must connect an AIProjectClient to the Helpdesk project. The supplied constructor requires a full project endpoint and a Microsoft Entra token-credential object; it does not accept an Azure OpenAI API-key string.

Configuration:

PROJECT_ENDPOINT=<verified Helpdesk project endpoint>
MODEL_ENDPOINT=<direct Azure OpenAI endpoint>
MODEL_KEY=<valid API-key string for direct model calls>
PROJECT_IDENTITY=<token credential authorized for Helpdesk>

client_type=AIProjectClient
endpoint=PROJECT_ENDPOINT
credential=MODEL_KEY

The current credential value has no token-acquisition method. Which focused correction establishes the required project connection?

Options:

• A. Use PROJECT_IDENTITY as the credential; switch to MODEL_ENDPOINT.

• B. Keep MODEL_KEY as the credential; switch to MODEL_ENDPOINT.

• C. Use PROJECT_IDENTITY as the credential; retain PROJECT_ENDPOINT.

• D. Use a refreshed MODEL_KEY as the credential; retain PROJECT_ENDPOINT.

Best answer: C

Explanation: Endpoint and credential must each match the client being initialized. This configuration already supplies the verified Helpdesk project endpoint, but its credential is a key string for direct model requests. The project client requires a token-credential object, and the supplied PROJECT_IDENTITY is authorized for the intended project. Replacing only the credential value addresses the demonstrated type mismatch.

A valid or refreshed API key remains the wrong credential type for this constructor. Switching to the direct model endpoint would also lose the required project scope.

• Refreshing the API key does not turn its string value into the required token-credential object.
• The authorized project identity does not make a direct model endpoint a valid project endpoint.
• Pairing the direct endpoint with its key still fails the required AIProjectClient connection contract.

Question 5

Topic: Foundry Implementation

An application must use the supplied team-language endpoint and matching API key with the Azure Language SDK.

Execution trace:

1. Deployment supplies:
   AZURE_LANGUAGE_ENDPOINT=<team-language endpoint>
   AZURE_LANGUAGE_KEY=<team-language key>
   TEST_LANGUAGE_KEY=<test-language key>
2. Startup loads AZURE_LANGUAGE_ENDPOINT and TEST_LANGUAGE_KEY.
3. TextAnalyticsClient is created with the loaded endpoint
   and AzureKeyCredential.
4. The client requests PII recognition.
5. The service returns 401 Unauthorized with no result.

Which stage is responsible for this result?

Options:

• A. Endpoint loading selects the team endpoint instead of the test endpoint.

• B. Client construction uses AzureKeyCredential instead of an Entra credential.

• C. Analysis invocation requests PII recognition before detecting the document language.

• D. Startup loading selects the test-resource key instead of the team key.

Best answer: D

Explanation: An Azure Language client must use an endpoint and credential accepted by the same service resource. The trace loads the required team-language endpoint but pairs it with the key for test-language. Client construction succeeds locally because the SDK cannot determine whether the key belongs to that endpoint. Authentication fails when the request reaches the service, producing 401 Unauthorized before PII analysis occurs.

Using AzureKeyCredential is appropriate when an API key is supplied. The fix is to load AZURE_LANGUAGE_KEY, not to change the requested analysis operation or target endpoint.

• Changing to the test endpoint would violate the requirement to use the supplied team resource.
• An API key is correctly wrapped in AzureKeyCredential; Entra authentication is not required here.
• PII recognition does not require a preceding language-detection call, and call order would not resolve an authentication failure.

Question 6

Topic: Foundry Implementation

A developer must use a direct OpenAI-compatible Responses client rather than a Foundry project client. The client accepts an endpoint, credential, and model deployment name.

Available configuration:

• Project endpoint: PROJECT_ENDPOINT
• Azure OpenAI endpoint: AZURE_OPENAI_ENDPOINT
• Deployment name: support-chat
• Base model name: gpt-4.1-mini
• Existing Entra credential: authorized for the Foundry resource
• OTHER_RESOURCE_KEY: belongs to a different resource

The developer configures PROJECT_ENDPOINT, the Entra credential, and support-chat. The request returns an HTTP 404 route error before producing a model response.

Which correction best matches the required workflow?

Options:

• A. Set the project endpoint, keep the Entra credential, and use support-chat as the deployment.

• B. Set the Azure OpenAI endpoint, keep the Entra credential, and use gpt-4.1-mini as the deployment.

• C. Set the Azure OpenAI endpoint, use OTHER_RESOURCE_KEY, and use support-chat as the deployment.

• D. Set the Azure OpenAI endpoint, keep the Entra credential, and use support-chat as the deployment.

Best answer: D

Explanation: A direct OpenAI-compatible model call must use configuration values from the same resource and client path. The Azure OpenAI endpoint is appropriate for the required direct client, while the project endpoint supports project-scoped client operations. The existing Entra credential is already authorized for the resource, so it does not need replacement. The model parameter must contain the configured deployment name, support-chat, rather than the underlying catalog model name. Mixing endpoint types, credentials from another resource, or base model names with deployment references prevents the request from reaching the intended deployment.

• Project endpoint continues mixing a project-scoped endpoint with the required direct client.
• Different resource key cannot authenticate access to the resource hosting the deployment.
• Base model name identifies the model family, not the callable support-chat deployment.

Question 7

Topic: AI Concepts

A support team must process chat transcripts to locate names, email addresses, phone numbers, and government identifiers, then return copies in which those details are masked. Which text analysis capability should the team use?

Options:

• A. Abstractive summarization and content condensation

• B. Key phrase extraction and relevance ranking

• C. Personally identifiable information detection and redaction

• D. Named entity recognition and entity categorization

Best answer: C

Explanation: Personally identifiable information (PII) detection is designed to locate sensitive personal details in text, such as names, contact information, and government identifiers. Its redaction capability can return a version of the text in which detected PII is masked, supporting safer storage or downstream processing. Detection results and redacted output should still be validated because repeatable processing does not guarantee complete identification of every sensitive value.

Named entity recognition can categorize entities, but it is not specifically designed to mask the full range of PII. The decisive requirement is both locating and redacting personal details.

• Entity recognition identifies categories such as people, places, and organizations but does not provide the required purpose-built PII masking.
• Key phrase extraction identifies important concepts in text rather than sensitive personal values.
• Abstractive summarization condenses content and does not reliably detect or redact personal details.

Question 8

Topic: AI Concepts

A customer-support generative AI application shifts 10% of traffic to a new model version. Monitoring uses the same evaluator and comparable request categories as before. Unsupported factual claims rise from 2% to 12% only in the new-version traffic. The prior version remains deployed, meets latency needs, and passed current reliability tests.

What is the most appropriate initial response?

Options:

• A. Revise the canary system prompt, continue the new version, and evaluate its production outputs.

• B. Lower canary temperature, continue the new version, and evaluate its production outputs.

• C. Stop canary traffic, restore the prior version, and evaluate the new version offline.

• D. Tighten canary content filters, continue the new version, and evaluate its production outputs.

Best answer: C

Explanation: The canary comparison provides strong evidence of a regression associated with the new model version: request categories and evaluation methods are comparable, while the unsupported-claim rate increased only for the new version. Reliability and safety incident response should first contain the known degradation by returning traffic to the validated version. Captured incident cases can then support targeted offline evaluation of the new version.

Monitoring and evaluators reveal problems but do not correct them. Changes to runtime parameters, prompts, or filters may be tested as possible mitigations, but they should not be treated as proven fixes while affected production traffic continues.

• Lowering temperature may reduce variability, but it does not establish factual correctness or justify continued exposure to the regression.
• Revising the system prompt could help, but the change requires validation before the new version resumes production traffic.
• Content filters address configured harmful-content categories and do not reliably detect unsupported factual claims.

Question 9

Topic: Foundry Implementation

An accounts-payable client maps invoice results from a custom Content Understanding analyzer. The destination requires a nonempty invoice ID, a total greater than 0, and an exception queue that preserves the original invoice and available analyzer evidence.

Current client mapping (simplified):

schema:
  invoiceNumber: string
  totalAmount: number
mapping:
  invoice_id: fields.invoiceNumber.value
  total_due: fields.totalAmount.value
post_when: analysis_completed
missing_or_invalid: substitute_zero
unsupported_or_failed: discard
retain_source_evidence: false

Which correction best aligns the client with the destination requirements?

Options:

• A. Treat completed analysis as field validation; queue unsupported or failed items and retain source plus available analyzer evidence.

• B. Validate mapped types before posting; substitute zero for missing amounts, queue analysis errors, and retain source plus available analyzer evidence.

• C. Validate mapped values before posting; queue incomplete, invalid, unsupported, or failed items and retain source plus available analyzer evidence.

• D. Validate mapped values before posting; queue all exceptions and retain the normalized destination record as the sole evidence.

Best answer: C

Explanation: Even a successful analysis status does not guarantee that every required field is present, valid, or suitable for the downstream contract. The client must validate mapped values, including the nonempty invoice ID and total greater than 0, before posting. Incomplete or invalid results and unsupported or failed analyses should enter the exception queue rather than be discarded or converted into invented values. The original invoice and available analyzer result, status, or extracted evidence should be retained for investigation and reprocessing.

A successful analysis status cannot replace field-level validation.

• Completion as validation overlooks missing or invalid fields in an otherwise completed analysis.
• Zero substitution produces a numeric value but violates the requirement that the total be greater than 0.
• Normalized record only removes the source and analyzer evidence required for reviewing extraction problems.

Question 10

Topic: Foundry Implementation

A museum guide uses Azure Speech Voice Live. Initially, the agent waits for a visitor to speak before responding.

The requirement changes: when a nearby visitor causes a Voice Live session to connect, the agent must deliver a welcome message before receiving any speech. How should the session configuration change?

Options:

• A. Increase the speech-recognition silence timeout for the session.

• B. Add a greeting instruction while keeping proactive engagement disabled.

• C. Enable interruption handling for the Voice Live session.

• D. Enable proactive engagement for the Voice Live session.

Best answer: D

Explanation: Proactive engagement determines whether a Voice Live agent can initiate an interaction. The original reactive behavior was appropriate while the agent needed to wait for visitor speech. Once the requirement changes so the agent must speak when the session connects, proactive engagement should be enabled.

Interruption handling controls whether user speech can interrupt an agent that is already speaking. A silence timeout affects speech-recognition timing. Instructions can define what the agent should say, but they do not independently trigger a conversational turn while proactive engagement remains disabled.

Enable proactive engagement only when agent-initiated speech is an intended part of the experience.

• Interruption handling supports user barge-in during agent speech; it does not cause the agent to begin speaking.
• Silence timeout changes when speech recognition treats input as complete; it does not initiate a welcome message.
• Greeting instruction alone defines intended behavior but does not trigger a turn while the session remains reactive.

Question 11

Topic: Foundry Implementation

Baseline: A help-desk application uses a project-derived Responses client to invoke the support-model deployment directly for generic responses.

Changed requirement: Each request must now execute the saved single agent HelpDeskAgent version 4, including its configured instructions and tools. How should the application change?

Options:

• A. Invoke HelpDeskAgent version 4 by its configured reference through the project-derived Responses client.

• B. Retrieve version 4 to find its model deployment, then invoke that deployment through the Responses client.

• C. Invoke support-model and include HelpDeskAgent version 4 as metadata on each model request.

• D. Invoke support-model and reproduce version 4’s instructions and tool calls in the application.

Best answer: A

Explanation: A model deployment and a saved agent are different invocation targets. Calling support-model directly performs inference with that deployed model, but it does not automatically apply HelpDeskAgent instructions or make its configured tools available. The application must invoke the saved agent through its configured identity and specified version using the project-derived Responses client.

Copying agent settings into application code might imitate some behavior, but it bypasses the saved agent configuration and can drift from it. The agent reference is what selects the intended reusable behavior.

• Reproducing instructions and tool calls in application code duplicates configuration instead of invoking the saved agent.
• Adding an agent name as model-request metadata does not change a direct model invocation into an agent invocation.
• Discovering and calling the underlying deployment still bypasses the agent’s configured instructions and tools.

Question 12

Topic: Foundry Implementation

A team selects a generative AI model from the Microsoft Foundry model catalog.

Workflow trace:

1. Review the model card and representative test results.
2. Open the Foundry portal playground and enter test prompts.
3. Find that the selected model is absent from the deployment list.
4. Attempt to run the prompt, but no inference response is generated.

Which stage is missing from the workflow?

Options:

• A. Create a model evaluation, select its result in the playground, and run the prompt.

• B. Create a model deployment, select it in the playground, and run the prompt.

• C. Configure the project endpoint, select it in the playground, and run the prompt.

• D. Create a saved agent, select it in the playground, and run the prompt.

Best answer: B

Explanation: Selecting a model from the model catalog does not make it callable. The team must create a model deployment, which exposes the selected model for inference. The deployment can then be selected in the Foundry portal playground, where the team submits a prompt and verifies that the model returns a response. Project endpoints support client access, evaluations measure model output, and agents add reusable instructions and tools, but none replaces the required model deployment.

The essential sequence is select, deploy, invoke, and verify the response.

• Project endpoint provides an access path but does not create a callable deployment of the selected model.
• Saved agent requires access to a deployed model and is unnecessary for testing the model directly.
• Model evaluation assesses model behavior but does not expose the model for playground inference.

Question 13

Topic: AI Concepts

A recycling app receives a photo containing one item and must assign one category, such as paper, plastic, glass, or metal, to the image. The app does not need the item’s location or pixel boundaries. Which computer vision task should the app use?

Options:

• A. Optical character recognition

• B. Object detection

• C. Image classification

• D. Semantic segmentation

Best answer: C

Explanation: Image classification is appropriate when the required output is an overall label for an image. Here, the app needs one material category and does not need spatial information about the item. Object detection would be appropriate if the app needed to identify and locate one or more items with bounding boxes. Semantic segmentation would classify individual pixels, which is useful when precise object boundaries are required. OCR extracts text and layout from visual content rather than assigning an overall material category. The required output granularity determines the computer vision task.

• Object detection adds location information that the recycling app does not require.
• Semantic segmentation produces pixel-level classifications rather than one image-level category.
• Optical character recognition extracts visible text instead of classifying the pictured material.

Question 14

Topic: Foundry Implementation

A team uses Azure Language sentiment analysis to return fixed sentiment and confidence fields for automated routing. The supported operation meets that baseline requirement.

Changed output requirement: Routing is removed. Analysts now want two differently worded, source-based interpretations of each message’s tone, with supporting phrases, for human review. There is no required label vocabulary, and variation in the prose is acceptable.

Which approach best fits the changed requirement?

Options:

• A. Use a general-purpose model to draft evidence-based narrative interpretations.

• B. Continue sentiment analysis and use its sentiment and confidence fields as the drafts.

• C. Use embedding similarity and present the nearest stored sentiment label as each draft.

• D. Use extractive summarization and present selected source sentences as the interpretations.

Best answer: A

Explanation: The original workflow benefits from a purpose-built operation with defined fields and repeatable output. The changed requirement instead asks for differently worded interpretations tied to source phrases. A general-purpose model can produce that flexible narrative form, with analysts checking that the interpretations remain supported by the message.

Probabilistic output is not inherently unsuitable when variation is acceptable and humans review it. The appropriate choice follows the required task and output contract, rather than treating either service type as universally preferable.

• Sentiment and confidence fields do not themselves provide the requested narrative interpretations.
• Extractive summarization selects existing sentences; those sentences need not interpret their own tone.
• Embedding similarity can identify related labels but does not generate evidence-based narrative drafts.

Question 15

Topic: Foundry Implementation

A team tests a single expense agent.

Goal: State policy limits only when supported by policy_lookup. Create a draft when requested, but call submit_expense only after the user explicitly confirms submission.

Current instructions: Help employees answer policy questions and prepare or submit expenses.

Test trace:

User: Is a client dinner capped at $100? Draft a $92 expense, but do not submit it.

Agent: Yes, the cap is $100. I submitted the $92 expense.

Tool calls: submit_expense only

Which TWO refinements most directly address the observed failures while preserving the intended capabilities? Select TWO.

Options:

• A. Disable submit_expense so the agent can draft but cannot submit expenses.

• B. Switch to a model with stronger tool-use benchmarks while leaving instructions unchanged.

• C. Require explicit submission confirmation after displaying the draft before calling submit_expense.

• D. Set temperature to zero so policy claims and action choices become deterministic.

• E. Require policy_lookup before stating limits and disclose insufficient retrieved evidence.

Correct answers: C and E

Explanation: The trace exposes two separate control gaps: grounding and action authorization. The agent stated a policy limit without using the available lookup tool, so its instructions should require retrieval and define behavior when evidence is insufficient. It also treated a drafting request as authorization to perform an external action, so its instructions need an explicit confirmation gate before submission.

These refinements preserve both intended capabilities: answering grounded policy questions and submitting expenses when authorized. Model selection and temperature can affect output quality or variability, but they do not define these required behavioral boundaries.

• Changing the model may improve tool use generally, but unchanged instructions still omit the grounding and confirmation rules.
• Lowering temperature reduces output variability but does not guarantee factual grounding or prevent unauthorized tool calls.
• Disabling submission prevents the observed action but also removes the required ability to submit confirmed expenses.

Question 16

Topic: Foundry Implementation

A developer uses a lightweight Responses client. The send_to_user function requires a string containing the generated reply.

response = client.responses.create(
    model=deployment_name,
    input="Suggest a name for a hiking app."
)
send_to_user(response)

Observed evidence:

Error: expected str, received Response
response.output: list
response.output[0]: message object
response.output[0].content: list
response.output_text: "TrailWise"

Which correction should the developer make?

Options:

• A. Call send_to_user(response.output_text).

• B. Call send_to_user(response.output[0].content).

• C. Call send_to_user(response.output[0]).

• D. Call send_to_user(response.output).

Best answer: A

Explanation: A Responses client returns a response object containing generated content and related structured data. Passing the complete object to a function that expects text causes the observed type error. In this client, output_text exposes the generated response as a convenient string, so it can be displayed or passed directly to another string-processing function.

The output property and nested content property are collections, while the first output element is a message object. Those values require further processing and do not satisfy the stated string requirement. The key step is retrieving the generated text rather than passing its containing response structure.

• Passing output supplies a list rather than the required generated-text string.
• Passing the first output element supplies a message object rather than its text.
• Passing the message’s content supplies a list of content items rather than a string.

Question 17

Topic: AI Concepts

An application currently makes direct model calls with this configuration:

• Model family: GPT-4.1
• Deployment name: support-prod
• Endpoint: Azure OpenAI endpoint
• Runtime: OpenAI-compatible Responses client

The deployment is replaced by another deployment of the same model family in the same Foundry resource. The new deployment is named support-v2, and support-prod is retired. The endpoint, authentication, and required client interface do not change.

Which implementation change should the developer make?

Options:

• A. Keep the client and deployment reference; change to the project endpoint.

• B. Keep the endpoint and deployment reference; replace the client with a project client.

• C. Keep the client and endpoint; change the deployment reference to support-v2.

• D. Keep the client and endpoint; change the deployment reference to GPT-4.1.

Best answer: C

Explanation: A model family identifies the underlying model type, while a deployment name identifies a configured, callable instance of that model. The endpoint identifies the resource interface, and the runtime client constructs and sends requests through that interface. Here, the model family, resource, endpoint type, authentication, and client interface remain unchanged. Only the callable deployment has been replaced and renamed, so the application must reference support-v2.

A model family name does not automatically substitute for a configured deployment name.

• Switching to the project endpoint changes the API path even though the scenario preserves direct Azure OpenAI calls.
• Replacing the runtime client is unnecessary because the required interface has not changed.
• Using GPT-4.1 confuses the model family identifier with the configured deployment name.

Question 18

Topic: AI Concepts

A support team creates a frequency report from device tickets. Corpus review shows that please is routine greeting language, while the frequent word fails identifies an important symptom. Product codes Z-1 and Z1 identify different devices. Whitespace differences carry no meaning. Which preprocessing policy best supports the report?

Options:

• A. Normalize whitespace, remove the validated greeting term, and preserve symptom words and product-code punctuation.

• B. Normalize whitespace, remove the validated greeting term, and strip punctuation from product codes.

• C. Normalize whitespace, remove all high-frequency words, and preserve product-code punctuation.

• D. Keep all original whitespace and greeting terms, and preserve symptom words and product codes.

Best answer: A

Explanation: Stop-word removal should depend on whether a term carries information for the actual analysis. A word is not dispensable merely because it occurs frequently: fails is common here because failures are a subject of interest. The corpus review supports removing please without losing the reported issue.

Whitespace can be normalized because it carries no meaning in this task. Product punctuation must remain because removing the hyphen would merge two distinct device codes. These decisions improve counting consistency without discarding the categories or symptoms that analysts need.

• Removing every frequent word would erase the important failure symptom along with greeting language.
• Stripping the product-code punctuation merges Z-1 with Z1, despite their different meanings.
• Preserving meaningless whitespace and greeting variation leaves avoidable noise in the frequency report.

Question 19

Topic: Foundry Implementation

A call-center app must transcribe a stored WAV recording rather than capture live microphone audio. The recording is in Canadian French, and its locale is known before recognition. Which speech recognition configuration should the developer use?

Options:

• A. Configure the WAV file as audio input and use en-CA.

• B. Configure the WAV file as audio input and use fr-CA.

• C. Configure the microphone as audio input and use fr-CA.

• D. Configure the WAV file as audio input and use fr-FR.

Best answer: B

Explanation: Speech recognition requires configuration for both the audio source and the language being recognized. Because the application receives a stored WAV recording, the recognizer must read from that file instead of the default microphone. The recognition locale should match the known spoken language and regional variant. Canadian French uses fr-CA; en-CA specifies Canadian English, while fr-FR specifies French as used in France.

The source configuration determines where audio comes from, and the recognition locale helps the service interpret that audio accurately.

• Using the microphone would capture live audio instead of processing the supplied recording.
• Using en-CA selects Canadian English rather than Canadian French.
• Using fr-FR selects a different regional variant of French than the known locale.

Question 20

Topic: AI Concepts

A team compares two models for summarizing specialized technical reports. Both models receive the same representative prompts, reports, and reference summaries. Higher evaluator scores are better.

The evaluator report identifies some factual omissions from both models. Which conclusion best uses this evidence?

Options:

• A. Neither result supports comparison; each candidate needs a tailored evaluation set.

• B. Candidate B leads for this workload; evaluator scoring has already corrected omissions.

• C. Candidate B leads for this workload; separately address omissions and re-evaluate.

• D. Candidate A leads for this workload; general benchmark evidence should take precedence.

Best answer: C

Explanation: Representative evaluation data is usually more relevant to a specific workload than a broad benchmark because it reflects the application’s actual prompts, inputs, and expected results. Candidate B has higher task-quality and groundedness scores on the same representative test set, making it the stronger candidate for this workload.

Evaluators measure characteristics such as quality, groundedness, and safety. They can reveal weaknesses and support comparison, but they do not change prompts, repair outputs, or improve a model automatically. The team must address the reported omissions through separate changes and then evaluate again.

• General benchmark priority overlooks that the workload-specific evaluation more closely represents the intended use.
• Automatic correction confuses measuring output problems with changing or repairing the outputs.
• Tailored test sets would weaken the comparison because the models would no longer be evaluated under the same representative conditions.

Question 21

Topic: AI Concepts

A payroll summarization application posts one report containing employee salaries to a collaboration channel that includes contractors. Access logs confirm two contractors opened it. Administrators can immediately restrict the report, and no other reports are affected. The organization has a documented privacy incident process. Which response is most appropriate?

Options:

• A. Restrict access to the report, preserve relevant evidence, ask about further use, and report only confirmed misuse.

• B. Restrict access to the report, preserve relevant evidence, report the exposure promptly, and correct the channel permissions.

• C. Suspend the application, preserve relevant evidence, report the exposure promptly, and block all users pending a full review.

• D. Redact the report, preserve relevant evidence, document the event in the project backlog, and correct the channel permissions.

Best answer: B

Explanation: Unauthorized access to sensitive personal data is a privacy and security incident, even when there is no evidence of further misuse. The team should immediately stop additional access, preserve logs and other evidence needed to determine the incident’s scope, and report the exposure through the documented process. It should then correct the channel permissions that caused the disclosure and verify the fix. Because the affected report can be restricted and no other reports are involved, suspending the entire application would be disproportionate. The response should combine prompt, scoped containment with formal reporting and targeted remediation.

• Recording the event only in a project backlog bypasses the documented privacy incident process.
• Waiting for confirmed misuse applies the wrong threshold because unauthorized viewing has already occurred.
• Blocking every user is broader than necessary when the affected report can be individually restricted.

Question 22

Topic: AI Concepts

Baseline: A news-monitoring application uses TextRank to rank connected sentences and returns the highest-ranked source sentences unchanged.

Changed requirement: Each summary must now combine facts from multiple sentences into concise wording that may not appear in the source.

How should the implementation change?

Options:

• A. Use a generative model only to rank and return the strongest source sentences.

• B. Cluster sentence embeddings and return one representative source sentence per cluster.

• C. Add an abstractive generation step that synthesizes the source into new wording.

• D. Retune TextRank to favor shorter sentences and return them in source order.

Best answer: C

Explanation: TextRank is a graph-based extractive technique. It represents sentences or terms as connected nodes, ranks them by importance, and selects highly ranked source content. Because its output is selected from the original text, TextRank alone does not combine ideas into newly written sentences.

The changed requirement calls for abstractive summarization. An abstractive model can synthesize facts from several source sentences and express them using new wording. TextRank could still help identify salient context, but a generation step is needed to produce the required summary.

• Generative ranking still returns original sentences, so the resulting summary remains extractive.
• Favoring shorter sentences changes TextRank’s selection but does not create synthesized wording.
• Embedding-based clustering identifies representative source sentences but still performs extractive selection.

Question 23

Topic: Foundry Implementation

A team is validating a customer returns agent in the Voice Live playground in the Foundry portal. What should the team do before the next validation run?

Evidence: Playground review

Options:

• A. Keep voice-support-v2, send returns-only instructions with each user request, enable interruption, and retest.

• B. Keep voice-support-v2, apply returns-only system instructions, enable proactive engagement, and retest.

• C. Keep voice-support-v2, apply returns-only system instructions, enable interruption, and retest.

• D. Switch to a different voice-capable model, apply returns-only system instructions, enable interruption, and retest.

Best answer: C

Explanation: Voice Live uses separate settings for the model, persistent agent behavior, and conversation handling. The selected deployment already matches the required model. A system prompt should define the reusable returns-only role and boundaries, while the interruption setting controls whether the agent stops speaking when the caller begins speaking. The evidence shows configuration mismatches in both of those areas, not a model failure. The team should correct those settings and repeat the spoken boundary and interruption tests.

Proactive engagement controls whether the agent initiates interaction; it does not replace interruption handling.

• Repeating instructions in user requests does not establish the required persistent role at the system level.
• Proactive engagement addresses agent-initiated conversation, not stopping output when the caller speaks.
• Changing deployments conflicts with the required model, and the evidence indicates configuration problems rather than inadequate model capability.

Question 24

Topic: Foundry Implementation

A developer adapts playground-generated code for a web form. The generated Responses client, endpoint, deployment, prompt, and runtime settings already match the tested setup. The supplied application contract says form_text contains the submitted question and reply.output_text contains the answer.

reply = tested_client.responses.create(
    model=deployment, instructions=rules,
    input="Explain the trial plan.",
    temperature=0.2, max_output_tokens=300
)
display(reply.output_text)

The form always receives the trial-plan answer. Which focused change uses the sample correctly?

Options:

• A. Replace the deployment reference with the model-family name; keep the literal input.

• B. Replace the literal input value with form_text; keep the tested client configuration.

• C. Replace instructions=rules with instructions=form_text; keep the literal input.

• D. Replace the tested client with a project client; keep the same literal input.

Best answer: B

Explanation: Playground code is a working starting point whose client, endpoint, deployment, instructions, and runtime settings should be carried into the application deliberately. The observed answer follows the literal sample question that remains in input; the web form’s text never reaches the request. Replacing that literal with form_text supplies the current question without changing the validated invocation path. Putting user text in the instructions field would alter its role and leave the old question present. Changing the deployment or client would likewise leave the actual input defect unresolved.

• Replacing system guidance with form text leaves the sample question as input and changes the instruction role.
• The configured deployment reference is already correct, and the sample input would remain unchanged.
• A different client does not connect the form value to the request and unnecessarily replaces the tested workflow.

Question 25

Topic: AI Concepts

A team is selecting a model for interactive customer-support summaries. Its requirements are:

• Quality score >= 87
• Safety score >= 95
• Median latency <= 900 ms
• Estimated cost <= $0.005 per request

The benchmark scores use the same representative dataset; higher scores are better.

Which model should the team advance to pilot testing based on this evidence?

Options:

• A. Advance Cobalt to pilot testing.

• B. Advance Birch to pilot testing.

• C. Advance Atlas to pilot testing.

• D. Advance Delta to pilot testing.

Best answer: B

Explanation: Model selection should reflect the constraints of the specific workload rather than a single leaderboard rank. Birch meets the minimum quality and safety scores while remaining within the latency and cost limits. Atlas has the highest quality score and fastest latency, but it misses the safety requirement. Cobalt offers the lowest cost and a strong safety score, but its quality is below the threshold. Delta satisfies quality, safety, and cost requirements but exceeds the latency limit.

Benchmark evidence narrows the candidates, while pilot testing with representative inputs confirms whether the selected model performs as expected in practice.

• Atlas fails because its safety score is below the required threshold.
• Cobalt fails because its quality score is below the required threshold.
• Delta fails because its median latency exceeds the permitted limit.

Questions 26-50

Question 26

Topic: AI Concepts

OCR correctly recognizes these lines on an invoice:

Pay by: April 3, 2026
Issued: March 4, 2026

A field-extraction stage must populate invoiceDate and dueDate. Which mapping correctly uses the recognized labels and their business meaning?

Options:

• A. invoiceDate: 2026-04-03; dueDate: 2026-04-03

• B. invoiceDate: 2026-03-04; dueDate: 2026-03-04

• C. invoiceDate: 2026-04-03; dueDate: 2026-03-04

• D. invoiceDate: 2026-03-04; dueDate: 2026-04-03

Best answer: D

Explanation: OCR provides recognized text and layout evidence, but field extraction assigns business meaning to that evidence. The label Issued associates March 4 with the invoice date, while Pay by associates April 3 with the due date. The payment deadline appears first on the page, so reading order alone should not determine which date belongs in which field. The values must be associated with their labels and mapped to the target schema. Copying one recognized date into both fields would discard the distinct meaning of the other date.

• Reversing the dates treats the first displayed date as the invoice date instead of interpreting its Pay by label.
• Using March 4 for both fields copies the issue date into the payment-deadline field.
• Using April 3 for both fields copies the payment deadline into the invoice-date field.

Question 27

Topic: AI Concepts

A home-design app receives a photograph of a furnished room. Which requested result requires visual generation rather than interpretation of the existing photograph?

Options:

• A. Report the current wall color and overall lighting conditions.

• B. Produce a revised room image with the walls changed to blue.

• C. Transcribe the words shown on signs and product labels.

• D. Mark each visible chair and table with a bounding box.

Best answer: B

Explanation: Visual interpretation derives information from existing pixels. Examples include describing image characteristics, detecting and locating objects, and extracting visible text. Visual generation produces new visual content, even when an existing image guides the result. Changing the walls to blue requires the system to synthesize a modified image rather than merely report what the original photograph contains.

The key distinction is the output: interpretation returns findings about the source image, while image-guided generation or editing returns new visual output.

• Reporting color and lighting describes properties already present in the photograph.
• Marking furniture with bounding boxes is object detection, which identifies and locates existing objects.
• Transcribing signs and labels is OCR, which extracts existing visual text.

Question 28

Topic: Foundry Implementation

A Foundry agent is configured to redact PII by using the managed Azure Language MCP server.

Execution trace:

1. The host sends the user’s text to the agent runtime.
2. The agent runtime, acting as an MCP client, selects the advertised PII tool.
3. The MCP client issues the tool call.
4. Missing stage
5. The agent model uses the returned analysis to compose a redacted response.

Which stage should replace the missing stage?

Options:

• A. The Azure Language MCP server runs PII analysis and returns results to the agent runtime.

• B. The MCP server returns the tool description, and the agent model infers the analysis result.

• C. The agent runtime runs PII analysis locally and returns results to the MCP server.

• D. The host invokes the deployed model again and treats its completion as the PII result.

Best answer: A

Explanation: MCP separates the caller from the tool provider. The agent runtime acts as the MCP client: it discovers an advertised tool, sends the tool call, and receives the result. The managed Azure Language MCP server exposes the PII capability, executes the requested analysis, and returns structured results. The agent model can then use those results to produce the final redacted response.

The key distinction is that the client requests tool execution, while the MCP server provides and executes the callable tool.

• Local execution by the agent reverses the client-server roles; the agent runtime requests the configured tool.
• Another model invocation would produce a model completion rather than the Azure Language tool result.
• A tool description explains availability and input structure, but it is not an executed analysis result.

Question 29

Topic: Foundry Implementation

A developer configures an OpenAI-compatible Responses API client for direct chat. The code reads AZURE_OPENAI_ENDPOINT, but the environment instead defines FOUNDRY_PROJECT_ENDPOINT; the credential is valid, and the configured support-chat deployment works in the model playground. Client startup reports that its required endpoint is missing. What should the developer change?

Options:

• A. Set AZURE_OPENAI_ENDPOINT to the Azure OpenAI endpoint.

• B. Set AZURE_OPENAI_ENDPOINT to the Foundry project endpoint.

• C. Replace support-chat with the base model name.

• D. Switch to a project client using FOUNDRY_PROJECT_ENDPOINT.

Best answer: A

Explanation: An OpenAI-compatible Responses API client requires the endpoint type expected by that client and the exact environment variable read by the application. A Foundry project endpoint does not substitute for an Azure OpenAI endpoint simply because both relate to the same project or resource. Here, the credential and deployment reference are already validated, while AZURE_OPENAI_ENDPOINT is absent. Adding the appropriate Azure OpenAI endpoint under that variable resolves the visible configuration mismatch.

The key diagnostic principle is to match the client path, endpoint type, and environment-variable name.

• Using the project endpoint supplies the wrong endpoint type for the existing client.
• Switching clients changes the interaction path rather than correcting the direct-model configuration.
• Using the base model name breaks the validated deployment reference and does not supply the missing endpoint.

Question 30

Topic: Foundry Implementation

A team is configuring one Foundry agent to answer questions from text or receipt images and call an order-lookup tool. The validated p95 response time must be <=2.0 seconds.

Observed behavior: Text requests succeed, but image requests return Unsupported image input.

Which correction best satisfies the agent’s requirements?

Options:

• A. Assign vision-no-tools-fast to the existing agent.

• B. Assign vision-tools-standard to the existing agent.

• C. Assign text-tools-fast to the existing agent.

• D. Assign vision-tools-fast to the existing agent.

Best answer: D

Explanation: Selecting a model deployment for a single agent requires satisfying every stated capability and operating requirement. Receipt processing requires image input, and order lookup requires callable-tool support. The deployment must also have a validated p95 response time no greater than 2.0 seconds. The image-input error indicates that a text-only deployment cannot provide the intended behavior, even though text requests work. Among the available deployments, vision-tools-fast meets the modality, tool-calling, and latency requirements simultaneously.

A multimodal deployment that lacks tool support or exceeds the latency limit does not fully satisfy the agent’s requirements.

• The text-only deployment meets the tool and latency requirements but cannot accept receipt images.
• The fast multimodal deployment accepts images but cannot call the required order-lookup tool.
• The standard multimodal deployment supports images and tools but exceeds the p95 latency limit.

Question 31

Topic: Foundry Implementation

A maintenance app should identify damaged parts in an equipment photo. The same prompt and JPEG work in the model playground with vision-inspect.

Application trace:

• Deployment: text-summary
• Input parts: instruction text and base64-encoded JPEG
• Error: Image inputs are not supported by this deployment.

What is the most likely cause of the application failure?

Options:

• A. The app omits the image and sends only the instruction text.

• B. The source image lacks enough detail for reliable interpretation.

• C. The app invokes a text-only deployment for a visual request.

• D. The app uses an unsupported base64 representation for the JPEG.

Best answer: C

Explanation: A deployed model must support every modality supplied in a request. Although the application includes both instruction text and a JPEG, it sends them to text-summary, whose deployment supports text only. The modality error occurs before the model can interpret the image. The successful playground test with vision-inspect confirms that the visual prompt and image can be processed by the image-capable deployment. The application should therefore reference vision-inspect while preserving the existing visual input. Image quality would affect interpretation accuracy, not produce an unsupported-modality error.

• Missing image does not fit because the trace confirms that the request contains a JPEG.
• Unsupported representation does not fit because the same JPEG succeeds with the image-capable deployment.
• Poor source quality could reduce result quality, but it would not cause the reported modality error.

Question 32

Topic: AI Concepts

A vision application resizes an RGB photo from 640 by 480 to 320 by 240 while preserving color. Which description correctly represents the resized digital image?

Options:

• A. A 320-by-240 grid, with numeric red, green, and blue values at each pixel

• B. A 640-by-480 grid, with smaller red, green, and blue values at each pixel

• C. A 320-by-240 grid, with a single numeric brightness value stored at each pixel

• D. A 320-by-240 grid, with numeric horizontal, vertical, and depth values at each pixel

Best answer: A

Explanation: A digital image is represented as numeric values arranged across a spatial grid. The resolution of 320 by 240 means the resized image contains 320 columns and 240 rows, or 76,800 pixel positions. Because the image remains RGB, each position has separate numeric values for the red, green, and blue channels. Resizing changes the number of spatial positions, while preserving RGB keeps the number and meaning of the color channels.

• A single brightness value describes a grayscale image rather than an RGB image.
• Resizing changes the spatial grid; it does not retain the original resolution and merely reduce channel values.
• RGB channels represent color components, not horizontal, vertical, and depth coordinates.

Question 33

Topic: AI Concepts

A help desk automatically routes messages when the language-detection confidence is at least 0.90. Otherwise, it sends them for manual language review. The standardized language code must be stored with each message.

Language-detection result:

Which routing action is best supported by this evidence?

Options:

• A. Route to the Spanish queue and store Spanish as the language code.

• B. Route to the Spanish queue and store es as the language code.

• C. Route to the Portuguese queue and store pt as the language code.

• D. Send to manual language review and store es as the language code.

Best answer: B

Explanation: Language detection returns a human-readable language or dialect, a standardized language code, and confidence evidence. Here, the service identifies Spanish and returns es as its code. The confidence score of 0.96 meets the stated 0.90 threshold, so the message qualifies for automatic routing to the Spanish queue. The application should store the returned code rather than substituting the language name. Manual review would apply only when the confidence falls below the routing threshold.

• Storing Spanish confuses the human-readable language name with the returned standardized code.
• Manual review is unnecessary because the confidence meets the stated automatic-routing threshold.
• Portuguese routing is unsupported because neither the detected language nor code indicates Portuguese.

Question 34

Topic: AI Concepts

A team is comparing models in the Microsoft Foundry model catalog for an assistant that must inspect Python snippets, identify likely defects, and produce corrected functions. Model names do not reveal their specialization. Which capability listed in the model cards should the team prioritize?

Options:

• A. Reasoning over multistep plans and logical constraints

• B. Natural-language generation for explanations and summaries

• C. Extraction of predefined fields from unstructured source content

• D. Coding for understanding, debugging, and generating source code

Best answer: D

Explanation: Model selection should begin with the required input, task, and output rather than a model’s name. This workload accepts source code, analyzes defects, and returns corrected source code, so the relevant model card should show strong coding capabilities such as code understanding, debugging, and generation. Representative tests using similar Python snippets can then confirm that capability.

General reasoning may support debugging, but reasoning evidence alone does not establish code proficiency. Likewise, natural-language generation focuses on prose, while extraction maps existing content to predefined fields instead of creating corrected functions. The key principle is to match documented and tested capabilities to the workload.

• Broad reasoning capability does not specifically demonstrate proficiency in understanding and correcting Python code.
• Natural-language generation could explain a defect but does not establish reliable corrected-code generation.
• Information extraction maps existing content into fields rather than synthesizing corrected functions.

Question 35

Topic: Foundry Implementation

An automated workflow receives plain-text support messages. It must detect each message’s language and identify PII, returning consistent structured fields that downstream code can process. Which approach should the team select?

Options:

• A. Use Azure Language for language detection and PII recognition.

• B. Use an embedding model with stored language and PII examples.

• C. Use Content Understanding with a custom ticket-field extraction schema.

• D. Use a generative model with a strict system prompt and low temperature.

Best answer: A

Explanation: Azure Language in Foundry Tools is designed for supported text-analysis tasks such as language detection and PII recognition. It returns defined result structures that application code can process predictably. A general-purpose generative model can analyze text, but prompting and a low temperature do not make its classification or output structure as purpose-built and repeatable. Structured output also does not guarantee perfect detection, so applications should still interpret confidence information and validate results when appropriate.

The key distinction is purpose-built structured analysis versus flexible, probabilistic text generation.

• Generative model uses probabilistic generation; low temperature and strict instructions do not make it equivalent to purpose-built text analysis.
• Content Understanding extracts schema-mapped information from documents and other media rather than serving as the primary tool for these supported plain-text tasks.
• Embedding model measures semantic similarity but does not directly perform language detection or PII recognition.

Question 36

Topic: AI Concepts

A support application is intended to use retrieval-augmented generation for current return policies. A test trace shows:

1. A search retrieves the relevant updated policy passage, which says returns are accepted for 30 days.
2. The model request contains the user’s question and a general instruction to answer helpfully, but no retrieved passage.
3. The generated answer states 60 days, and the application appends a link to the retrieved passage afterward.

Which change completes the intended grounding process?

Options:

• A. Increase the number of search results while keeping the model request unchanged.

• B. Add the retrieved passage to the displayed source links after generating the answer.

• C. Save the retrieved passage as a training example while keeping the current request unchanged.

• D. Include the retrieved policy passage in the prompt used to generate the answer.

Best answer: D

Explanation: RAG combines retrieval with generation: relevant external content is retrieved and supplied as context in the model prompt. The trace already shows successful retrieval, but the source passage never reaches the generation request. Adding a citation afterward does not make the earlier answer source-based. Passing the retrieved passage into the prompt closes this missing connection without retraining the model.

A grounded response still needs appropriate evaluation; supplying a source helps the model use current facts but does not guarantee perfect output.

• More search results do not help if none of the retrieved content is supplied to the model.
• Post-generation source links can show a document but cannot change the evidence used to produce an already generated answer.
• A saved training example is not prompt grounding and does not change the context of the current inference request.

Question 37

Topic: AI Concepts

A support team saves an agent with instructions defining its role and response boundaries. Each customer starts a separate conversation. The agent should remember product details mentioned earlier in the current conversation while applying the same support behavior in every conversation that uses the agent.

Which statement correctly explains these capabilities?

Options:

• A. Saved instructions provide reusable behavior; conversation context retains current-conversation details.

• B. Saved instructions retain current-conversation details; conversation context provides reusable behavior.

• C. The model deployment provides reusable behavior; saved instructions retain current-conversation details.

• D. Conversation context provides both reusable behavior and details across separate conversations.

Best answer: A

Explanation: A saved agent definition contains persistent instructions that establish reusable behavior, such as its role, tone, and boundaries. Those instructions apply whenever an application invokes that agent. Conversation context serves a different purpose: it carries relevant messages or retained state so the agent can interpret follow-up requests using details from the current conversation. Starting a separate conversation does not automatically transfer the previous conversation’s details.

The key distinction is reusable agent behavior versus conversation-specific working context.

• Reversing the roles incorrectly treats temporary conversation details as part of the saved instructions.
• A model deployment makes a model callable but does not define the saved agent’s behavior or conversation history.
• Conversation context does not automatically combine details from separate customer conversations.

Question 38

Topic: Foundry Implementation

A developer is configuring a single agent in Microsoft Foundry to perform PII recognition through Azure Language. The agent must use the managed Azure Language MCP server, avoid custom service-call code, and use the existing Foundry resource named contoso-foundry.

Which TWO actions should the developer take?

Options:

• A. Add Azure Language as a knowledge source for the agent.

• B. Add the managed Azure Language MCP server as an agent tool.

• C. Provision an Azure Language resource and configure its service endpoint.

• D. Configure the MCP server with the model deployment name.

• E. Configure the MCP server with the contoso-foundry resource name.

Correct answers: B and E

Explanation: The managed Azure Language MCP server lets a single agent call supported Azure Language capabilities without custom service-call code or a separately provisioned Azure Language resource. The server must be added as an available agent tool, and its connection context must identify the existing Foundry resource by name. This configuration enables the agent to invoke capabilities such as PII recognition through the managed integration.

A knowledge source supplies grounding context rather than callable language-analysis operations, while a model deployment name identifies an inference deployment rather than the resource connection required by the MCP server.

• Provisioning a separate Azure Language resource is unnecessary because the managed server uses the existing Foundry resource context.
• Adding a knowledge source would provide context, not expose PII recognition as a callable agent tool.
• Supplying the model deployment name does not identify the Foundry resource used for the managed server connection.

Question 39

Topic: Foundry Implementation

A lightweight agent client provides lookup_order and cancel_order tools. An authenticated user has confirmed, “Cancel order 4821,” but the agent calls lookup_order, receives eligible_for_cancellation: true, and replies that the order was canceled. What should the client do next?

Options:

• A. Report cancellation eligibility and ask the user to confirm the request again.

• B. Route to cancel_order, verify completion, and report the actual outcome.

• C. Regenerate the response from the lookup result and report the cancellation.

• D. Repeat lookup_order, verify eligibility, and report the order as canceled.

Best answer: B

Explanation: An agent’s natural-language response must be validated against the tool that was called and the result it returned. lookup_order only checks the order and establishes cancellation eligibility; it does not change the order’s state. Because the authenticated user already confirmed the request, the client should route the interaction to cancel_order, inspect its result, and report whether cancellation actually completed. A failed result should be reported or handled rather than converted into a success claim.

Repeated lookup results cannot substitute for evidence that the requested action occurred.

• Repeating the lookup reconfirms eligibility but cannot prove that cancellation occurred.
• Regenerating text changes the wording, not the order state or supporting evidence.
• Requesting confirmation again leaves the action incomplete because the user already confirmed it.

Question 40

Topic: AI Concepts

A generative AI assistant drafts employee travel guidance:

Employees may claim up to $2,500 per trip under Policy section 7.4.

The current authoritative policy sets a $1,500 limit and contains no section 7.4. Which two responses should the team take before publishing the guidance? Select TWO.

Options:

• A. Publish with an AI-generated disclaimer while preserving the wording.

• B. Regenerate with lower temperature and accept a repeated amount.

• C. Ask another generative model to confirm the amount before approval.

• D. Verify the amount and citation against the authoritative policy.

• E. Escalate the discrepancy to the policy owner before approval.

Correct answers: D and E

Explanation: Generative models can produce fluent claims and realistic-looking citations that are unsupported or fabricated. Here, the stated limit conflicts with the authoritative policy, and the cited section does not exist. Material claims should be checked against trusted sources, and unresolved discrepancies should be reviewed by the person accountable for the policy before publication. Lower temperature can reduce response variability, but it does not establish factual accuracy. Another model is also not an authoritative source, and a disclaimer does not make incorrect operational guidance acceptable.

• Lower temperature may produce more consistent wording, but consistency does not verify the amount or citation.
• Second-model confirmation can repeat the same unsupported claim and does not replace an authoritative source.
• AI disclaimer discloses how content was produced but does not correct inaccurate policy guidance.

Question 41

Topic: Foundry Implementation

A support application sends both a system prompt and a user prompt with each model call. The model must consistently act as a triage assistant, follow defined boundaries, and return a fixed output structure. The customer ticket changes with each call.

Which prompt design should the developer use?

Options:

• A. Place the reusable role and ticket in the system prompt; place the boundaries and format in the user prompt.

• B. Place all instructions and the ticket in the user prompt; reserve the system prompt for conversation summaries.

• C. Place the ticket and format in the system prompt; place the reusable role and boundaries in the user prompt.

• D. Place the reusable role, boundaries, and format in the system prompt; place the ticket in the user prompt.

Best answer: D

Explanation: A system prompt establishes the model’s role, behavioral instructions, boundaries, and output expectations for the interaction. These requirements are reusable across changing requests. The user prompt should contain the immediate task and its relevant source input, such as the current customer ticket.

Separating these concerns makes the application’s intended behavior consistent while allowing each user request to vary. System guidance must still be supplied or otherwise applied to the relevant interaction; it does not independently preserve unrelated conversation history. The key distinction is reusable behavioral guidance versus request-specific input.

• Putting the changing ticket in the system prompt mixes request-specific data with reusable behavioral guidance.
• Putting reusable boundaries in the user prompt treats application-level controls as part of the changing request.
• A system prompt is not reserved for summaries; conversation continuity and persistent behavior are separate concerns.

Question 42

Topic: Foundry Implementation

A speech synthesis app produces a zero-byte audio file while reporting success.

Workflow trace:

1. Configure the synthesizer, voice, and output format.
2. Await the speech synthesis request.
3. Save returned audio bytes and report success.
4. Inspect the synthesis result state.
5. If canceled, read the cancellation evidence.

Observed evidence:

Result state: Canceled
Error category: AuthenticationFailure
Details: Supplied credential was rejected
Saved output: 0 bytes

Which workflow correction best addresses both the failed synthesis and the false success report?

Options:

• A. Repeat stage 1 with another output format; validate new audio before inspecting cancellation evidence.

• B. Keep stage 3 first; treat nonempty audio as completion and inspect cancellation only if empty.

• C. Retry stage 2 before stage 5; inspect cancellation evidence only after another failed attempt.

• D. Move stages 4-5 before stage 3; fix authentication, then validate audio after completion.

Best answer: D

Explanation: A completed request call does not necessarily mean speech synthesis succeeded. The app should first inspect the synthesis result state. If the result is canceled, it should read the cancellation category and details instead of saving or playing the returned bytes as valid audio. Here, the evidence identifies rejected authentication, so the credential must be corrected before retrying. Only a completed synthesis result should proceed to output validation, such as confirming that audio data or the intended destination contains usable output.

Checking cancellation evidence before processing audio prevents canceled requests from being reported as successful.

• Treating byte count as the primary status can misclassify a canceled request and delays use of explicit error evidence.
• Retrying before examining cancellation details ignores an actionable authentication failure and is unlikely to change the result.
• Changing the output format does not address the reported credential rejection.

Question 43

Topic: AI Concepts

A customer-support assistant can check delivery status using an order number and postal code. The same conversation previously covered a billing dispute and an accessibility request, neither of which is needed for the delivery check.

To apply data minimization when preparing the next model request, what should the application include?

Options:

• A. The current request and the masked full conversation transcript

• B. The current request and the ten most recent messages

• C. The current request and a summary of all earlier topics

• D. Only the current request, order number, and postal code

Best answer: D

Explanation: Data minimization means processing only information necessary for a stated purpose. The delivery check requires the current request, order number, and postal code, so unrelated billing and accessibility details should not be supplied as model context. Necessary personal data can still be used when the task depends on it, but its inclusion should be limited by purpose rather than conversation length, convenience, or masking alone.

Summarization and masking can reduce privacy risk, but they do not justify sending irrelevant information.

• Summarizing every earlier topic preserves unrelated information even if the resulting context is shorter.
• Selecting recent messages limits context by age rather than by relevance to the task.
• Masking a full transcript may reduce identifiability but still exposes unnecessary conversation content.

Question 44

Topic: AI Concepts

A retailer needs a customer-facing application that compares two product photos in one request and answers in French. Commercial use must be permitted.

The team assigns aliases to these model-catalog candidates:

Which candidate best fits all requirements?

Options:

• A. Select Maple Vision.

• B. Select Pine Vision.

• C. Select Orchid Vision.

• D. Select Cedar Vision.

Best answer: C

Explanation: Model selection should consider the complete model-card evidence, not capability alone. The application requires two images in one request, French output, commercial licensing, and suitability for visual comparison. Orchid Vision satisfies all four conditions. A model that accepts images may still be unsuitable because of its image-count limitation, supported output languages, license, or intended use.

The key takeaway is to evaluate capabilities, limitations, licensing, and intended use together before deployment.

• Pine Vision supports multilingual visual questions, but its one-image limit prevents the required two-photo comparison.
• Maple Vision supports commercial visual comparison, but its English-only output does not meet the French requirement.
• Cedar Vision meets the technical requirements, but its noncommercial research license excludes the customer-facing commercial use.

Question 45

Topic: Foundry Implementation

An insurance company receives claim forms and repair estimates as PDFs with varying layouts. The solution must extract policy details and a collection of repair items into a consistent JSON structure. The required fields are specific to the company’s claims process, and free-form text output is not acceptable.

Which approach should the company use?

Options:

• A. Use OCR and Azure Language entity recognition to assemble the required structure.

• B. Use a custom Content Understanding analyzer with the required nested schema.

• C. Use a prebuilt Content Understanding invoice analyzer and rename its output fields.

• D. Use a multimodal model with instructions to generate the required JSON structure.

Best answer: B

Explanation: Azure Content Understanding in Foundry Tools extracts structured information from documents by applying an analyzer and schema. A custom analyzer is appropriate because the company needs organization-specific policy fields and a nested collection of repair items across documents with varying layouts. The schema defines the consistent structure expected by the downstream claims process.

A prebuilt analyzer is better when its supported common scenario and existing fields match the business requirement. OCR extracts text and layout evidence but does not by itself map that evidence into the required business fields. A general multimodal model can produce JSON, but free-form generation is not the purpose-built structured extraction approach required here.

• Prebuilt invoice analyzer provides predefined invoice fields that do not match the company-specific claims schema.
• OCR and entity recognition can identify text and entities but do not directly perform layout-aware mapping to the nested schema.
• Multimodal generation may follow JSON instructions, but it is not the purpose-built analyzer for consistent field extraction.

Question 46

Topic: AI Concepts

A company uses a third-party AI system to rank job applicants. A recruiter enters the job criteria, a data scientist evaluates ranking quality, and a hiring manager authorizes applicant rejections based on the rankings.

Under the responsible AI principle of accountability, who remains responsible for the resulting hiring decisions?

Options:

• A. The employer and the hiring manager who authorizes the decisions

• B. The evaluation team and the data scientist who assesses ranking quality

• C. The recruiting team and the recruiter who enters the job criteria

• D. The vendor and the product manager who supplies the ranking system

Best answer: A

Explanation: Accountability means that people and organizations remain responsible for AI-assisted outcomes. An AI system can provide rankings or recommendations, but it cannot assume responsibility for a business decision. Here, the employer uses the system in its hiring process, and the hiring manager has authority to approve rejections. They therefore remain accountable for the decisions, including appropriate oversight and review.

Vendors, evaluators, and recruiters have responsibilities for their own work, but those duties do not transfer ownership of the final hiring outcome away from the decision-making organization and its authorized manager.

• The vendor supplies the system but does not make or authorize the employer’s applicant decisions.
• The evaluation team assesses model quality but lacks authority over the final hiring outcome.
• The recruiter provides job criteria, while the hiring manager authorizes the resulting rejections.

Question 47

Topic: Foundry Implementation

A retailer analyzes customer reviews. For each review, it must identify product features such as “battery life” and determine the positive or negative opinion associated with each feature, rather than only the review’s overall sentiment. Which Azure Language capability best meets this requirement?

Options:

• A. Named entity recognition

• B. Key phrase extraction

• C. Extractive summarization

• D. Sentiment analysis with opinion mining

Best answer: D

Explanation: Sentiment analysis determines whether text expresses positive, neutral, or negative sentiment. Opinion mining adds aspect-level detail by identifying a target, such as “battery life,” and connecting it to an assessment, such as “excellent” or “poor.” This directly supports the requirement to determine sentiment about individual product features instead of reporting only an overall review score.

Key phrase extraction can identify prominent topics, but it does not associate each topic with an opinion. The deciding distinction is the required relationship between a feature and its expressed sentiment.

• Key phrase extraction identifies important terms but does not connect them to positive or negative assessments.
• Named entity recognition identifies categorized entities such as people, organizations, and locations, not feature-level opinions.
• Extractive summarization selects important source sentences but does not produce structured aspect-and-sentiment relationships.

Question 48

Topic: Foundry Implementation

A team is selecting a model to summarize supplied customer-support policies. The release gate requires at least 90% of tested summaries to preserve every required policy condition, a safety pass rate >=97%, p95 latency <=2 seconds, and support for inputs up to 8,000 tokens.

Observed: Model A was chosen from its broad benchmark alone, but pilot summaries omitted important conditions.

Evidence: The evaluation used 200 representative cases with the intended settings. Condition-preservation pass rate measures the proportion of summaries containing every required condition, checked against a reference checklist.

Which correction should the team make before deployment?

Options:

• A. Select Model B, balancing representative quality with stated operating constraints.

• B. Select the model with the highest unweighted average across reported metrics.

• C. Select Model C, prioritizing its highest condition-preservation pass rate.

• D. Select Model A, prioritizing its broad benchmark and lower latency.

Best answer: A

Explanation: The release gate combines representative task quality with operating constraints. Model A fails the 90% condition-preservation requirement even though it leads the broad benchmark. Model C preserves conditions most often but exceeds the two-second latency limit. Model B meets the stated thresholds and supports the required input length.

The condition-preservation checklist directly measures the observed omission problem. Groundedness would instead assess support for the claims a summary includes; it would not by itself establish that all required information is present. These test results support selection for this workload, not a guarantee about every future response.

• Model A’s 82% condition-preservation rate fails the 90% release gate despite its broad benchmark advantage.
• Model C’s 2.7-second p95 latency exceeds the two-second maximum despite its higher preservation rate.
• An unweighted average mixes measurement scales and can hide failure of a mandatory threshold.

Question 49

Topic: Foundry Implementation

A company creates a custom Content Understanding document analyzer with an amountDue field. Test invoices use different layouts. One says “Balance to remit: $1,240.00”; another says “Please pay $860.00 by June 30” without an amount-due label. Both values appear in amountDue in the structured results. Which capability best explains this behavior?

Options:

• A. Position-based extraction maps fixed page regions directly to the amountDue field.

• B. Semantic extraction maps contextual meaning to the schema-defined amountDue field.

• C. Free-form summarization produces text that the application maps to the amountDue field.

• D. Label matching maps recognized document captions directly to the amountDue field.

Best answer: B

Explanation: Schema-based semantic extraction maps source content to fields according to meaning and context. The source does not need to use the schema field’s exact name or provide a separate label. Here, both “Balance to remit” and the payment sentence communicate an amount owed, so the analyzer can place each value in amountDue. OCR can recognize the visible text, but recognition alone does not assign its business meaning. The varying layouts also rule out reliance on fixed coordinates, while the structured result differs from free-form summarization.

• Label matching cannot explain the unlabeled amount in the second invoice.
• Fixed-position extraction is unsuitable because the invoices use different layouts.
• Free-form summarization would not directly produce the stated schema-mapped result.

Question 50

Topic: Foundry Implementation

A company processes supplier invoices containing multiple line items. Lot code and expiry date appear in labeled columns and must remain associated with the corresponding item. The application requires the nested result shown and will not parse free-form text.

configured_analyzer: prebuilt-invoice
prebuilt_line_items:
  - description
  - quantity
  - amount
required_line_items:
  - lot_code
  - expiry_date

Which configuration correction should the team make?

Options:

• A. Keep the prebuilt analyzer and add lot code and expiry date as output aliases.

• B. Use a custom analyzer with lot code and expiry date nested per line item.

• C. Keep the prebuilt analyzer and obtain lot code and expiry date from description.

• D. Use a custom analyzer with separate top-level arrays for lot codes and expiry dates.

Best answer: B

Explanation: A prebuilt analyzer is appropriate when its predefined schema covers a supported common extraction scenario. Here, the prebuilt invoice line-item schema does not include two required fields. Because each invoice can contain multiple items, the values must also be represented inside each corresponding line-item object. A custom analyzer allows the team to define that nested schema and extract the labeled values into structured results.

Separate arrays would not satisfy the required object structure, while aliases cannot add extraction capabilities to a prebuilt schema. Reading the description field would require unsupported assumptions or additional parsing.

• Separate arrays fail to preserve the required nested line-item structure directly.
• Output aliases can rename available data but cannot extend the prebuilt analyzer with new extracted fields.
• Description parsing does not produce the required schema-mapped lot and expiry fields.

Turn your attempt into a study plan

Use the domain label beside each question to split your tally, and record missed or guessed question numbers for each domain. The questions are mixed throughout the set; Questions 1–25 and 26–50 are navigation groups, not separate domains.

These raw practice tallies do not convert to Microsoft’s scaled exam score . Choose your next session from the missed and guessed topics, even when the combined total looks strong.

Track concepts and implementation separately. A stronger total on an immediate repeat can reflect memory of these answers; it is not an official pass prediction.

Report a question issue

Email a question report with the question number, what seems incorrect or unclear, and a supporting Microsoft reference if available. The draft includes this page’s address; you do not need to copy the question. You can also write directly to support@masteryexamprep.com .

For sign-in or paid-access problems, follow the account support steps .

Continue in the web app

Use IT Mastery for interactive AI-901 practice with mixed sets, timed mocks, topic drills, explanations, and progress tracking.

Try AI-901 on Web