AI-901 Implement AI solutions by using practice questions
7-day money-back guarantee — full refund within 7 days of purchase if you've completed under 20% of the questions. See pricing →
Certifications Tools Flashcards Career Paths Exam Guides Blog Pricing For Teams About
Try 20 questions free →

AI-901 Implement AI solutions by using Microsoft Foundry: 754 practice questions

AI-901 754 questions 12 shown free

12 of the 754 Implement AI solutions by using Microsoft Foundry questions in the Certsqill AI-901 bank, shown in full below. Each one carries an explanation for every option, not just the correct one — the wrong answers are where the marks go.

Preparing for AI-901? Take the free 5-min readiness check →

1. Select a deployed model after checking that its supported: What should the team do next?

Hard
A sports club needs a Foundry chat application to summarize submitted practice notes. It will use an authorized identity and must choose an appropriate model before coding. What should the team do next?
  1. Choose a model solely because it offers the highest provisioned capacity, without testing representative practice notes.
    Capacity supports predictable throughput but does not establish summary quality or suitability for the workload.
  2. Select a deployed model after checking that its supported task, context, language, quota, and cost fit the summary workload. ✓
    Model selection should match capabilities and operational constraints before the team builds the chat application.
  3. Use any catalog model because generative models provide equivalent summaries regardless of supported capabilities or context.
    Models differ in supported tasks, context, quality, latency, cost, and availability, so arbitrary selection is unsuitable.
  4. Call the project endpoint directly and rely on the endpoint to select an available deployment for each request.
    An endpoint identifies the project service; the application still needs a supported model deployment or explicitly supported access path.
The trap
Assumes all generative models behave equivalently. Confuses a project endpoint with automatic model selection. Confuses capacity planning with model quality and task fit.

Choose a supported deployed model by matching capabilities and operational constraints to the chat requirement.

2. Choose the subscription: Select TWO appropriate next actions.

Hard
A manufacturing team is starting a Foundry solution for inspection-note analysis. It needs an organized project and a usable model workflow. Select TWO appropriate next actions.

Select two. More than one option is correct — every correct one is ticked below.

  1. Start coding against a model name before checking whether that model is available in the selected region.
    Availability, quota, region, and model capabilities should be checked before application development depends on a deployment.
  2. Grant every project user unrestricted tool permissions so analysis can proceed without approval checks.
    Tools should receive only necessary permissions, with results checked and suitable oversight for consequential actions.
  3. Choose the subscription, resource group, resource, and region, then create or use a Foundry project. ✓
    These resource choices establish the project context that organizes models, resources, and related assets.
  4. Create a new chat conversation to retrain the selected model on the team's inspection notes.
    Starting a new chat clears conversation context; it does not retrain model weights.
  5. Discover candidate models, review their capabilities and availability, and deploy a suitable model for testing. ✓
    Model discovery and deployment provide an appropriate model endpoint for the team's inspection-note analysis.
The trap
Underestimates deployment availability and regional constraints. Confuses conversation reset with model training. Confuses convenience with least-privilege and responsible tool use.

Create the project context and then discover, evaluate, and deploy a suitable model for the workload.

3. Confirm task fit: What should it check first?

Hard
A retailer selected a catalog model for a customer-support prototype and wants to know whether requests can be sent reliably from its chosen Foundry resource. What should it check first?
  1. Confirm task fit, then check the model's applicable access path, region, quota, and request identifier. ✓
    These checks establish whether the model fits the workload and whether the selected deployment or supported instant-access path is usable.
  2. Use the project endpoint as the model identifier and skip capability checks in the application.
    The endpoint identifies the project, while the applicable deployment or access identifier selects the model; capability checks remain necessary.
  3. Assume readiness because the model appears in the catalog and the project was created successfully.
    Catalog visibility and project creation do not prove capability fit, regional availability, quota, or usable access.
  4. Increase provisioned capacity and infer that support answers will become more accurate.
    Provisioned capacity supports predictable throughput but does not guarantee model accuracy or task suitability.
The trap
Confuses catalog discovery with inference readiness. Confuses capacity with answer quality. Confuses endpoint identity with model access and omits suitability checks.

Check task fit, applicable model access, region, quota, and the identifier required by that access path.

4. Open the model in the Foundry playground: What should the team do next?

Medium
A museum has deployed a chat model and wants to evaluate instructions for visitor questions before integrating code. What should the team do next?
  1. Start a new chat repeatedly and treat each cleared conversation as evidence that the model learned the instructions permanently.
    New chat clears conversation context; it does not retrain the model or permanently change its behavior.
  2. Add unrestricted web-search access before testing so every response is assumed to be grounded and correct.
    Tools require suitable configuration and permissions, and grounding does not guarantee truth or replace testing.
  3. Open the model in the Foundry playground, apply the instructions, and test typical and off-topic visitor prompts. ✓
    The playground enables prompt experimentation and representative testing before the museum commits to application integration.
  4. Publish the model immediately because a successful response to one visitor question proves reliable behavior.
    One successful response cannot establish performance across representative, ambiguous, or off-topic questions.
The trap
Confuses conversation reset with model learning. Confuses a single example with meaningful evaluation. Treats web access as automatic correctness.

Use the playground to test instructions against typical and off-topic prompts before integrating the deployed model.

5. Define system instructions that specify the role: What should you do next?

Easy
A nonprofit wants an assistant that always acts as a benefits-information guide, stays within its scope, and uses a consistent helpful tone. What should you do next?
  1. Deploy a larger model and rely on its learned behavior to maintain the nonprofit’s scope.
    Model size does not ensure adherence to organization-specific role, scope, or tone requirements without suitable instructions.
  2. Attach a knowledge file and assume retrieval will define the assistant’s role and communication style.
    Knowledge sources provide information, but they do not replace explicit behavioral and output instructions.
  3. Add the requirement as a user message each time someone asks a question.
    User messages request individual tasks; they do not reliably establish persistent role, scope, and behavior across conversations.
  4. Define system instructions that specify the role, permitted scope, and desired communication behavior. ✓
    System instructions establish the assistant’s role, scope, behavior, and output constraints for subsequent user tasks.
The trap
Learners may confuse a task request with persistent assistant behavior. A larger model can appear more capable but does not automatically implement organizational policy. Grounding content is confused with system-level behavioral guidance.

Use system instructions for persistent role, scope, and behavioral requirements rather than repeating them in individual tasks.

6. Add the required labels and section order to the system: What should the team do next?

Medium
A recruitment team needs every candidate summary to contain exactly three labeled sections: Skills, Experience, and Follow-up. What should the team do next?
  1. Increase the output token limit so the model has room to produce all sections.
    A larger token limit permits longer output but does not require specific labels or section ordering.
  2. Tell the model to be creative so summaries vary between candidates.
    Creative wording does not establish the required labels or guarantee a consistent three-section structure.
  3. Add the required labels and section order to the system instructions or task prompt. ✓
    Explicit format instructions communicate the required structure, labels, and ordering for generated summaries.
  4. Clear the conversation before every candidate so the model learns the format again.
    Clearing history removes context; it does not teach or enforce a desired response structure.
The trap
Variation is mistaken for compliance with a specified output format. Output capacity is confused with output-format control. New conversation context is incorrectly treated as format training.

State the required labels and ordering explicitly in instructions so generated summaries follow the desired format.

7. Instruct the assistant to say when the supplied material: Select TWO actions.

Easy
A finance operations team wants an assistant to answer invoice questions using supplied policy text, while avoiding unsupported conclusions. Select TWO actions.

Select two. More than one option is correct — every correct one is ticked below.

  1. Ask for a confident answer even when the policy contains no applicable information.
    Confidence cannot establish correctness when the supplied evidence does not address the invoice question.
  2. Instruct the assistant to say when the supplied material does not support an answer. ✓
    An uncertainty instruction encourages acknowledgment of insufficient evidence instead of unsupported conclusions.
  3. Include the relevant policy text with the user’s invoice question as source context. ✓
    Supplying relevant source material gives the model evidence to use when answering the invoice question.
  4. Raise the temperature so the assistant considers more possible policy interpretations.
    Sampling variation does not provide missing evidence and may make unsupported interpretations less consistent.
  5. Tell the model that its general training is an authoritative replacement for the policy text.
    Generative models are not guaranteed factual databases and should not replace the supplied governing policy.
The trap
General model knowledge is mistaken for authoritative, current organizational evidence. More varied generation is confused with better evidence-based reasoning. Confidence is incorrectly treated as proof of factual support.

Provide relevant source context and require uncertainty acknowledgment when that context does not support an answer.

8. Provide representative examples showing the desired: What should the team do next?

Medium
A document-processing team wants an extraction assistant to produce outputs matching a demonstrated invoice style. What should the team do next?
  1. Remove all instructions so the examples are the only source of behavior.
    Examples show style but may not define scope, exceptions, audience, or other behavioral requirements.
  2. Use unrelated examples because any additional text improves extraction reliability.
    Unrelated examples can introduce conflicting patterns and do not reliably demonstrate the required document style.
  3. Provide representative examples showing the desired input-to-output format. ✓
    Examples demonstrate the expected style and structure, helping the model pattern its response accordingly.
  4. Retrain the model’s weights using the examples directly in the prompt.
    Prompt examples guide generation temporarily; they do not retrain or modify model weights.
The trap
In-context examples are confused with model fine-tuning or retraining. Examples are incorrectly treated as a complete replacement for instructions. More examples are assumed helpful even when they are not representative.

Use representative examples to demonstrate the desired response style and structure without implying model retraining.

9. Define the classifier role in system instructions: What should you do next?

Medium
A document-processing assistant must consistently identify itself as an internal document classifier, while users submit individual documents for classification. What should you do next?
  1. Put each document in system instructions and the role in the user prompt.
    Documents are task-specific input, while the persistent role belongs in system instructions.
  2. Ask users to restate the classifier role whenever they submit a document.
    Repeated user wording is unnecessary and makes consistent behavior depend on every user.
  3. Put the classifier identity in each user's document-submission message.
    User messages should provide changing task content, not the assistant's persistent role definition.
  4. Define the classifier role in system instructions; send each document as user input. ✓
    System instructions establish persistent role and boundaries, while user input supplies each document and task.
The trap
Confuses persistent role instructions with the current user task. Mistakes user repetition for reliable system configuration. Reverses the roles of contextual data and persistent behavior.

Use system instructions for the persistent role and user input for each document task.

10. Start a new chat so the current conversation context is: What should you do?

Easy
A travel company’s chat assistant has accumulated an incorrect destination preference in its current conversation. The next traveler should start without that context. What should you do?
  1. Increase the model temperature to make it forget the preference.
    Temperature changes sampling variation; it does not remove messages or reliably erase conversation context.
  2. Start a new chat so the current conversation context is cleared. ✓
    Starting a new chat clears the current conversation context for the next interaction.
  3. Deploy the same model again so the new deployment has no conversation memory.
    Deployment selection does not itself clear application-managed conversation history or establish a new conversation.
  4. Add another user message instructing the model to ignore every previous message.
    An instruction may be followed imperfectly, but previous messages remain in the conversation context.
The trap
Sampling behavior is confused with conversation-state management. A prompt instruction is mistaken for actual history deletion. Model deployment is confused with conversation lifecycle control.

Start a new chat to remove the accumulated conversation context; this does not retrain the model.

11. Retain only conversation turns relevant to the current: Select TWO actions.

Easy
A customer support assistant receives long conversations containing obsolete troubleshooting attempts. The team wants to reduce irrelevant history while preserving useful context. Select TWO actions.

Select two. More than one option is correct — every correct one is ticked below.

  1. Retain only conversation turns relevant to the current support issue. ✓
    Removing unrelated turns reduces context consumption and limits distraction from obsolete or irrelevant information.
  2. Attach every previous conversation because more context always improves accuracy.
    Excessive or irrelevant history consumes finite context and can introduce conflicting or private information.
  3. Summarize useful earlier details before sending them with the current request. ✓
    A concise summary preserves important context while using fewer tokens than the complete conversation history.
  4. Start a new chat for every message, even when prior details remain necessary.
    Starting over removes useful context and may force users to repeat information unnecessarily.
  5. Increase temperature so irrelevant history has less influence on the answer.
    Temperature changes response variation but does not remove irrelevant messages or reduce their context consumption.
The trap
Sampling controls are confused with history filtering. More context is incorrectly assumed to be universally beneficial. History reduction is confused with indiscriminate loss of all context.

Filter irrelevant turns and summarize useful history to conserve context while retaining information needed for the current issue.

12. Acknowledge the gap and request supporting records: What should the team do next?

Medium
A finance operations assistant cannot find evidence in supplied records for a requested payment-status conclusion. What should the team do next?
  1. Raise temperature to produce several possible payment statuses.
    Generating alternatives does not supply missing records and may increase unsupported financial speculation.
  2. Ask it to infer the status from general model knowledge.
    General model knowledge is not authoritative evidence for a specific payment-status conclusion.
  3. Acknowledge the gap and request supporting records or human review. ✓
    The assistant should disclose insufficient evidence and seek appropriate support or oversight.
  4. Treat a high confidence score as proof of the payment status.
    A confidence score is not proof when the supplied records do not support the conclusion.
The trap
Confuses plausible inference with evidence-based verification. Mistakes multiple guesses for responsible uncertainty handling. Treats model confidence as factual validation.

Require transparent uncertainty and seek supporting records or appropriate human review.

742 more Implement AI solutions by using Microsoft Foundry questions

The remaining 742 questions in this domain are part of the full AI-901 bank — 1300 questions, every option explained. Start with the free five-minute check and see your score per domain.

Test your AI-901 readiness — free

Other AI-901 domains

Part of the Certsqill AI-901 question bank · Implement AI solutions by using Microsoft Foundry · Every answer, right and wrong, comes with its own explanation.