Microsoft AI-300 Dumps
Microsoft AI-300 Practice Test 35
Question 340
A legal research assistant requires accurate grounded citations and must compare candidate foundation models against task-relevant false citation and cost metrics. Which combined implementation and verification plan is the best match for this workload?
Answer & Explanations
Correct answer: B
The action is applicable: Compare candidate Foundry foundation models using representative prompts, grounded domain tasks, output safety measures, latency and expected token cost before choosing a production deployment. But its proposed test—Run load tests with realistic prompt and output lengths, measure tokens per minute, throttling and p95 latency, then verify the selected quota meets the traffic SLO.—addresses “Provision Foundry token throughput capacity for stable peak inference”. This requirement is confirmed instead by: Run a blind paired evaluation on a fixed representative set, record quality and risk metrics alongside measured p95 latency and token cost and approve the best feasible candidate.
The question is set in this situation: A legal research assistant requires accurate grounded citations and must compare candidate foundation models against task-relevant false citation and cost metrics. Compare candidate Foundry foundation models using representative prompts, grounded domain tasks, output safety measures, latency and expected token cost before choosing a production deployment. Confirm it through this test: Run a blind paired evaluation on a fixed representative set, record quality and risk metrics alongside measured p95 latency and token cost and approve the best feasible candidate. Model size or catalog popularity alone does not determine fit; task success and operational limits depend on input context, evaluation criteria and workload-specific performance.
The test would be relevant: Run a blind paired evaluation on a fixed representative set, record quality and risk metrics alongside measured p95 latency and token cost and approve the best feasible candidate. But the offered implementation—Estimate requests per second, token input-output mix, rate limits and desired service latency, then configure supported provisioned throughput capacity or other appropriate deployment quota for a high-volume model.—addresses “Provision Foundry token throughput capacity for stable peak inference”. The action needed here is: Compare candidate Foundry foundation models using representative prompts, grounded domain tasks, output safety measures, latency and expected token cost before choosing a production deployment.
This action/test pair addresses “Design resilient model endpoint fallback without silently changing data policy”. Define approved fallback model deployments with compatible inputs, data handling and quality gates, and trigger fallback only on supported operational errors rather than masking every low-quality answer. Simulate a primary deployment rate limit in staging, verify only a preapproved compatible endpoint handles the request and confirm the returned format and audit trail remain valid. The case instead calls for: Compare candidate Foundry foundation models using representative prompts, grounded domain tasks, output safety measures, latency and expected token cost before choosing a production deployment.
Question 341
An insurer wants a hosted sentiment model without maintaining serving VMs but its data scientist also evaluates an open-weight model with custom runtime requirements. Which Azure AI operations action best addresses this requirement?
Answer & Explanations
Correct answer: B
Estimate requests per second, token input-output mix, rate limits and desired service latency, then configure supported provisioned throughput capacity or other appropriate deployment quota for a high-volume model. This addresses “Provision Foundry token throughput capacity for stable peak inference”, where the governing principle is: Model serving capacity is consumed by tokens and request processing rather than just the number of subscribed users, and underprovisioned resources can throttle requests despite healthy application hosts. The stated case instead calls for: Use a Foundry-supported serverless or model-as-a-service API deployment for an eligible hosted model, or managed compute where the chosen supported model and runtime require controlled resources.
Use a Foundry-supported serverless or model-as-a-service API deployment for an eligible hosted model, or managed compute where the chosen supported model and runtime require controlled resources. Serverless hosted inference abstracts underlying model serving capacity while managed compute deployment exposes different cost, scaling and lifecycle responsibilities; not every catalog model supports every deployment type. Verify the result with this test: Deploy a small candidate using the documented supported option, verify its request shape, authenticated response and billing/scale behavior against the alternative in a pilot.
Commit prompt templates, parameter names, model deployment references and evaluation dataset identifiers to a reviewed Git repository, assigning immutable release identifiers rather than editing only in a live chat UI. This addresses “Version production prompt templates in Git with documented release identity”, where the governing principle is: Prompt text and variables are part of an AI application's behavior, so untracked instruction changes can alter safety and answer quality without any model binary change. The stated case instead calls for: Use a Foundry-supported serverless or model-as-a-service API deployment for an eligible hosted model, or managed compute where the chosen supported model and runtime require controlled resources.
Define approved fallback model deployments with compatible inputs, data handling and quality gates, and trigger fallback only on supported operational errors rather than masking every low-quality answer. This addresses “Design resilient model endpoint fallback without silently changing data policy”, where the governing principle is: A backup model may differ in region, cost, privacy conditions and output schema; blind fallback to an arbitrary public endpoint can violate residency or application validation assumptions. The stated case instead calls for: Use a Foundry-supported serverless or model-as-a-service API deployment for an eligible hosted model, or managed compute where the chosen supported model and runtime require controlled resources.
Question 342
An insurer wants a hosted sentiment model without maintaining serving VMs but its data scientist also evaluates an open-weight model with custom runtime requirements. The engineer proposes the following change: Use a Foundry-supported serverless or model-as-a-service API deployment for an eligible hosted model, or managed compute where the chosen supported model and runtime require controlled resources. Which mechanism explains why it works?
Answer & Explanations
Correct answer: C
A backup model may differ in region, cost, privacy conditions and output schema; blind fallback to an arbitrary public endpoint can violate residency or application validation assumptions. That explains “Design resilient model endpoint fallback without silently changing data policy”, not the mechanism behind the described change. The relevant principle is: Serverless hosted inference abstracts underlying model serving capacity while managed compute deployment exposes different cost, scaling and lifecycle responsibilities; not every catalog model supports every deployment type.
Prompt text and variables are part of an AI application's behavior, so untracked instruction changes can alter safety and answer quality without any model binary change. That explains “Version production prompt templates in Git with documented release identity”, not the mechanism behind the described change. The relevant principle is: Serverless hosted inference abstracts underlying model serving capacity while managed compute deployment exposes different cost, scaling and lifecycle responsibilities; not every catalog model supports every deployment type.
Serverless hosted inference abstracts underlying model serving capacity while managed compute deployment exposes different cost, scaling and lifecycle responsibilities; not every catalog model supports every deployment type. The corresponding implementation is: Use a Foundry-supported serverless or model-as-a-service API deployment for an eligible hosted model, or managed compute where the chosen supported model and runtime require controlled resources.
Model serving capacity is consumed by tokens and request processing rather than just the number of subscribed users, and underprovisioned resources can throttle requests despite healthy application hosts. That explains “Provision Foundry token throughput capacity for stable peak inference”, not the mechanism behind the described change. The relevant principle is: Serverless hosted inference abstracts underlying model serving capacity while managed compute deployment exposes different cost, scaling and lifecycle responsibilities; not every catalog model supports every deployment type.
Question 343
An insurer wants a hosted sentiment model without maintaining serving VMs but its data scientist also evaluates an open-weight model with custom runtime requirements. After implementing this change—Use a Foundry-supported serverless or model-as-a-service API deployment for an eligible hosted model, or managed compute where the chosen supported model and runtime require controlled resources.—which test most directly validates its intended outcome?
Answer & Explanations
Correct answer: A
Deploy a small candidate using the documented supported option, verify its request shape, authenticated response and billing/scale behavior against the alternative in a pilot. This check is decisive because serverless hosted inference abstracts underlying model serving capacity while managed compute deployment exposes different cost, scaling and lifecycle responsibilities; not every catalog model supports every deployment type.
Restore a previous tagged prompt release in staging, rerun its evaluation fixtures and confirm production traces identify the exact active prompt version. That would help test “Version production prompt templates in Git with documented release identity”, not the outcome required in this scenario. The relevant acceptance check is: Deploy a small candidate using the documented supported option, verify its request shape, authenticated response and billing/scale behavior against the alternative in a pilot.
Simulate a primary deployment rate limit in staging, verify only a preapproved compatible endpoint handles the request and confirm the returned format and audit trail remain valid. That would help test “Design resilient model endpoint fallback without silently changing data policy”, not the outcome required in this scenario. The relevant acceptance check is: Deploy a small candidate using the documented supported option, verify its request shape, authenticated response and billing/scale behavior against the alternative in a pilot.
Run load tests with realistic prompt and output lengths, measure tokens per minute, throttling and p95 latency, then verify the selected quota meets the traffic SLO. That would help test “Provision Foundry token throughput capacity for stable peak inference”, not the outcome required in this scenario. The relevant acceptance check is: Deploy a small candidate using the documented supported option, verify its request shape, authenticated response and billing/scale behavior against the alternative in a pilot.
Question 344
An insurer wants a hosted sentiment model without maintaining serving VMs but its data scientist also evaluates an open-weight model with custom runtime requirements. Which proposed shortcut most directly risks unavailable generative inference or unexpected infrastructure expense from the wrong deployment type and should therefore be rejected?
Answer & Explanations
Correct answer: D
Buy capacity based solely on a count of application users and claim large-context generations cannot exhaust provisioned token throughput. That shortcut risks rate limiting and latency spikes after misestimating peak model token workloads. Here the specific shortcut is: Assume every open-weight foundation model can be invoked through a serverless endpoint with no quota, account setup or supported deployment configuration.
When a Foundry deployment is throttled, send sensitive prompts to any public consumer chat API without checking location, model compatibility or approved data usage. That shortcut risks confidential data exposure and corrupted responses during model service failover. Here the specific shortcut is: Assume every open-weight foundation model can be invoked through a serverless endpoint with no quota, account setup or supported deployment configuration.
Edit the system prompt in production by hand and expect the team to reconstruct its prior wording from the model's chat history during an incident. That shortcut risks untraceable generative AI regressions after uncontrolled prompt changes. Here the specific shortcut is: Assume every open-weight foundation model can be invoked through a serverless endpoint with no quota, account setup or supported deployment configuration.
Assume every open-weight foundation model can be invoked through a serverless endpoint with no quota, account setup or supported deployment configuration. The foreseeable failure is unavailable generative inference or unexpected infrastructure expense from the wrong deployment type. The safer response is: Use a Foundry-supported serverless or model-as-a-service API deployment for an eligible hosted model, or managed compute where the chosen supported model and runtime require controlled resources.
Question 345
An insurer wants a hosted sentiment model without maintaining serving VMs but its data scientist also evaluates an open-weight model with custom runtime requirements. Which combined implementation and verification plan is the best match for this workload?
Answer & Explanations
Correct answer: A
Use a Foundry-supported serverless or model-as-a-service API deployment for an eligible hosted model, or managed compute where the chosen supported model and runtime require controlled resources. Confirm it through this test: Deploy a small candidate using the documented supported option, verify its request shape, authenticated response and billing/scale behavior against the alternative in a pilot. Serverless hosted inference abstracts underlying model serving capacity while managed compute deployment exposes different cost, scaling and lifecycle responsibilities; not every catalog model supports every deployment type.
The test would be relevant: Deploy a small candidate using the documented supported option, verify its request shape, authenticated response and billing/scale behavior against the alternative in a pilot. But the offered implementation—Estimate requests per second, token input-output mix, rate limits and desired service latency, then configure supported provisioned throughput capacity or other appropriate deployment quota for a high-volume model.—addresses “Provision Foundry token throughput capacity for stable peak inference”. The action needed here is: Use a Foundry-supported serverless or model-as-a-service API deployment for an eligible hosted model, or managed compute where the chosen supported model and runtime require controlled resources.
This action/test pair addresses “Design resilient model endpoint fallback without silently changing data policy”. Define approved fallback model deployments with compatible inputs, data handling and quality gates, and trigger fallback only on supported operational errors rather than masking every low-quality answer. Simulate a primary deployment rate limit in staging, verify only a preapproved compatible endpoint handles the request and confirm the returned format and audit trail remain valid. The case instead calls for: Use a Foundry-supported serverless or model-as-a-service API deployment for an eligible hosted model, or managed compute where the chosen supported model and runtime require controlled resources.
The action is applicable: Use a Foundry-supported serverless or model-as-a-service API deployment for an eligible hosted model, or managed compute where the chosen supported model and runtime require controlled resources. But its proposed test—Run load tests with realistic prompt and output lengths, measure tokens per minute, throttling and p95 latency, then verify the selected quota meets the traffic SLO.—addresses “Provision Foundry token throughput capacity for stable peak inference”. This requirement is confirmed instead by: Deploy a small candidate using the documented supported option, verify its request shape, authenticated response and billing/scale behavior against the alternative in a pilot.
Question 346
A research laboratory is comparing API-hosted and dedicated managed compute approaches for a specialized language model and needs a test of actual supported deployment types. Which Azure AI operations action best addresses this requirement?
Answer & Explanations
Correct answer: B
Record exact foundation model identifiers and deployment versions, use a controlled candidate deployment or routing layer, and promote after regression tests while retaining a supported prior version for rollback. This addresses “Pin model versions and use staged rollout for Foundry production endpoints”, where the governing principle is: Hosted model family names can map to changing behavior and capabilities, and routing different versions without a gate may alter response formatting or safety without an application code change. The stated case instead calls for: Use a Foundry-supported serverless or model-as-a-service API deployment for an eligible hosted model, or managed compute where the chosen supported model and runtime require controlled resources.
The question is set in this situation: A research laboratory is comparing API-hosted and dedicated managed compute approaches for a specialized language model and needs a test of actual supported deployment types. Use a Foundry-supported serverless or model-as-a-service API deployment for an eligible hosted model, or managed compute where the chosen supported model and runtime require controlled resources. Serverless hosted inference abstracts underlying model serving capacity while managed compute deployment exposes different cost, scaling and lifecycle responsibilities; not every catalog model supports every deployment type. Verify the result with this test: Deploy a small candidate using the documented supported option, verify its request shape, authenticated response and billing/scale behavior against the alternative in a pilot.
The question is set in this situation: A research laboratory is comparing API-hosted and dedicated managed compute approaches for a specialized language model and needs a test of actual supported deployment types. Estimate requests per second, token input-output mix, rate limits and desired service latency, then configure supported provisioned throughput capacity or other appropriate deployment quota for a high-volume model. This addresses “Provision Foundry token throughput capacity for stable peak inference”, where the governing principle is: Model serving capacity is consumed by tokens and request processing rather than just the number of subscribed users, and underprovisioned resources can throttle requests despite healthy application hosts. The stated case instead calls for: Use a Foundry-supported serverless or model-as-a-service API deployment for an eligible hosted model, or managed compute where the chosen supported model and runtime require controlled resources.
Assign the supported Azure managed identity and scoped RBAC roles for Foundry projects or deployed application workloads to access only intended model endpoints and backing services. This addresses “Use managed identities and RBAC for Foundry application connections”, where the governing principle is: An Azure managed identity receives time-limited resource tokens without embedded reusable application secrets, but those tokens still require data-plane role assignments for the specific service. The stated case instead calls for: Use a Foundry-supported serverless or model-as-a-service API deployment for an eligible hosted model, or managed compute where the chosen supported model and runtime require controlled resources.
Question 347
A research laboratory is comparing API-hosted and dedicated managed compute approaches for a specialized language model and needs a test of actual supported deployment types. The engineer proposes the following change: Use a Foundry-supported serverless or model-as-a-service API deployment for an eligible hosted model, or managed compute where the chosen supported model and runtime require controlled resources. Which mechanism explains why it works?
Answer & Explanations
Correct answer: C
The question is set in this situation: A research laboratory is comparing API-hosted and dedicated managed compute approaches for a specialized language model and needs a test of actual supported deployment types. Model serving capacity is consumed by tokens and request processing rather than just the number of subscribed users, and underprovisioned resources can throttle requests despite healthy application hosts. That explains “Provision Foundry token throughput capacity for stable peak inference”, not the mechanism behind the described change. The relevant principle is: Serverless hosted inference abstracts underlying model serving capacity while managed compute deployment exposes different cost, scaling and lifecycle responsibilities; not every catalog model supports every deployment type.
An Azure managed identity receives time-limited resource tokens without embedded reusable application secrets, but those tokens still require data-plane role assignments for the specific service. That explains “Use managed identities and RBAC for Foundry application connections”, not the mechanism behind the described change. The relevant principle is: Serverless hosted inference abstracts underlying model serving capacity while managed compute deployment exposes different cost, scaling and lifecycle responsibilities; not every catalog model supports every deployment type.
The question is set in this situation: A research laboratory is comparing API-hosted and dedicated managed compute approaches for a specialized language model and needs a test of actual supported deployment types. Serverless hosted inference abstracts underlying model serving capacity while managed compute deployment exposes different cost, scaling and lifecycle responsibilities; not every catalog model supports every deployment type. The corresponding implementation is: Use a Foundry-supported serverless or model-as-a-service API deployment for an eligible hosted model, or managed compute where the chosen supported model and runtime require controlled resources.
Hosted model family names can map to changing behavior and capabilities, and routing different versions without a gate may alter response formatting or safety without an application code change. That explains “Pin model versions and use staged rollout for Foundry production endpoints”, not the mechanism behind the described change. The relevant principle is: Serverless hosted inference abstracts underlying model serving capacity while managed compute deployment exposes different cost, scaling and lifecycle responsibilities; not every catalog model supports every deployment type.
Question 348
A research laboratory is comparing API-hosted and dedicated managed compute approaches for a specialized language model and needs a test of actual supported deployment types. After implementing this change—Use a Foundry-supported serverless or model-as-a-service API deployment for an eligible hosted model, or managed compute where the chosen supported model and runtime require controlled resources.—which test most directly validates its intended outcome?
Answer & Explanations
Correct answer: D
Send a fixed regression set through current and candidate deployments, inspect quality and latency, switch a small permitted traffic cohort and successfully restore the previous route. That would help test “Pin model versions and use staged rollout for Foundry production endpoints”, not the outcome required in this scenario. The relevant acceptance check is: Deploy a small candidate using the documented supported option, verify its request shape, authenticated response and billing/scale behavior against the alternative in a pilot.
The question is set in this situation: A research laboratory is comparing API-hosted and dedicated managed compute approaches for a specialized language model and needs a test of actual supported deployment types. Run load tests with realistic prompt and output lengths, measure tokens per minute, throttling and p95 latency, then verify the selected quota meets the traffic SLO. That would help test “Provision Foundry token throughput capacity for stable peak inference”, not the outcome required in this scenario. The relevant acceptance check is: Deploy a small candidate using the documented supported option, verify its request shape, authenticated response and billing/scale behavior against the alternative in a pilot.
Run a harmless test application under its deployed identity, verify the model or storage call succeeds and an unrelated resource rejects the same identity's attempt. That would help test “Use managed identities and RBAC for Foundry application connections”, not the outcome required in this scenario. The relevant acceptance check is: Deploy a small candidate using the documented supported option, verify its request shape, authenticated response and billing/scale behavior against the alternative in a pilot.
The question is set in this situation: A research laboratory is comparing API-hosted and dedicated managed compute approaches for a specialized language model and needs a test of actual supported deployment types. Deploy a small candidate using the documented supported option, verify its request shape, authenticated response and billing/scale behavior against the alternative in a pilot. This check is decisive because serverless hosted inference abstracts underlying model serving capacity while managed compute deployment exposes different cost, scaling and lifecycle responsibilities; not every catalog model supports every deployment type.
Question 349
A research laboratory is comparing API-hosted and dedicated managed compute approaches for a specialized language model and needs a test of actual supported deployment types. Which proposed shortcut most directly risks unavailable generative inference or unexpected infrastructure expense from the wrong deployment type and should therefore be rejected?
Answer & Explanations
Correct answer: D
The question is set in this situation: A research laboratory is comparing API-hosted and dedicated managed compute approaches for a specialized language model and needs a test of actual supported deployment types. Buy capacity based solely on a count of application users and claim large-context generations cannot exhaust provisioned token throughput. That shortcut risks rate limiting and latency spikes after misestimating peak model token workloads. Here the specific shortcut is: Assume every open-weight foundation model can be invoked through a serverless endpoint with no quota, account setup or supported deployment configuration.
Paste a global administrator service principal secret into every prompt template and assume the secret is required for the project to call models. That shortcut risks overprivileged generative AI workloads and credential leakage through application artifacts. Here the specific shortcut is: Assume every open-weight foundation model can be invoked through a serverless endpoint with no quota, account setup or supported deployment configuration.
Point all production requests at an unqualified latest model alias and remove the previous deployment before confirming format and safety compatibility. That shortcut risks unexpected assistant behavior with no safe rollback after an unreviewed foundation model change. Here the specific shortcut is: Assume every open-weight foundation model can be invoked through a serverless endpoint with no quota, account setup or supported deployment configuration.
The question is set in this situation: A research laboratory is comparing API-hosted and dedicated managed compute approaches for a specialized language model and needs a test of actual supported deployment types. Assume every open-weight foundation model can be invoked through a serverless endpoint with no quota, account setup or supported deployment configuration. The foreseeable failure is unavailable generative inference or unexpected infrastructure expense from the wrong deployment type. The safer response is: Use a Foundry-supported serverless or model-as-a-service API deployment for an eligible hosted model, or managed compute where the chosen supported model and runtime require controlled resources.
Question 350
A research laboratory is comparing API-hosted and dedicated managed compute approaches for a specialized language model and needs a test of actual supported deployment types. Which combined implementation and verification plan is the best match for this workload?
Answer & Explanations
Correct answer: B
This action/test pair addresses “Provision Foundry token throughput capacity for stable peak inference”. Estimate requests per second, token input-output mix, rate limits and desired service latency, then configure supported provisioned throughput capacity or other appropriate deployment quota for a high-volume model. Run load tests with realistic prompt and output lengths, measure tokens per minute, throttling and p95 latency, then verify the selected quota meets the traffic SLO. The case instead calls for: Use a Foundry-supported serverless or model-as-a-service API deployment for an eligible hosted model, or managed compute where the chosen supported model and runtime require controlled resources.
The question is set in this situation: A research laboratory is comparing API-hosted and dedicated managed compute approaches for a specialized language model and needs a test of actual supported deployment types. Use a Foundry-supported serverless or model-as-a-service API deployment for an eligible hosted model, or managed compute where the chosen supported model and runtime require controlled resources. Confirm it through this test: Deploy a small candidate using the documented supported option, verify its request shape, authenticated response and billing/scale behavior against the alternative in a pilot. Serverless hosted inference abstracts underlying model serving capacity while managed compute deployment exposes different cost, scaling and lifecycle responsibilities; not every catalog model supports every deployment type.
The action is applicable: Use a Foundry-supported serverless or model-as-a-service API deployment for an eligible hosted model, or managed compute where the chosen supported model and runtime require controlled resources. But its proposed test—Send a fixed regression set through current and candidate deployments, inspect quality and latency, switch a small permitted traffic cohort and successfully restore the previous route.—addresses “Pin model versions and use staged rollout for Foundry production endpoints”. This requirement is confirmed instead by: Deploy a small candidate using the documented supported option, verify its request shape, authenticated response and billing/scale behavior against the alternative in a pilot.
The test would be relevant: Deploy a small candidate using the documented supported option, verify its request shape, authenticated response and billing/scale behavior against the alternative in a pilot. But the offered implementation—Record exact foundation model identifiers and deployment versions, use a controlled candidate deployment or routing layer, and promote after regression tests while retaining a supported prior version for rollback.—addresses “Pin model versions and use staged rollout for Foundry production endpoints”. The action needed here is: Use a Foundry-supported serverless or model-as-a-service API deployment for an eligible hosted model, or managed compute where the chosen supported model and runtime require controlled resources.
Operationalizing Machine Learning and Generative AI Solutions (AI-300) Practice Tests
- Microsoft AI-300 Practice Test 33 - 10 Questions (Last Reviewed 20-Oct-2026)
- Microsoft AI-300 Practice Test 34 - 10 Questions (Last Reviewed 19-Oct-2026)
- Microsoft AI-300 Practice Test 35 - 10 Questions (Last Reviewed 18-Oct-2026)
- Microsoft AI-300 Practice Test 36 - 10 Questions (Last Reviewed 17-Oct-2026)
- Microsoft AI-300 Practice Test 37 - 10 Questions (Last Reviewed 16-Oct-2026)