Microsoft Designing and Building Integrated AI Agent Solutions in Copilot Studio AB-620 Exam Questions

Page: 1 / 14
Total 93 questions

Want more questions? Get Premium Access.
()

Question 1

You create a test set in Copilot Studio to evaluate an agent that answers policy questions and retrieves data from connected knowledge sources. The evaluation threshold is set to 0.80.

You run the evaluation and review the following result for one test case:

Expected response: Digital products are non-refundable after download.

Actual response: Digital products cannot be refunded once downloaded.

Result: Pass

Score: 0.92

Reasoning: The actual response semantically matches the expected response.

Knowledge sources used: RefundPolicy_KB

Tools invoked: None

You need to determine whether additional corrective action is required based on the evaluation results.

Which two actions should you perform? Each correct answer presents a complete solution. Choose two.

NOTE: Each correct selection is worth one point.

ARerun the evaluation because the score is below 1.00.

BConfirm that the connected knowledge source is functioning as expected.

CInvestigate external tool configuration for this test case.

DModify the test case because the response is not an exact match.

EAccept the result because semantic matching is sufficient for a passing evaluation.

Answer : B, E

Comprehensive and Detailed Explanation From Microsoft AB-620 Study Guide: The test passed because the score of 0.92 exceeds the configured threshold of 0.80 and the evaluation method recognizes semantic equivalence. The actual answer conveys the same policy meaning as the expected response even though the wording is different, so an exact textual match is unnecessary. The recorded knowledge source is RefundPolicy_KB, which is the expected grounding path; confirming that source behavior is therefore appropriate. No tool was expected or invoked, so investigating tool configuration would add noise rather than address a defect. A score below 1.00 is not itself a failure, and rewriting the expected answer to mirror every valid paraphrase would undermine a meaning-based evaluation. The result can be accepted while still being retained as regression evidence. Reviewers should confirm that the cited or retrieved policy content supports the non-refund statement and that the chosen threshold reflects the business risk. For fixed codes or legally prescribed phrases, exact match might be preferable, but this natural-language policy answer is correctly evaluated by semantic similarity. Study Guide alignment: Test and manage agents > Evaluate agent performance > Review test results.

===============

Question 2

An agent in Copilot Studio must retrieve real-time data from an external system.

The external system exposes a REST API.

The API returns structured JSON data and requires authentication.

The agent must:

Call the API during conversations to fulfil user requests.

Retrieve real-time data from the external system by calling an authenticated REST API during a conversation.

The solution must use a supported mechanism that executes the REST API call, and the call must be explicitly configured to run at runtime.

You need to determine which approaches add the REST API to the agent.

Which two solutions meet the goal? Each correct answer presents a complete solution.

NOTE: Each correct selection is worth one point.

ACreate a global variable to store the API URL and call it by using a Redirect node.

BAdd the REST API as a tool and call it by using a topic or agent instruction.

CEnable generative answers and rely on the model to fetch the REST API data automatically.

DConfigure an HTTP action and invoke it from a topic during the conversation.

EAdd the REST API as a knowledge source so the agent can query it during conversations.

FCreate a topic variable and use a formula to parse the API endpoint string.

Answer : B, D

Adding the REST API as a tool is one supported solution. Copilot Studio can use an API definition to expose selected REST operations as callable tools. The tool can be invoked explicitly from a topic or selected through generative orchestration based on the tool description and agent instructions. Authentication, inputs, outputs, and completion behavior must be configured so the operation can execute during the conversation.

A Send HTTP request action in a topic is also a complete runtime approach. The author specifies the endpoint URL, HTTP method, authentication headers, request body, and expected response type. A sample JSON response can be used to generate a structured Power Fx record, enabling the topic to process the returned data and present it to the user.

Variables merely store values and do not execute network requests. Redirect nodes transfer control between topics, while formulas manipulate data but cannot independently call an API. Generative answers do not automatically discover and invoke arbitrary REST endpoints. A transactional or real-time API should be implemented as a tool or HTTP action rather than treated as a conventional indexed knowledge source.

Question 3

A company uses an agent in Copilot Studio to generate structured responses for internal users. The agent must meet the following requirements:

Consistently follow a reusable instruction template for response tone and structure.

Use an enterprise-approved foundation model for responses generated through the template.

You need to configure custom prompts so that the agent uses the Microsoft Foundry model catalog. Which two actions should you perform? Each correct answer presents part of the solution. Choose two.

NOTE: Each correct selection is worth one point.

ACreate a custom prompt template and apply it where the agent generates the structured response.

BPublish the agent to apply the Foundry model selection automatically.

CConfigure Azure AI Search as the grounding source so that the model choice is enforced.

DEnable citations so that the model is restricted to approved outputs.

ESelect a model from the Foundry model catalog for the custom prompt configuration.

FAllow topic authors to define tone and structure directly in topic steps.

Answer : A, E

A custom prompt provides a reusable instruction template that controls how the model performs a specific task. The template can define tone, output structure, formatting, constraints, examples, and fallback behavior. It can then be added as an agent-level tool, called from a topic, or used within an agent flow. This makes option A the correct mechanism for applying consistent instructions wherever the structured response is generated.

The prompt's model is configured separately in the prompt editor. The builder selects the required enterprise-approved model from the Model field or connects an approved deployment from the Microsoft Foundry model catalog. After that model is selected for the custom prompt, every execution of that prompt uses the selected model, making option E correct.

Publishing does not automatically select an appropriate Foundry model; the model must be explicitly configured. Azure AI Search controls retrieval and grounding, not model selection. Citations provide source traceability but do not limit execution to approved models. Defining instructions independently in individual topics would duplicate configuration and weaken consistency.

Question 4

A solution containing an agent in Copilot Studio is moved across environments. The agent relies on configuration values that differ between environments and are managed as environment variables.

You must ensure the following:

Configuration values are managed through environment variables that support ALM.

Runtime behavior reflects updates to environment variable values.

You need to implement environment-variable usage that supports ALM and ensures correct runtime behavior. Which three actions should you perform? Each correct answer presents part of the solution.

NOTE: Each correct selection is worth one point.

AAssume that all environment variable updates take effect immediately at runtime for all variable types without republishing.

BReplace environment variables with hard-coded values inside the agent so that runtime behavior is always consistent across environments.

CStore external reference values as environment variables so the agent solution can move between environments while keeping the agent definition the same.

DAfter an administrator updates a non-secret environment variable, republish any agents that use the variable so the change is effective at runtime.

EAllow builders to change environment variable values directly in Copilot Studio during topic authoring to speed deployment.

FUse environment variables as read-only in Copilot Studio and change their values in the appropriate administrative authoring experience.

Answer : C, D, F

Environment variables separate deployment-specific configuration from the agent's solution components. Therefore, external references---such as endpoint URLs, resource identifiers, and other environment-dependent settings---should be stored as environment variables. The same agent definition can then move through development, testing, and production while administrators provide the appropriate value in each target environment. This satisfies option C and is fundamental to Power Platform application lifecycle management.

Environment variables exposed inside Copilot Studio are read-only. Agent authors can consume their values but cannot edit them directly while authoring topics. Administrators or authorized makers must modify the current values through the appropriate Power Platform administrative or solution-authoring experience, making option F correct and option E incorrect.

When a non-secret environment variable is updated, an already published agent does not automatically begin using the new value. The agent must be republished so its published runtime configuration incorporates the change, which makes option D correct. Secret environment-variable values are handled differently because they can be retrieved dynamically at runtime.

Question 5

Users can add content to a paragraph field in a team-managed project, but they cannot @mention other users.

Identify the cause.

AGlobal permissions

BField settings

CIssue-level security settings

DNotification settings

EScreen settings

FProject permissions

Answer : A

Comprehensive and Detailed Explanation From [Atlassian Jira Cloud Documentation]

The cause is the Jira Browse users and groups global permission. Atlassian states that users with this permission can view user names when working with user-picker fields or @mentioning users. Without it, Jira cannot provide the affected users with the complete user directory required to search for and select another person through an @mention.

The fact that users can enter and save ordinary text in the paragraph field confirms that the field is present, editable, and available on the applicable work-item layout. Therefore, field settings and screen settings are not responsible. Notification settings determine how users receive updates after an event and do not control whether names appear in the mention picker.

Project permissions and issue-level security can affect whether a particular person can access the project or work item. However, a general inability to search for or mention other users indicates that the initiating users lack the site-wide user-browsing capability.

A Jira administrator should review Settings System Global permissions and grant Browse users and groups to an appropriate group containing the affected users, subject to the organization's privacy policies.

Relevant documentation topic: Jira Cloud administration Global permissions Browse users and groups. See What are global permissions, and what do they control?.

Question 6

You need to enable collaboration requirements for Operations Concierge

Which two actions should you perform? Each correct answer presents part of the solution Choose two. NOTE: Each correct selection is worth one point

AEnable allow connections from other agents on the Service Desk Agent.

BReplace the Service Desk Agent with an MCP server.

CPublish Operations Concierge to enable delegation automatically.

DAdd the Service Desk Agent as an indexed knowledge source

EPublish the Service Desk Agent

Answer : A, E

Comprehensive and Detailed Explanation From Microsoft AB-620 Study Guide: The Service Desk Agent must be both published and available for connections from other agents. Publishing establishes a runtime version that the primary agent can call; an unpublished draft is not a valid connected-agent target. On the Service Desk Agent, the setting that allows other agents to connect and use it must also be enabled. Microsoft additionally requires the connected agent to be in the same environment and the maker to own it or have it shared, but those conditions are not offered as choices. Replacing the agent with an MCP server would discard the existing agent's orchestration and ticket-routing logic. Adding it as a knowledge source would expose content for retrieval, not delegate an IT request to another agent. Publishing Operations Concierge alone does not satisfy the prerequisites of the target agent. After connection, the local description in Operations Concierge should clearly define when IT-specific work is delegated, because descriptions influence routing. Any later changes to the Service Desk Agent must be published before the main agent can use the updated version. Study Guide alignment: Integrate and extend agents in Copilot Studio > Configure multi-agent collaboration from Copilot Studio > Integrate an existing agent in Copilot Studio.

===============

Question 7

An agent in Copilot Studio is evaluated by using a fixed test set and an automated evaluation method.

After running the evaluation several times, the team observes that the same three interactions fail in every run, while all other interactions pass consistently.

You need to use the evaluation results to determine an accurate conclusion.

What should you conclude from the evaluation results?

AThe outcomes isolate which portions of the agent's functional scope require remediation.

BThe agent should not be deployed.

CThe agent's overall design is flawed and requires a full redesign.

DThe failing interactions are caused by an issue in the underlying language model.

Answer : A

When the same three test cases fail consistently while the remaining cases pass, the evaluation has isolated a repeatable weakness within a specific portion of the agent's functional scope. The appropriate next step is targeted triage and remediation of those interactions rather than redesigning the entire agent.

The team should first verify that each failing test case has a current expected response, realistic input, and an appropriate evaluation method. If the evaluation setup is valid, the builder can inspect the generated responses, transcripts, activity maps, knowledge retrieval, topic routing, tool selection, and tool parameters to identify the shared cause. The issue might involve a missing instruction, inaccessible knowledge, overlapping topic triggers, or an incorrectly configured tool.

The results alone do not prove that the underlying language model caused the failures. They also do not automatically establish that deployment must be blocked; deployment readiness depends on whether the failures affect safety, compliance, or critical business scenarios and whether organizational acceptance thresholds are met. Consistent passing results elsewhere indicate that the agent's entire design is not necessarily defective.