AB-620 — frequently asked questions

DLP policies operate at the Power Platform environment level to control data flow permissions. They act as a foundational gatekeeper that must exist before the agent can securely interact with restricted data sources referenced in prompts.

No. Adversarial testing is a validation step used to check if your configuration works correctly against malicious inputs. It is not a configuration action you perform to set up the agent.

No. Prompts cannot control the underlying model selection. You must configure the approved model in the agent's settings via the Foundry catalog.

No. Azure AI Search is used for grounding responses with enterprise data, not for controlling which LLM model processes the prompt.

No. You must establish a valid connection first so the system can query and display the available tables.

The tables selected will not be linked to the agent, and the agent will not have access to the live enterprise data for grounding.

Use the 'Configure run after' property or specific error scopes to catch exceptions, then use actions like 'Send email' or 'Post message' to notify users before deciding whether to continue or end.

It leaves users confused with no feedback, breaking trust and making debugging difficult. Agents must always provide clear status updates or error messages.

Custom Connectors require explicit user interaction to invoke tools. Copilot Connectors are designed specifically for passive knowledge grounding where the agent automatically retrieves indexed data.

No. Copilot Connectors support various sources including SharePoint, OneDrive, and approved third-party systems that expose data via supported interfaces.

A Connector Knowledge Source queries live data from a connected application (like Dataverse) without copying it, whereas an Index copies data into a vector store for faster but static retrieval.

It uses the authenticated user's identity to query the source system, ensuring the agent only returns data the user is authorized to view, preserving original security boundaries.

Tools are for invoking specific actions or APIs. For grounding LLM answers in existing indexed data, the Knowledge feature is designed to handle retrieval and citation automatically.

No. Since the connector already indexes content into Microsoft Graph, adding the source links the agent to that pre-existing, permission-trimmed index without manual intervention.

Environment variables store configuration values that can be changed per environment during import, whereas custom connectors define the connection logic itself. Env vars keep the agent definition static.

Yes, you export the agent as a Solution package and import it into each environment (Dev, Test, Prod). You then map the environment variables differently in each target instance.

The case study explicitly requires avoiding manual edits to individual topics to ensure consistent tone and disclaimers across all generative responses efficiently.

No, Responsible AI filters are configured at the agent level in Copilot Studio. You do not enable them individually for each channel like Teams or Web Chat.

Adding disclaimers to each topic requires manual editing of every topic, which violates the requirement to avoid manual edits and ensures inconsistent application.

No, the technical requirements explicitly state that HTML is not supported and Markdown must be used for formatting such as bold and bullet points.

Technically, Graph connectors index Microsoft 365 data. However, in this exam context, Option A is the only 'connector-based' choice left after eliminating manual Q&A.

Yes, but this specific case study forbids it to ensure scalability and maintainability for large document sets.

A greeting only appears once when the chat starts. The requirement is for a disclaimer with every AI-generated response.

No. Disabling them removes the AI capability entirely, violating the need for scalable knowledge integration and self-service.

Anonymous access is allowed only for general inquiries that do not involve personal or sensitive data, such as flight status or public policies.

Internal staff authenticate via Microsoft Entra ID, ensuring secure access to internal support features and data within the Teams channel.

No, you must authenticate first. The system requires a valid connection to retrieve and display the available tables for knowledge grounding.

You can find it under the 'Knowledge' experience section of your agent builder interface, specifically by selecting 'Add knowledge'.

Validation ensures the flow executes correctly with sample data and handles errors gracefully. Skipping this step can lead to failures when the actual agent invokes the flow during conversations.

No. The invocation trigger defines how the agent calls the flow. While you might technically build actions in isolation, the standard preparation sequence requires defining the entry point (trigger) before finalizing the internal logic (actions).

Generative answers nodes allow granular control over which sources are used for specific topics, ensuring precise grounding rather than broad availability.

No, prompts guide the model but do not provide the factual data retrieval mechanism needed for grounding in approved content.

It only enforces HTTPS encryption. It does not verify who the user is or restrict access based on their organizational affiliation.

No. Personal accounts are intended for external or consumer-facing scenarios. Internal employees should use their organizational Microsoft Entra ID accounts.

No. In the Power Platform, agents must be contained within a solution to be exported as a package for ALM purposes.

A solution becomes managed when an unmanaged solution is imported into a target environment, such as Production.

A Retry Scope handles transient failures automatically but does not inherently trigger a notification or log detailed errors unless configured with a separate On Failure branch.

If retries are exhausted, the flow will still fail. Without an explicit 'On Failure' handler, the agent terminates silently, violating the prompt's constraints.

Publishing provisions the app registration and channel endpoint that produce the tenant ID, app ID, and bot ID. Until the agent is published and its channel is enabled, those configuration values do not exist to retrieve.

Validation runs after the SDK is implemented using the retrieved configuration and before the web app is deployed to production, so a failing conversation flow never reaches production.

A generative answers node produces a response from configured knowledge and performs no action, so it fits a purely explanatory step. A tool node would invoke an agent flow needlessly, adding latency and consuming API quota.

If the step must execute or reuse backend logic implemented as an agent flow, use a tool node. If the step only produces a natural language answer without running anything, use a generative answers node.

Azure AI Search grounds generative answers from the indexed policy library; the case explicitly delegates specialized summarization to Foundry, so grounding there would bypass the required architecture.

Citations govern how traceability renders on knowledge-based answers, addressing the case's traceability gap, but they neither select a summarization model nor create a reusable prompt template.

Publishing exposes the agent's API endpoint and metadata, making it discoverable and callable by other agents or external systems.

Adding registers the agent as a reusable component within the primary agent's design, while configuring connections sets up the specific communication channels for runtime invocation.

Publishing deploys the agent's current configuration as a version; it does not repoint generative answers to an approved model. Model and governance settings are configured separately and must be verified before publishing.

Citations add source traceability for audit evidence, and Azure AI Search controls the grounding/retrieval source. Neither constrains which model generates the response, so neither satisfies the approved-model requirement.

Publishing the primary agent (sender) is good practice, but the restriction preventing delegation is on the target agent. If the Service Desk Agent isn't published or connected, the call will fail.

No. MCP is for tool/API integration. The case study requires delegating functional requests (ticket creation) to an existing agent, which uses the Agent-to-Agent protocol.

The agent will not generate a response based on its internal training data, effectively refusing to answer to prevent hallucination.

No, you can ground against other data sources like SharePoint or Dataverse, but the question specifically mandates Azure AI Search.

You must add the tool first to create the object in the agent's canvas. Configuration settings like endpoints and authentication are properties of that specific tool instance.

No, testing should be the final step. You need to configure both the endpoint and authentication successfully to ensure the tool can communicate securely and correctly.

No, you can reuse existing connections if the same credentials apply. However, each tool instance requiring authentication must have a connection reference assigned.

The tool will appear in the topic editor, but the connector call may fail or return default/null values if required inputs are not bound to topic variables or user prompts.

It is a feature that enables agents to perform UI automation on desktop applications by simulating user interactions like clicks and keystrokes, useful for apps without APIs.

You can access the remote session within the testing pane to view the real-time execution of the agent's desktop interactions during test conversations.

If validation fails, the deployment process is halted before it reaches the target environment. This prevents broken solutions with missing connections from being deployed to production.

While you can set environment-specific variables, the requirement for a 'single, unified location' refers to Centralized pipeline configuration, which manages the overall pipeline structure and approval workflows centrally.

It indicates the binary Pass or fail status of a specific test case, determining if the agent's output satisfied the predefined success criteria.

Intelligent agents can be non-deterministic. Repeated failures or varied results across runs indicate instability or intermittent outcomes that require troubleshooting.

You must have an existing Application Insights resource to serve as the telemetry sink. The agent configuration points to this resource ID; it cannot create the resource itself during the monitoring setup.

No. Monitoring settings are applied when the agent is published. You must publish a new version of the agent for the telemetry to start flowing to Application Insights.