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.