A Copilot Studio agent uses Classic mode with generative answers enabled and a company SharePoint site as the knowledge source. During testing, the developer notices that when a user asks a question completely outside the SharePoint content — such as asking about the weather — the agent generates a plausible but fabricated answer using its built-in AI knowledge rather than returning 'I don't know.' How should the developer prevent the agent from generating answers outside its authorized knowledge source?
Select an answer to reveal the explanation.
Short Explanation
Turn off the generative answers setting that lets the model use its own general knowledge. That keeps your agent inside its lane, so it answers only from the SharePoint source and falls back when there's no match.
Full Explanation
In Copilot Studio's generative answers configuration, there is a setting that controls whether the model can supplement knowledge source content with its own built-in general knowledge. When this setting is enabled (it may be labeled 'Allow the AI to use its own general knowledge,' 'Use Azure OpenAI knowledge,' or similar depending on the Copilot Studio version), the agent can generate answers about topics not covered by the configured knowledge sources, which causes the fabrication issue described. Disabling this toggle constrains the generative answers model to only produce responses grounded in the configured knowledge sources, such as SharePoint. If the SharePoint content does not contain relevant information for a query, the generative answers node returns 'no answer found,' which then triggers the configured fallback behavior rather than hallucinating an answer. The approach of disabling generative answers entirely and relying only on pre-built topics is incorrect because it is an overreaction that removes a valuable capability. The correct fix is a configuration toggle, not removal of the feature, and pre-built topics alone cannot cover all HR or company knowledge questions. The approach of switching the agent to Orchestration mode is incorrect because Orchestration mode does not automatically restrict answers to configured knowledge sources only. Orchestration mode determines how topics and actions are selected; the general knowledge restriction is a separate configuration on the generative answers node that applies in both modes. The approach of adding a content moderation topic that intercepts off-topic questions before they reach the generative answers node and redirects them to a fallback message is incorrect because it is complex, fragile, and incomplete. It requires the developer to enumerate what constitutes 'off-topic,' which is impossible to do comprehensively. The correct solution is a single configuration toggle on the knowledge grounding setting. Exam tip: The 'allow general knowledge' toggle is a key data governance control for enterprise agents. Know that turning it off forces the agent to ground all answers in the configured knowledge sources, preventing hallucination of out-of-scope content.