Generative Answers in Agents

Overview

Authored topics give you precise control over specific conversations, but they only cover the questions you anticipated and took the time to build. For a typical government agent, that might be ten or twenty common scenarios. But users have hundreds of questions, and your organization already has the answers scattered across SharePoint sites, policy documents, and knowledge bases.

Generative answers bridge this gap. When a user asks a question that does not match any authored topic, the agent searches your connected knowledge sources and uses AI to generate a natural language answer grounded in that content. The user gets a helpful, relevant response, and you did not have to anticipate or script it in advance.

This video covers how to enable, configure, and control generative answers in your Copilot Studio agents.

What You’ll Learn

• What generative answers are: How AI creates responses from your knowledge sources
• Source configuration: How to connect and prioritize the right content
• Behavior controls: How to manage quality, scope, and risk
• Hybrid architecture: How to combine generative answers with authored topics
• Government considerations: Responsible AI, compliance, and data handling

Script

Hook: Beyond scripted responses

Authored topics cover the questions you anticipated. Your password reset topic handles password questions. Your leave request topic handles leave questions. But what about the dozens of other questions your users will ask? What is the telework policy? When is the next all-hands meeting? How do I submit a facilities request?

You could spend weeks building a topic for every possible question. Or you could let your agent generate answers from the content your organization already has.

Generative answers let your agent create responses from your connected knowledge on the fly. In the next eight minutes, you will learn how to enable, configure, and control this capability so your agent can answer the questions you never anticipated.

What are generative answers?

Generative answers are AI-generated responses based on your connected knowledge sources. When a user asks a question that does not match any authored topic, the agent does not simply say “I don’t understand.” Instead, it searches your connected knowledge sources, such as SharePoint sites, uploaded documents, and websites, for relevant content. It then uses generative AI to synthesize a natural language answer grounded in that content. The answer includes citations so users can verify the information by clicking through to the source document.

This is fundamentally different from authored topics. With authored topics, you control every word the agent says. The conversation follows a specific flow you designed. With generative answers, the agent creates the response dynamically based on the question and the available content. You control what sources it draws from, but the specific wording is generated by AI.

The value proposition is significant. Instead of building hundreds of topics to cover every question, you connect your knowledge sources and let generative answers handle the long tail of user inquiries. Your ten authored topics might cover sixty percent of interactions, and generative answers from knowledge sources cover most of the rest.

For government environments, this is grounded AI. The agent generates answers from your organization’s content, not from the open internet. It draws from the SharePoint sites, documents, and approved sources you connected. This is a critical distinction that separates it from general-purpose AI chatbots.

Configuring generative answer sources

Enabling generative answers starts with your knowledge source configuration. The quality of your generative answers depends entirely on the quality and relevance of your connected sources.

Navigate to your agent’s settings and ensure generative answers are enabled. Then review your connected knowledge sources. These are the same sources you added in the Knowledge section: SharePoint sites, uploaded documents, public websites, and Dataverse tables. Every source you connect becomes content the agent can draw from when generating answers.

Source selection matters. The agent searches connected sources when a user asks a question without a matching topic and uses the most relevant content it finds to formulate the response. If you have connected ten sources but only three are relevant to a particular question, the agent identifies and uses the relevant ones.

Best practices for source selection: use authoritative, well-maintained sources. Government policy sites that are regularly updated produce better answers than archived sites with outdated content. Avoid connecting sources that contain conflicting information because the agent may surface inconsistent answers. Keep sources focused and relevant to your agent’s purpose. If your agent handles IT support, connecting an HR policy site will dilute relevance and may produce confusing results for IT-related questions.

For government environments, prioritize .gov and .mil sources for any external content. Use your agency’s maintained SharePoint sites rather than general repositories. And before connecting any source, verify it is approved for the audience your agent serves. If your agent is available to the public, ensure the knowledge sources contain only publicly releasable information.

Controlling generative answer behavior

Enabling generative answers is straightforward. Controlling their behavior to meet government standards requires more deliberate configuration.

Content moderation settings determine how strictly the agent filters generated responses. Copilot Studio provides moderation levels that control how closely the generated answer must match the source content. Stricter settings mean the agent only generates answers when it has high confidence the source content supports the response. More permissive settings allow the agent to synthesize answers from less directly relevant content. For government agents, start with stricter settings and loosen them only after testing confirms the quality meets your standards.

You can control the scope of generative answers by specifying which topic areas they should cover. If certain categories of questions should only be answered through authored topics, such as anything involving PII collection or security procedures, configure your agent so generative answers do not attempt to address those areas.

Quality and accuracy require ongoing attention. Generative answers are only as good as your connected sources. If the sources contain errors, the generated answers will contain errors. Review the answers your agent generates regularly. Provide a mechanism for users to report inaccurate answers, such as a thumbs-up and thumbs-down feedback option.

Address hallucination risk directly. Grounding generative answers in your organization’s content significantly reduces the risk of fabricated information compared to open-ended AI. But it does not eliminate the risk entirely. Add disclaimers to generative responses: “This answer was generated from agency documentation. Please verify with the official policy for your specific situation.” Direct users to authoritative sources for any decisions with significant consequences.

For government-specific controls, use stricter moderation on any public-facing agent. Establish a review process before enabling generative answers for agents that handle sensitive use cases. And maintain audit trails to track what content is being generated and from which sources.

Combining generative answers with authored topics

The most effective Copilot Studio agents use a hybrid approach. Authored topics handle the interactions where you need precise control, and generative answers handle the broader range of knowledge questions.

Here is how they work together in practice. When a user sends a message, the agent first checks for a matching authored topic. If the message matches a topic’s triggers, that topic takes over and the conversation follows the flow you designed. If no authored topic matches, the agent falls back to generative answers and searches your connected knowledge sources. If neither authored topics nor generative answers can address the question, the Fallback topic catches it and provides escalation options.

This creates a three-tier response strategy. Design it intentionally. Author topics for critical, high-frequency interactions where you need structured flows, data collection, or specific step-by-step guidance. Password resets, software requests, and hardware issue reports are good candidates for authored topics. Use generative answers for broad knowledge questions where users need information from your documentation. Policy questions, procedure inquiries, reference lookups, and general information requests are ideal for generative answers. Customize the fallback for everything else, providing clear escalation paths to human support.

Here is an example architecture. An IT support agent has authored topics for password reset, software request, hardware issue, and VPN troubleshooting. These are the top five requests and each follows a specific process. Generative answers handle questions like “What is the acceptable use policy?” and “How do I configure my email on a mobile device?” and “What are the approved browsers?” These answers come from the IT knowledge base in SharePoint. The fallback catches everything else and routes to the help desk team.

Test the combined experience thoroughly. Verify that authored topics still trigger correctly and that generative answers do not override them. Test generative answer quality across different question types. Pay attention to the transitions: when a user goes from an authored topic interaction to asking a knowledge question, the experience should feel seamless.

Government considerations

Deploying generative answers in government environments carries specific responsibilities.

Responsible AI is not optional. Generative answers must be accurate and trustworthy. Include citations with every generated answer so users can verify the information. Add a standard disclaimer to generated responses that identifies them as AI-generated from agency documentation. This transparency builds trust and sets appropriate expectations.

Include generative answers in your agent governance plan. Define who approves the knowledge sources that power generated answers. Establish a process for monitoring answer quality over time. Create a review checkpoint before enabling generative answers for any new agent, especially those serving sensitive use cases or public audiences.

On data handling, generative answers in government clouds stay within your cloud boundary. Your content is used at query time to generate the answer. It is not used to train the underlying AI models. This is a key assurance for government stakeholders concerned about data handling, and it is one you should communicate clearly to your leadership and compliance teams.

Close: The best of both worlds

Let us recap. Generative answers extend your agent’s capabilities far beyond what authored topics alone can cover. They let your agent create natural language responses from your connected knowledge sources, handling the long tail of user questions that you did not anticipate or script.

Combined with authored topics, you get precision where it matters most and flexibility everywhere else. Authored topics handle the structured, repeatable interactions. Generative answers handle the broad knowledge questions. And the fallback catches the rest.

Here are your next steps. Enable generative answers on your agent if you have not already. Test with a range of questions that your knowledge sources should be able to answer. Review the generated answers for quality, accuracy, and appropriate citations. Add disclaimers and feedback mechanisms so users know the answers are AI-generated and can report issues.

Generative answers let your agent be as knowledgeable as your organization, without scripting every answer by hand.

Sources & References

• Boost conversations with generative answers — Official documentation for enabling and configuring generative answers in Copilot Studio
• Knowledge sources in Copilot Studio — Documentation for knowledge sources that power generative answers
• Microsoft Copilot Studio documentation — Comprehensive documentation hub for all Copilot Studio capabilities