Explore AI capabilities in Copilot Studio

AI capabilities are what make Copilot Studio more than a scripted chatbot. With generative AI at its core, Copilot Studio isn't designed to be heavily authored, topic by topic. Instead, it's focused on authored business-critical experiences that work side by side with AI-driven knowledge, orchestration, and automation. This approach means agents no longer rely on hundreds of rigid topics to cover every possible question. Instead, they combine curated topics with flexible generative responses grounded in trusted sources.

Summary of AI capability options

Capability	What it enables	Limits or considerations
Generative orchestration	Multi-intent recognition, dynamic chaining of topics, knowledge, and actions	Limits on orchestration length; careful testing needed for complex flows
Generative answers	Dynamic answers from connected knowledge sources, with summarization and citations	Accuracy depends on data quality; requires planning for follow-up questions
Generative builder	Natural language topic creation and updates	Human validation required; generated dialogs might need refinement
Computer use	Operate a computer by using natural language discussions	Need to provision and maintain a machine for the agent to use
AI prompts	Generate a response from an AI model given a set of instructions	Performance and capabilities vary based upon the chosen AI model
AI approvals and other AI capabilities	Decision automation, sentiment triggers, transcript analysis, AI Builder integration	Still emerging features; ensure governance and monitoring are in place

Generative actions and orchestration

By using generative orchestration, agents can automatically chain together topics, actions, and knowledge sources. This approach replaces the traditional Natural Language Understanding (NLU) model with one powered by large language models that can recognize multiple intents in a single utterance and extract multiple entities, even of the same type. When paired with actions, agents can call APIs, run agent flows, or trigger connectors. The agent prompts the user only for missing details and then summarizes outputs. This approach allows agents to not just answer, but also act across systems.

Learn more: Apply generative orchestration capabilities

Generative answers

By using generative answers, agents can respond dynamically by searching over knowledge sources such as Dataverse, SharePoint, OneDrive, public websites, or other custom data. Rather than hand-building FAQs, you configure which sources the agent should use, and the AI generates answers. It summarizes retrieved content, validates results, and even generates citations where possible. You can place generative answers in multiple points of a topic flow or in the Conversational Boosting topic before Fallback. This approach gives broad coverage without the maintenance overhead of manual topics.

Generative builder

Authoring agents is no longer limited to manually defining every message or dialog path. The generative builder lets you describe what you want in natural language and the AI helps build the agent for you. This approach reduces the time to create and update topics, provides suggested variations to make the agent feel more natural, and can even help produce adaptive cards that consolidate information. While human review is always needed, the generative builder accelerates design for both technical and business users.

Computer use

Computer use enables an agent to perform actions on a computer by visually interpreting the screen and carrying out tasks step‑by‑step, without requiring APIs or scripted automation. When invoked, the agent analyzes images of the active window, determines the next action, and continues until the task is complete. Computer use is helpful for integrating legacy applications, desktop‑only workflows, and multi‑application processes that lack modern integrations. This capability extends Copilot Studio beyond conversational logic, enabling real task automation across web and desktop environments where traditional connectors or HTTP calls aren't available.

Learn more: Automate desktop processes by using the computer use tool

AI prompts

AI prompts let makers directly control how an AI model should interpret an instruction, format a response, or perform a reasoning task. Unlike the orchestrator, AI prompts give you full control over the model's input, including tone, constraints, examples, and structured output formats such as JSON. Prompts can use a wide range of models, including Microsoft Foundry models, and support advanced capabilities like Dataverse grounding and code interpreter when enabled. This control makes AI prompts ideal for scenarios that require precise language generation, structured data extraction, or highly customized outputs that would be difficult to accomplish through generative orchestration alone.

Learn more: Generate a response by using AI prompts

Other AI features

Beyond conversational generation, Copilot Studio continues to expand into adjacent AI scenarios. For example:

• By using multistage and AI approvals in agent flows, the agent can automatically evaluate and decide on approval requests by using predefined criteria.
• Sentiment analysis, where negative tone can trigger escalation to a live agent.
• Transcript analysis for new topic discovery.
• Integration with AI Builder for image recognition and text extraction.

Together, these features extend agents beyond dialog into decision automation and content understanding.

Responsible AI and compliance

All generative AI features in Copilot Studio run on Microsoft Foundry with Microsoft's Responsible AI standards. The system grounds responses where possible, validates them to reduce incorrect information, and moderates harmful or noncompliant content. The system temporarily stores conversations only for operations support, and Customer Lockbox can control any support access. Agents don't contribute tenant data to model training, and environment settings allow admins to disable features that require data movement across regions.

Learn more: Responsible AI FAQs for Copilot Studio