Unlocking the Power of Custom Copilot Agents in Microsoft 365

As AI capabilities advance at a breakneck speed, businesses are increasingly viewing Microsoft 365 Copilot not as a mere built-in productivity assistant, but as a customizable platform. In our recent webinar, Bespoke AI: Customizing Microsoft 365 Copilot with Agents, our CEO, Michal Pisarek, and Microsoft MVP Andrew Connell shared key strategies and insights on crafting custom agents—declarative and custom-engine—that turn Copilot into a powerful, personalized tool. Read on for the rundown.

Understanding the Copilot Stack and Semantic Index

Copilot's architecture places an emphasis on the role of the semantic index—a dual-index system built from Microsoft Graph content that supports personalized, security-trimmed responses. It enables Copilot to ground its outputs in both public and organizational data, providing contextually relevant answers. Crucially, even a single Microsoft 365 Copilot license can activate this index tenant-wide.

While Copilot uses foundational models like GPT-4o, Microsoft's orchestrator governs the interaction between user prompts, organizational data, and the model itself. This layered architecture is what sets Microsoft 365 Copilot apart from standalone tools like ChatGPT.

Why Custom AI Agents Matter

Do you find default Copilot applications—like the one in Outlook—underwhelming? You're not alone. The real value emerges when you build agents tailored to your business scenarios. As Andrew put it, “Copilot becomes truly useful when it’s working with your data, your rules, and your workflows.”

Two main customization paths exist:

Declarative agents

• Built using JSON/YAML and Copilot Studio or Teams Toolkit, these allow you to define instructions, data sources, and actionable skills

Custom engine agents

• Full-stack implementations that let you control the foundational model and orchestrator, useful when working outside the Microsoft ecosystem or needing full autonomy

Injecting Knowledge and Skills into Copilot Agents

Andrew detailed two customization vectors:

Custom Knowledge

• Use Microsoft Graph Connectors to bring in data from systems like JIRA, Confluence, or secure environments not natively integrated with Microsoft 365. These connectors index external content into the semantic index, making it available for Copilot queries while respecting granular access controls.

Custom Skills (Actions)

• Define REST APIs that Copilot can call in real time—ideal for frequently changing or structured data. For example, an internal CRM system or a custom Azure AI search endpoint can provide up-to-date responses Copilot would otherwise miss.  

These enhancements also solve practical limitations. For instance, the semantic index only processes the beginning portion of large documents. A custom agent can bypass this by retrieving full content via custom APIs, enabling comprehensive document parsing.

Best Practices and Licensing Insights

Key operational advice included:

• Always have at least one M365 Copilot license. Keeping one on your tenant will unlock the semantic index and SharePoint Premium features. 
• Be strategic with connectors. Use Graph Connectors for static knowledge and actions for dynamic or secure integrations.  
• Manage your Swagger files smartly. Copilot Studio uploads entire API schemas, which may overwhelm the agent. Teams Toolkit offers more control, allowing selective exposure of skills. 

On the licensing side, it's important to note a few things: For one, dev tenants can't purchase Copilot licenses. Also, the new consumption model lacks grounding unless paired with a standard Copilot license. And finally, Power Platform licenses may be needed for agents writing to services like Dataverse. 

The Future of Agents: Adaptive, Autonomous, and Embedded

Microsoft’s pivot from "custom copilots" to "agents" hints at a more autonomous future—where agents act independently on behalf of users. Andrew previewed that upcoming features—likely to be announced at Microsoft Build—will allow more nuanced deployment and integration, including differential surfacing across Teams and other apps.

Custom Copilot agents represent a shift from AI convenience to AI empowerment. By tailoring agents with organizational knowledge and business-specific actions, companies can unlock transformative value from their Microsoft 365 investments.

As Andrew reminded us, building with these tools isn’t just about keeping up—it’s about leading the curve. Whether you're a developer or an IT leader, now's the time to start shaping Copilot into your organization's smartest assistant.

Want to learn more? Andrew's currently offering his 6-hour pre-recorded workshop, Build Declarative Agents for Microsoft 365 Copilot, for 20% off at https://vtns.io/orchestry-april-2025. And if you want to go even deeper, he'll also be hosting a 2-day Copilot developer workshop on building declarative agents—sign up here.