Guide assumptions

1. Why you should read this chapter first

When you first encounter Copilot Studio, the expectation “What can I build with this?” is immediately followed by the question “Then can we build our system with it too?” This chapter clearly separates what Copilot Studio can and cannot do, and summarizes the boundary with custom development (SI) using practical criteria.

The conclusion first:

Copilot Studio is an “enterprise agent platform for building quickly and using safely,” not an “all-purpose AI development tool that does everything for you.”

If you understand this positioning clearly from the start, much of the later confusion around licensing, features, cost, and security is greatly reduced.

2. What Copilot Studio can do

Microsoft Copilot Studio is a low-code agent builder. With graphical tools, you can implement conversational AI agents without writing code, or with only minimal configuration. It is designed so agents can be built easily without data scientists or professional developers, while also allowing pro-code extensions when needed.

📖 Reference: Copilot Studio overview – What is an agent?

2.1 Natural-language conversational agents

An agent is an AI-powered conversational bot that understands users’ natural-language questions and provides appropriate responses. It is used in scenarios such as:

• FAQ bot: business hours, product information, and automated answers to frequently asked questions
• Internal guide bot: HR policies, IT manuals, and benefits guidance
• Customer support bot: order tracking, returns processing, and account inquiries
• Business automation bot: approval requests, email sending, and data lookup

Copilot Studio handles conversations in two ways:

1. Topics: pre-designed conversation flows following structured patterns such as “trigger phrases → questions → conditional branching → actions → responses”
2. Generative Answers: answers automatically generated by an LLM based on Knowledge sources

When Generative Orchestration is enabled, LLM-based multi-intent recognition runs instead of traditional NLU (Natural Language Understanding). It identifies multiple intents in a single utterance and automatically chains the appropriate topics, actions, and knowledge.

📖 Reference: Explore AI capabilities in Copilot Studio

2.2 Responses based on internal data (Knowledge)

When you connect Knowledge sources to an agent, it provides generative answers based on internal documents and data. It works through RAG (Retrieval-Augmented Generation), summarizing and validating retrieved content before responding with citations.

Supported Knowledge sources:

📖 Reference: Knowledge sources summary · Quotas and limits

When Work IQ (Tenant Graph Grounding) is enabled, Copilot Studio agents use Semantic Search when searching SharePoint. This can improve search performance. (Enabled by default.)

Also, when referencing SharePoint, the agent inherits the original data permissions as-is. Documents and sites that a user cannot access are not included in responses, so you get the same permission-management experience as the existing M365 environment without managing separate permissions for the agent.

2.3 Automation through Power Platform integration

• From agent tools, you can use flows to run automation workflows such as approval requests, email sending, and data recording directly from the agent through native integration with Power Automate.
• You can integrate internal systems through standard and premium connectors such as Outlook, Teams, SharePoint, Dataverse, SAP, and Salesforce.
• With Agent Flow, you can run Power Automate flows within the Copilot Studio context and return results to the conversation.

📖 Reference: Use actions with custom agents · Call Power Automate flows

2.4 External system integration

• Call external REST APIs through HTTP requests, or integrate with a wide range of services through more than 1,000 standard and premium connectors.
• Create custom connectors to exchange data with internal and external systems.
• Connect external MCP servers directly as tools through MCP (Model Context Protocol).

📖 Reference: Custom connectors overview · Connector reference

2.5 Advanced AI-powered capabilities

Copilot Studio provides a variety of AI capabilities beyond a simple chatbot.

2.6 Multi-channel deployment

Created agents can be deployed to multiple channels at the same time:

• Microsoft Teams — the most common internal deployment channel
• Website — iframe embed or custom web chat
• Microsoft 365 Copilot — exposed as an agent plugin in Copilot Chat
• Facebook Messenger, Slack, WhatsApp — B2C customer engagement
• Direct Line API — custom app integration

📖 Reference: Publish and deploy your agent

2.7 Multilingual support

• Copilot Studio supports many languages, including Korean.
• When creating an agent, you can set the primary language to create an agent that can recognize and respond in Korean.
• Korean reference materials (Knowledge) also work well.

3. Where Copilot Studio has limits

Copilot Studio is a powerful tool, but it is not suitable for every scenario. The following limits come from official documentation, practical experience, and internal customer cases.

3.1 Complex business logic processing

• If multi-step conditional branching, complex calculations, or large-scale data processing is required, it is difficult to implement with topic flows alone.
• In this case, you should separate the logic into external components such as Power Automate or Azure Functions.
• For long-running stateful behavior or complex transaction processing, custom development (Azure/SI) is recommended.

3.2 Advanced UI/UX customization

• Adaptive Card support is limited, and if you need custom rendering beyond basic text formatting, you need development that connects the frontend through a separate Agent SDK.
• Fine-grained developer control over API calls, actions, and response formatting is limited.

📖 Reference: Choose the right tool – Copilot Studio Pros/Cons

3.3 Data processing constraints (specific numbers)

📖 Reference: Quotas and limits for Copilot Studio

Additional constraints:

• Loop files cannot be used directly as Knowledge sources (currently Feature In Review).
• Classic ASPX pages, modern pages containing SPFx components, and custom CSS/accordion navigation sites in SharePoint are not used for Generative Answers.
• Questions that directly reference a filename (for example, “What does file-name.pdf say?”) cannot be answered.

3.4 Real-time bidirectional data synchronization

• Copilot Studio basically works in a request-response pattern.
• WebSocket-based real-time synchronization, event streaming, and similar mechanisms are not supported.

3.5 Source control and CI/CD

• Power Platform officially supports solution-based ALM, and you can configure CI/CD pipelines through Power Platform Build Tools (Azure DevOps) and GitHub Actions for Power Platform.
• With the Power Platform CLI (pac), you can unpack solutions into individual files, commit them to Git, and apply PR-based code review workflows.
• Using Pipelines in Power Platform, you can configure a no-code deployment pipeline across Dev → Test → Prod environments within the Admin Center.
• Through the Copilot Studio VS Code extension, you can edit agents locally as YAML files and integrate with Git by using pac copilot commands to push/pull.
• However, because Copilot Studio UI does not include built-in functionality for direct Git integration or PR creation, you must compose the tools above to build a separate ALM pipeline.

3.6 On-premises-only environments

• Because Copilot Studio is a cloud-based SaaS service, it cannot run in a fully on-premises environment.
• However, hybrid configurations using On-Premises Data Gateway and Azure vNET are possible.

📖 Reference:
Power Platform vNet setup and configuration
Hands-on: Integrating Copilot Studio with VNet

4. Boundary criteria: Copilot Studio ↔ custom development (SI)

4.1 Decision criteria table

4.2 Copilot Studio vs custom development (SI) vs Azure AI Foundry

The following summarizes the three options most often compared in practice.

Important: Copilot Studio is an agent productization tool, Foundry is an AI engine, and SI is system implementation. The three are complementary, not competing. Depending on who will develop the agent, the purpose, and the difficulty, you can select and combine them appropriately to enable agent development and operations at optimized cost.

4.3 Practical decision-making framework

Requirements analysis
  ├─ ① Can Studio cover 70% or more?
  │    ├─ Yes → Copilot Studio baseline + supplement the rest with PA/Azure
  │    └─ No → ② Is conversational UI the core requirement?
  │         ├─ Yes → Hybrid architecture (Studio frontend + Azure backend)
  │         └─ No → Custom development (SI)
  │
  ├─ ③ Is maintenance owned by business users?
  │    ├─ Yes → Prioritize Copilot Studio (autonomous citizen-developer operations)
  │    └─ No → Custom development is more likely
  │
  └─ ④ Is there a company-wide rollout plan?
       ├─ Yes → Studio (built-in governance and channel management)
       └─ No → Choose based on purpose

4.4 Feedback from real enterprise customers

The following summarizes feedback that repeatedly appears when enterprise customers in Korea evaluate adopting Copilot Studio.

👍 Positive feedback (common):

• Perceived PoC speed: “I have never seen an agent run this quickly”
• Business-side autonomous modification: “We can modify it directly without going through IT”
• Combination with M365 Copilot: “Having our dedicated agent attached to the existing Copilot Chat increases perceived value”

👎 Recurring issues:

• Difficulty understanding the licensing/permission structure
• Misunderstandings about premium connector warning messages
• Conflicts with network/security teams (CA policies, proxy settings)

❓ Frequently asked questions:

• “Can it call external systems too?” → Systems can be called by using connectors.
• “Will costs explode?” → M365 Copilot users get B2E free usage (conditions apply; explained in detail in Ch1)
• “Will it run automatically if a person does not approve it?” → Constraints must be clearly set in the instructions