Limitations of Microsoft Copilot Studio
Verification note
This document was prepared from the provided draft and refined against Microsoft Learn references available as of 10 June 2026. Preview features, especially Work IQ MCP tools, can change and should be rechecked before production rollout.
1. Executive Summary
Microsoft Copilot Studio is a strong platform for building AI-powered conversational agents across Microsoft 365, Power Platform, Dynamics 365, Dataverse, SharePoint, Teams, and custom business applications. It can combine generative AI, topics, knowledge sources, tools, Power Automate flows, MCP servers, and connected agents.
The platform is powerful, but it is not an unlimited automation runtime, universal document parser, or fully deterministic orchestration engine. Enterprise teams must design around hard limits, connector behavior, authentication boundaries, user permissions, file-size constraints, preview feature risks, and ALM dependencies.
Enterprise takeaway
Most Copilot Studio limitations are not blockers. They are architecture constraints. When they are handled during design, agents can scale predictably. When discovered during UAT or production, they become delivery risk.
References: https://learn.microsoft.com/en-us/microsoft-copilot-studio/requirements-quotas, https://learn.microsoft.com/en-us/microsoft-copilot-studio/knowledge-copilot-studio, https://learn.microsoft.com/en-us/microsoft-copilot-studio/knowledge-unstructured-data
2. Scope and Assumptions
This guide focuses on standard Copilot Studio agents created and managed through copilotstudio.com. It is intended for agents deployed through Microsoft 365 Copilot, Microsoft Teams, web chat, Direct Line, Power Apps/custom applications, or related enterprise channels.
| In scope |
| Standard Copilot Studio agents |
| Microsoft 365 Copilot integration and Work IQ |
| Built-in Word, OneDrive, and SharePoint connector limitations |
| Agent flows, MCP, and connected agents |
The document intentionally separates connector actions from knowledge sources. This distinction is critical: a connector action performs an operation, while a knowledge source helps the agent generate grounded answers from indexed or retrieved content.
3. Quick Reference: Key Limits
| Feature | Limit | Enterprise design impact |
| Instructions for a Copilot agent | 8,000 characters | Prompt engineering must be modular; avoid putting the entire enterprise process in one instruction block. |
| Connector payload | 5 MB public cloud; 450 KB GCC | Large JSON, base64 files, binary content, and broad list responses must be filtered or summarized. |
| File upload size | 512 MB | Applies to file upload limits; actual grounding support depends on source type and license. |
| Uploaded files | 500 files | Does not apply to SharePoint used directly as a knowledge source. |
| Skills/tools | 100 per agent | Practical orchestration quality may degrade before the hard limit if tools overlap. |
| Topics | 1,000 per agent | Large topic libraries require strict naming and descriptions. |
| Trigger phrases | 200 per topic | Trigger phrases are not regex or deterministic business rules. |
| Dataverse knowledge | 2 sources; 15 tables per source | Use focused tables/views, not the full enterprise data model. |
References: https://learn.microsoft.com/en-us/microsoft-copilot-studio/requirements-quotas, https://learn.microsoft.com/en-us/microsoft-copilot-studio/advanced-flow-create
4. Platform, Orchestration, Topic, and Tool Limits
4.1 Generative Orchestration Is Powerful but Not Fully Deterministic
Generative orchestration allows the agent to dynamically select a topic, tool, knowledge source, child agent, connected agent, or direct generative answer. This improves user experience, but the selection is AI-driven, not a traditional if-else routing engine.
- Tool names and descriptions must be clearly differentiated.
- Similar tools can confuse routing, especially when they read related systems or folders.
- Long or conflicting instructions can reduce consistency.
- A user request with multiple intents may trigger the wrong path or require clarification.
- Knowledge source descriptions become important when many sources are available.
Design guidance
Treat every tool description like an API contract: what it does, when to use it, what inputs it needs, what output it returns, and when not to use it.
References: https://learn.microsoft.com/en-us/microsoft-copilot-studio/requirements-quotas, https://learn.microsoft.com/en-us/microsoft-copilot-studio/knowledge-copilot-studio, https://learn.microsoft.com/en-us/microsoft-copilot-studio/authoring-add-other-agents
4.2 Too Many Choices Can Reduce Routing Quality
Microsoft recommends considering separate connected agents when the main agent’s ability to differentiate between tools, topics, and agents’ starts to degrade. As a rule of thumb, this can happen around 30-40 choices, especially when descriptions are similar.
The hard tool limit is not the same as the recommended design limit. An agent may technically support up to 100 skills/tools, but routing quality can degrade earlier when tools overlap.
References: https://learn.microsoft.com/en-us/microsoft-copilot-studio/requirements-quotas, https://learn.microsoft.com/en-us/microsoft-copilot-studio/authoring-add-other-agents
5. Agent Flows Limitations
Agent flows are commonly used as Copilot Studio tools. They are best suited for deterministic actions such as querying Dataverse, reading metadata, creating records, sending messages, creating files, or calling APIs.
| Area | Limitation | Recommended design |
| Payload size | Returned data still falls under connector/action payload constraints. | Filter, paginate, summarize, and avoid returning raw file content. |
| Large processing | Reading many files, converting documents, or processing large Excel/PDF content can exceed the response window. | Move heavy processing to asynchronous flows, Azure Functions, Logic Apps, or document services. |
| Tool discoverability | A flow must use the appropriate agent for triggering and response action to be added as a tool. | Use clear flow names and descriptions in Copilot Studio. |
References: https://learn.microsoft.com/en-us/microsoft-copilot-studio/advanced-flow-create, https://learn.microsoft.com/en-us/microsoft-copilot-studio/flow-agent, https://learn.microsoft.com/en-us/power-automate/limits-and-config
6. Built-in Connector Limitations: Word, OneDrive, and SharePoint
This section covers the practical limitations most frequently seen when makers expose Microsoft 365 connectors as Copilot Studio tools. The biggest misconception is that file actions automatically allow the agent to understand document content. In many cases, file actions return binary content, which is useful for storage and transfer but not directly useful for reasoning.
Important distinction
Read file bytes is not the same as understand the document. For Word, PDF, Excel, and PowerPoint content, an additional extraction, knowledge, or parsing layer is usually required.
6.1 Word Online (Business) Connector
The Word Online (Business) connector is a Premium connector in Copilot Studio. It works with Word files in document libraries supported by Microsoft Graph, including OneDrive for Business, SharePoint Online Sites, and Office 365 Groups.
References: https://learn.microsoft.com/en-us/connectors/wordonlinebusiness/
6.1.1 Action: Create a Microsoft Word Document With the Given Content
This Copilot Studio-specific action creates a Microsoft Word file with provided content in the root directory and returns a Word document URL. It is useful for simple draft generation, but it is not a complete enterprise document assembly capability.
| Point | Documented behavior / limitation | Enterprise impact |
| Creates a Word file from content | Supported. The action accepts document content and an optional file name. | Useful for quick draft generation from an agent. |
| Target location | The action creates the file in the root directory; it does not expose a target SharePoint library/folder parameter. | For governed storage, move/copy the output to the correct SharePoint library after creation. |
| Output | Returns a Microsoft Word document URL. | Return the URL to the user or use a follow-up flow step to relocate the file. |
| Formatting | Better for simple HTML/plain text style generation than complex branded documents. | Use templates or custom document service for complex proposals, contracts, or reports. |
| Editing | The generated file opens in view/read-only mode in Word Online – to edit content, the user must open it in the Microsoft Word desktop application. | Use Word Work IQ (Preview) |
References: https://learn.microsoft.com/en-us/connectors/wordonlinebusiness/
6.1.2 Action: Populate a Microsoft Word Template
Populate a Microsoft Word template is one of the most important Word actions for enterprise document generation. It reads a Word template, fills supported content controls with dynamic values, and returns a generated Word document as binary.
| Area | Documented behavior / limitation | Enterprise impact |
| Template preparation | The template must be built with Word content controls. The Developer tab is not available on Word Online. | Templates should be authored and validated in the Word desktop client. |
| Storage location | Templates can be stored in Microsoft Graph-supported libraries such as OneDrive for Business, SharePoint Sites, and Office 365 Groups. | Maintain an approved template library with ownership and versioning. |
| Output | The action returns the populated Word document as a binary. | You must explicitly save the binary output to OneDrive, SharePoint, or another repository. |
| Supported controls | Plain Text, Combo Box, Drop-Down List, Image, and Repeating Section content controls. | Design templates for supported controls only. |
| Unsupported controls | Rich Text, Building Block Gallery, Date Picker, Check Box, and legacy form controls are not supported. | Avoid unsupported controls in production templates. |
| File size | The maximum input file size is 10 MB. | Large templates need to be simplified or generated through another service. |
| Unique names | Content control names must be unique. | Duplicate names can cause unintended duplicate replacement. |
| Images | Nested image content controls and images in headers are not supported. | Validate signature/logo/image scenarios early. |
| Editing output | Microsoft recommends using the Word desktop client when editing generated documents; generated content controls can be locked or greyed out. | Avoid promising an easy Word Online editing experience for template-generated files. |
| MFA conditional access | Populate a Microsoft Word template cannot be used if an MFA conditional access policy is enabled. | Validate tenant security policies before selecting this connector pattern. |
| Timeout | Documents must convert in less than two minutes, or the request times out. | Keep templates lightweight and add retry/error handling. |
References: https://learn.microsoft.com/en-us/connectors/wordonlinebusiness/
6.1.3 Action: Convert Word Document to PDF
The Convert Word Document to PDF action gets a PDF version of a selected Word file and returns PDF binary content.
- PDF conversion returns binary output, not automatically a saved file.
- Display Barcode field codes are not supported.
- Files marked with sensitivity labels such as Confidential or Highly Confidential are not supported for this action.
- Password-protected, digitally signed, or rights-managed documents can fail in common document conversion patterns.
- For production document generation, save the intermediate Word file, validate it exists, then convert and save the PDF separately.
References: https://learn.microsoft.com/en-us/connectors/wordonlinebusiness/
6.2 OneDrive for Business Connector
The OneDrive for Business connector is a Standard connector and is useful for personal file automation. It should not be treated as the main enterprise repository for shared process documents.
| Action/area | Limitation | Enterprise impact |
| Create file | Creates a file in the connected user’s OneDrive context. | Good for personal drafts; use SharePoint for governed team files. |
| Cross-drive | Cross-drive functionality isn’t supported; connector operates on the account used for the connection. | A flow that works for the maker can fail for other users or shared shortcuts. |
| Cross-tenant shared files | Accessing shared files across tenant hostname boundaries isn’t supported. | Avoid designing cross-tenant shared-folder automation on OneDrive. |
| Get file content | Returns binary content. | Not the same as text extraction or document understanding. |
| Triggers | Created/modified triggers skip files larger than 50 MB. | Do not use these triggers as the only control for critical large-file automation. |
| Copy file | Can time out for larger files depending on service load. | Add retry and completion checks. |
| Extract archive | Maximum archive size is 50 MB and 100 files inside. | Avoid using it for large migration or bulk extraction scenarios. |
| Upload from URL | Reports success after 20 seconds regardless of actual result. | Verify file existence and metadata before downstream steps. |
References: https://learn.microsoft.com/en-us/connectors/onedriveforbusiness/
6.3 SharePoint Connector
The SharePoint connector is usually a better fit than OneDrive for team-owned files, project folders, templates, approvals, and governed enterprise records. However, it still has file, list, trigger, permission, and conditional access limitations.
| Area | Limitation | Enterprise impact |
| Connector class | SharePoint is a Standard connector in Copilot Studio. | Usually preferred for shared enterprise content. |
| Get file content | Returns file content as a binary. | Use separate extraction/knowledge/parsing for content understanding. |
| Large files | For some Logic Apps scenarios, files over connector size limits need Graph API or intermediate storage. | Use Graph upload sessions or backend services for large files. |
References: https://learn.microsoft.com/en-us/connectors/sharepointonline/
7. Knowledge Source Limitations
Knowledge sources are used for grounding generative answers. They are not the same as connector actions. A connector’s action retrieves or changes data; a knowledge source helps the agent answer questions from indexed or searched content.
7.1 Uploaded Files, OneDrive, and SharePoint Upload Knowledge
| Area | Limitation | Enterprise impact |
| Uploaded files | Files are stored and indexed through Dataverse for semantic retrieval. | Good for controlled knowledge; not always real time. |
| Sync frequency | Connected files refresh automatically; OneDrive/SharePoint upload knowledge typically refreshes every four to six hours. | Do not expect immediate answers from newly updated files. |
| OneDrive limits | Up to 1,000 files, 50 folders, and 10 subfolder levels per source; 512 MB per file. | Avoid large, messy, personal OneDrive folders. |
| OneDrive shared folders | Shared folders are not supported as OneDrive knowledge sources. | Use SharePoint for shared enterprise knowledge. |
| Unsupported/protected files | Confidential/Highly Confidential sensitivity labels, password protection, unsupported types, or missing permissions can prevent answers. | Validate source readiness during UAT. |
References: https://learn.microsoft.com/en-us/microsoft-copilot-studio/requirements-quotas, https://learn.microsoft.com/en-us/microsoft-copilot-studio/knowledge-unstructured-data
7.2 SharePoint as a Knowledge Source
SharePoint knowledge retrieval has multiple modes. Some options copy selected files/folders into Dataverse and refresh every four to six hours; the SharePoint connector option can query SharePoint search infrastructure directly and support more advanced query filters.
| Area | Limitation | Enterprise impact |
| 500-file upload limit | The 500-file upload limit does not apply to SharePoint used directly as a knowledge source. | Large SharePoint repositories still need information about architecture discipline. |
| File size without M365 Copilot/Work IQ | Generative answers can only use SharePoint files under 7 MB for makers without the required M365 Copilot setup. | Large policy/SOW/proposal files may need splitting or Work IQ enablement. |
| File size with M365 Copilot + Work IQ | SharePoint and Microsoft Copilot connectors support larger files when Work IQ is enabled. | Better retrieval quality but possible latency increase. |
| SharePoint lists | Only the first 2,048 rows are returned for list queries. | Use Dataverse/API-backed tools for large operational tables. |
| List attachments | Attachments to list items aren’t indexed for reasoning. | Store important document content in document libraries or Dataverse/file stores. |
| File-name questions | Questions that reference only a file name may not always be answered reliably. | Ask content-based questions and use metadata/search tools for file discovery. |
| Guest users | Guest users have restrictions on internal SharePoint knowledge even with SSO-enabled apps. | Test external/guest access explicitly. |
References: https://learn.microsoft.com/en-us/microsoft-copilot-studio/requirements-quotas, https://learn.microsoft.com/en-us/microsoft-copilot-studio/knowledge-copilot-studio, https://learn.microsoft.com/en-us/microsoft-copilot-studio/knowledge-unstructured-data
7.3 Dataverse as a Knowledge Source
Dataverse knowledge is useful for structured enterprise data, but it is not intended to expose an unlimited relational schema to an agent.
- Maximum two Dataverse knowledge sources per agent.
- Maximum 15 Dataverse tables per knowledge source.
- The maker must have the required read permissions.
- Synonyms and glossaries have character limits and should be curated carefully.
- For complex relational queries, a Dataverse tool/flow/API is usually better than direct knowledge of grounding.
References: https://learn.microsoft.com/en-us/microsoft-copilot-studio/requirements-quotas, https://learn.microsoft.com/en-us/microsoft-copilot-studio/knowledge-copilot-studio
8. Work IQ and Standard MCP Limitations
Work IQ MCP tools extend Copilot Studio agents with Microsoft 365 context and actions. They can help agents interact with SharePoint, OneDrive, Word, and Microsoft 365 Copilot search. However, these capabilities are currently preview and must be treated with production caution.
Preview warning
Microsoft states that preview MCP features can have restricted functionality, and that tool names and parameters might change. Avoid hard-coded dependencies around preview tool names and schemas.
| MCP / Work IQ area | Documented behavior / limitation | Enterprise impact |
| Work IQ SharePoint | Preview MCP server for SharePoint sites, lists, libraries, files, folders, and sharing. File operations are limited to files 5 MB or less; folder listing returns top 20 by default. | Good for lightweight SharePoint operations; not a large-file processing engine. |
| Work IQ OneDrive | Targets the authenticated user’s personal OneDrive. Upload/download operations are limited to files 5 MB or less; folder listing returns top 20 by default. | Good for personal file operations; not team document management. |
| Work IQ Word | Creates new documents at the root of the user’s OneDrive; retrieves DOCX content from SharePoint/OneDrive sharing URLs; supports comments. | Useful for lightweight Word workflows, not full enterprise document assembly. |
| Work IQ Copilot search | Searches Microsoft 365 content such as SharePoint, OneDrive, email, Teams chats, PDFs, spreadsheets, presentations, and sites. | Powerful for discovery, but search ranking and user permissions affect results. |