I have been trying to replicate the Claude Projects experience using Copilot Studio, but I have not yet been able to achieve comparable results.
My use case is to have one agent per project and a separate chat for each incident. Every chat should have access to the project’s shared context, including documentation, configurations, history, images, and previous decisions. This would allow the agent to assist with troubleshooting without requiring me to provide the same context again whenever I start a new incident.
Claude Projects handles this workflow very well. The project instructions and knowledge remain available across multiple conversations, while each conversation can focus on a specific incident. Files and images are also easy to add as part of the troubleshooting process.
I first looked at Copilot Notebooks. Although they allow users to collect sources and work with a defined set of content, they seem more focused on exploring, analysing, and summarising information. They do not appear to provide the same project-based workspace with multiple independent conversations sharing a common context.
I then created a dedicated Copilot Studio agent for the project, which seemed closer to what I needed. However, I have encountered several limitations. Response times are often slow, and answers can be incorrect or poorly grounded even when the relevant information exists in the configured knowledge sources. Context is not always maintained consistently across conversations, and there appears to be no straightforward way to copy and paste images directly into the chat for analysis. The interface also occasionally freezes or stops responding, making iterative troubleshooting less fluid.
Has anyone successfully implemented a Claude Projects-style workflow within the Microsoft ecosystem? I am particularly interested in maintaining persistent context for each project while keeping individual incident chats isolated. The project’s knowledge base should be available to every chat, but the history of one incident should not affect another.
Is there a recommended architecture using Copilot Studio, Microsoft 365 Copilot, Copilot Notebooks, SharePoint, Dataverse, Azure AI Search, or another Microsoft service? I would also like to know whether the slow response times, image limitations, and occasional freezing are known issues, and whether there are settings or architectural choices that can improve accuracy and reduce hallucinations.
This is not intended as a promotional comparison. I am genuinely trying to understand whether I am using the wrong tools, whether this requires a different architecture, or whether these are simply current limitations of the Microsoft platform.
Has anyone implemented this use case successfully? If so, what architecture and configuration would you recommend?