I’ve spent days troubleshooting an issue with an Employee Assessment Copilot app that uses Teams authentication and leverages data from SharePoint and across the enterprise M365 apps using the Graph API. The app is installed through the Teams Store for our company.
Initial Success in Testing Environment:
When running the original 8K prompt I had built for an Employee Self-Assessment prompt in Teams Copilot, the entire sequence executed as expected in the traditional Copilot AI sense.
I then took that over to Copilot Studio to build a Copilot App and began the process of breaking down the Prompt instructions into multiple Topics, each representing a segment of the overall flow. This allowed me to create a streamlined 1900-character objective and instructions, with a defined process flow, states and stages, a main loop for operational control, Error handling and global variable management in the Copilot Instructions.
The states and stages component is a logical process progression of Topics where each Topic redirects at the end > to the next Topic in sequence - based on when the user is ready to continue on from one stage to another.
The Main Loop in the instruction, watching over each Topic transition to make sure that everything happens cleanly. The User has complete control of the pacing in CoPilot app, and the app performs well as a Prompt, and in testing inside of Copilot Studio.
Core Problem - Environment Mismatch:
The main issue arises when deploying to Teams. The initialization logic behaves differently in the live Teams environment as compared to Testing in Copilot Studio (which basically runs the Copilot Instructions and Topic flow so they are visible to the developer, and you can test variables, choices, values along the way). There is no Copilot Simulator / Sandbox available (that I can find) that would allow the developer to run the Copilot and interact with it (using file uploads and other knowledge references to simulate an actual user interaction) as a User would see it LIVE.
Attempted Solution - CustomInitialize Topic
To address the issue, I created a CustomInitialize topic to centralize the initialization process after the Conversation Starter topic initializes the Copilot. I moved all the app’s initialization logic into this separate CustomInitialize topic, centralizing the setup of critical variables, authentication checks, and user guidance. This approach aimed to provide smoother transitions and a better user experience in Teams.
However, while this approach worked well in Copilot Studio, it failed in the Teams environment. Upon launching the app, a new chat window appears, but the Conversation Starter topic, CustomInitialize topic, and subsequent redirects don’t trigger. I even tried using trigger words to initiate the dialog, but nothing activated in the Teams app environment, despite everything working correctly during testing in Copilot Studio.
Other Details
Variable Initialization Issues
The Global.CurrentYear variable, crucial for constraining Graph API data analysis and guiding the user through their assessment, may not be initializing as expected in Teams. If this variable isn’t set correctly at the start, subsequent processes fail silently, causing the user to see a blank or unresponsive screen.
Authentication Challenges
Ensuring that the user is authenticated before proceeding with any data analysis or initialization is critical. Any delay or failure in authentication with Teams when the Copilot app opens could interrupt the intended flow. The user.token check needs to be seamlessly integrated with the initialization sequence to ensure the Copilot does not proceed until Teams authentication is confirmed.
Redirect Strategy
The new strategy redirects users from the Conversation Starter topic directly into CustomInitialize, simplifying the entry point and ensuring that all prerequisite conditions (like setting Global.CurrentYear and confirming authentication) are handled in one place before the Copilot progresses to the main assessment logic (Topics).
Fallback Logic
This made me realize the importance of having robust error handling to prevent the Copilot from becoming unresponsive due to silently failing initialization or authentication failures. This led to implementing retry mechanisms and user notifications to ensure that errors are communicated clearly, instead of leaving the user in a stalled state.
Adaptive Feedback
To enhance user experience, I included adaptive cards and confirmation messages, giving users a sense of progress as the initialization completed. This would keep users informed and engaged during any delays or necessary authentication processes.
I am looking for some advice on how to proceed from here. Right now, the Copilot is unusable because it’s just an app in Teams with a new User Chat message and no way to ‘start’ the Copilot instruction or the rest of the sequence. Has anyone experienced this and has some advice for me?
I’ve been teaching Prompt Engineering at work for about two years now. I’ve personally built about 400 prompts that our engineers, product development teams, Test Automation teams, and PMO use every day. This is not new to me by any means, but I’m simply missing something that’s probably obvious to some of you.
Thanks in advance,
Alan
2
Share
Report
All responses (
Answers (
