How to Rebuild a Custom GPT in Copilot Agent Builder
A step-by-step way to move a ChatGPT custom GPT into a Microsoft Copilot agent: what to copy, where each part goes, and what needs Copilot Studio.
OpenAI has scheduled custom GPTs to retire on December 11, 2026, across all ChatGPT plans (Enterprise workspaces with an approved deferral have until February 11, 2027). If your company runs on Microsoft 365, the natural new home for those GPTs is a Microsoft Copilot agent. There is no import button, so each GPT is rebuilt. The good news: for most GPTs that takes less time than it sounds, because the parts map almost one to one.
This guide walks through the rebuild in Agent Builder, the agent tool built into Microsoft Copilot, and shows where a GPT needs Copilot Studio instead.
What maps to what
| Custom GPT | Copilot agent | Notes |
|---|---|---|
| Name and description | Name and description | Copy as is, then tighten the description so people know when to use it |
| Instructions | Instructions | Agent Builder takes up to 8,000 characters |
| Knowledge files | Uploaded files or a SharePoint library | Up to 20 uploaded files; a SharePoint library keeps one current copy |
| Conversation starters | Starter prompts | Reuse the ones people actually click |
| Actions (API calls) | Copilot Studio connectors or MCP servers | Not available in Agent Builder |
| Sharing settings | Sharing with specific people, or organization-wide publishing | Organization-wide publishing goes through Copilot Studio with admin approval |
Step 1: Capture the GPT before you touch anything
OpenAI doesn't document a way to export a GPT's configuration or its knowledge files, so capture them by hand while the GPT still exists:
- Open the GPT's configuration view and copy the instructions into a text file.
- Copy the conversation starters.
- Note every action: what system it calls and what it does.
- Collect the original knowledge files from whoever uploaded them. They are usually sitting in someone's OneDrive or SharePoint already.
- Write down who uses the GPT and for what. That tells you who to test with and who to share the agent with.
Do this for every GPT your team depends on, not only the popular ones. The one a single person in finance uses every month-end is often the one nobody remembers until it's gone.
Step 2: Rewrite the instructions for Copilot
Most GPT instructions work in Copilot with light edits. Check three things:
- References to ChatGPT features. Lines like "use the code interpreter" or "browse the web" should describe the task instead, for example "calculate the totals from the attached rate card".
- References to knowledge files by name. If the files move to a SharePoint library, point the instructions to it in plain words: "Use only the pricing documents in the Sales Pricing library."
- Length. Agent Builder accepts up to 8,000 characters. If a GPT's instructions are longer, move reference material out of the instructions and into a knowledge file.
If you want a head start, paste a GPT's instructions into the chat on our ChatGPT to Copilot page and it rebuilds them as a Copilot agent you can download for Agent Builder. Our agent examples show six common GPTs side by side with their rebuilt agents.
Step 3: Build the agent in Agent Builder
- In Microsoft Copilot, open Agent Builder and create a new agent.
- Paste the name, description and rewritten instructions.
- Add knowledge: upload the files (up to 20), or point the agent to the SharePoint or OneDrive location where they live. Agent Builder can also use websites, Teams, Outlook and connectors as knowledge.
- Add the starter prompts.
- Test it with real questions from the people who used the GPT, side by side with the GPT's answers while you still have it.
A SharePoint library is usually the better choice for knowledge. People update one file in one place, and Copilot only shows content to people who already have permission to see it. That last point is also why it pays to check SharePoint permissions before you roll agents out: an agent can't overshare a file nobody overshared.
Step 4: Decide who gets it
Agent Builder lets you share an agent with specific people. To publish an agent across the whole organization in Teams and Copilot, you build or finish it in Copilot Studio, and an admin approves it.
Cost matters here too. Microsoft Copilot Chat comes with Microsoft 365 business subscriptions at no extra cost, and agents that use only instructions or public websites run on it for free. Agents that use company data such as SharePoint are metered in Copilot Credits for Copilot Chat users. Knowing which of your GPTs touch company data tells you which agents need a budget or the paid Microsoft Copilot license for their users.
When Agent Builder isn't enough
Move the GPT to Copilot Studio instead if it:
- Calls other systems. GPT actions become Copilot Studio connectors or MCP servers.
- Needs more than 20 files or many knowledge sources. Copilot Studio supports up to 500 per agent.
- Should be available to everyone from day one, through Teams and Copilot.
What doesn't come across
A few things have no direct equivalent, so plan for them:
- Chat history with the GPT. OpenAI says existing conversations stay accessible after the retirement, but only while you keep ChatGPT. Plan how to keep what you need before you cancel seats; Business workspaces have no self-service export.
- The GPT's selected model. It doesn't carry over, so judge the agent by its answers in testing, not by the model name.
- Personal memory. Copilot can't import memory from another AI tool. Each person sets their preferences again as Copilot custom instructions.
Doing this for dozens of GPTs
One GPT is an afternoon. Forty GPTs across five departments is a project: someone has to find them all, decide which are worth rebuilding, rebuild them, test them with the people who use them, and make sure the knowledge files land with the right permissions. That's the work our AI Migration Assessment starts with: an inventory of every GPT, a map of where each one lands in Microsoft 365, and a dated plan to get it done before December 11.