How to migrate from ChatGPT Business or Enterprise to Open WebUI

Moving a team from ChatGPT to Open WebUI is less a data migration than a product change. Very little is stored in ChatGPT that people need to keep, and what there is often cannot be exported. What people do have is habits, a set of custom GPTs, and an expectation of how fast and how good the answers are. This guide covers what can come across, how custom GPTs map onto Open WebUI, and the rollout order that keeps people using the new tool.

Why teams move from ChatGPT to a self-hosted interface

• Where prompts go. In ChatGPT, every prompt and every pasted document is processed by OpenAI under its terms. With Open WebUI on your own servers or in a region you choose, prompts to local models never leave that environment.
• Answers over your own documents. Open WebUI connects models to knowledge collections you control, with access decided by your directory groups.
• Choice of model. One interface can serve open-weight models on your hardware and, if you choose, hosted models through an API. The choice is made per model, by an administrator.
• Cost per seat. A self-hosted interface has no per-user licence. The cost is the hardware and the operations.

The trade-off is model quality at the top end, covered in step 3.

Step 1: inventory what your team uses ChatGPT for

Before exporting anything, find out what the tool is doing today. Enterprise workspace owners can export usage analytics as CSV from the workspace dashboard: users, GPTs and projects. On Business, ask team leads directly.

Write down:

1. The custom GPTs in use, who built each one, and who relies on it.
2. The features people depend on: file upload, image generation, web search, code execution, voice.
3. The documents people repeatedly paste or upload. These become your first knowledge collections.
4. Any GPT actions that call internal or third-party APIs.

This list decides the model sizing, the choice between Open WebUI and LibreChat, and what goes in the first release.

Step 2: export ChatGPT conversations, where you can

What you can export depends on the plan.

ChatGPT Business (formerly Team)

OpenAI states that data export is not available in ChatGPT Business workspaces. There is no self-service download of conversation history from settings. Tell people early, set a date, and ask them to copy out the conversations they genuinely want to keep. In our experience that is a small number: prompts that worked, and answers that became documents.

ChatGPT Enterprise

Enterprise workspaces do not offer self-service export either, but a workspace owner can enable the Compliance API with an admin key. It returns time-stamped records of conversations, uploaded files, workspace GPT configuration and users. Those records are not in the format Open WebUI imports, so they are converted into Open WebUI’s chat format with a script and loaded into each user’s account.

Decide whether you need this at all. Bulk-importing every conversation fills the new system with history nobody reads. Many teams import nothing and archive the Compliance API output for their records.

Personal ChatGPT accounts

People who used a personal account can export their data from Settings, Data controls. The zip contains a conversations.json file. In Open WebUI each user opens Settings, Data Controls and uses Import Chats on that file: ChatGPT exports are detected and converted automatically. Imports are added alongside existing chats, and importing the same file twice creates duplicates.

Step 3: choose and size the models

This step decides whether the rollout works. A private deployment that is slower or clearly worse than what people had is abandoned quickly.

• Size the hardware against the model. Which models your GPUs can serve, at what context length, for how many people at once. Answer this before installing anything. vLLM is built for many concurrent users; Ollama is simpler to run and suits a pilot or a small team.
• Be explicit about the gap. Open-weight models that fit on one or two GPUs are good at drafting, summarising, translation and answering over documents. They are weaker than the largest hosted models at long multi-step reasoning and complex code. Say so to users in the announcement.
• Decide on hosted models per use case. Open WebUI can connect to any OpenAI-compatible API alongside local models. Pilae can route such traffic to OpenAI, but only if you enable it; by default nothing leaves your environment.

Step 4: rebuild custom GPTs as Open WebUI models

There is no export button for a custom GPT. Each one is rebuilt, and the mapping is direct.

1. Copy the configuration. Open each GPT in the editor and copy its name, instructions, conversation starters and enabled capabilities into your inventory.
2. Collect the knowledge files from their source, not from the GPT. The originals in your document store are the ones that stay current.
3. Create a knowledge collection in the Open WebUI workspace and upload the files. Open WebUI chunks and indexes them for retrieval.
4. Create the custom model. Choose a base model, paste the instructions as the system prompt, attach the knowledge collection, add prompt suggestions, and grant access to the right groups.
5. Rebuild actions as tools. A GPT action is already described by an OpenAPI schema. Open WebUI can call an OpenAPI tool server directly, so the same schema usually carries across with new authentication.
6. Test with real questions from the people who used the GPT, and compare answers side by side. Instructions tuned for one model often need adjusting for another.

Timing matters. OpenAI announced in September 2026 that it plans to retire custom GPTs in favour of plugins, with Enterprise workspaces given the earliest dates. Check the notice in your own workspace, and copy each GPT’s configuration while the editor is still available. See OpenAI’s custom GPT retirement FAQ for the current schedule.

Step 5: connect sign-on and roll out

• Single sign-on first. Open WebUI supports OIDC, so accounts come from your directory through Keycloak or Entra ID. Groups decide who reaches which model and which knowledge. A leaver loses access with everything else.
• Start with one team that has a clear use case, usually one of the custom GPTs from step 1. Fix what they find, then widen.
• Run both for a set period, then close ChatGPT seats team by team. Keep the Compliance API archive, if you took one, under your own retention rules.

Open WebUI or LibreChat

Open WebUI keeps local models, knowledge collections and model administration in one interface, and suits teams whose main goal is private models over their own documents. It ships under the Open WebUI License: BSD-3-clause with one added condition, that above fifty users in any thirty-day period the project’s branding stays. LibreChat is MIT-licensed, also imports ChatGPT exports, and is strong where people move between several hosted providers and build agents with tools. Pilae operates either. For the wider picture, see our pages on ChatGPT Enterprise alternatives and private AI.

Let Pilae run the move

Pilae runs ChatGPT to Open WebUI migrations at a fixed price: the inventory and model sizing, the export and conversion where your plan allows it, the rebuild of your custom GPTs as models with knowledge, sign-on, and a staged rollout. Then we operate it on your premises or in Pilae Cloud, with daily encrypted backups and a monthly restore test. See migration services for what the fixed price covers, or start with one team.