If your team built a custom GPT to write proposals in the house voice, it is going to have to move. OpenAI plans to retire custom GPTs on December 11, 2026, and Google plans to start replacing Gems with skills in November. For a business wondering whether a configured assistant is enough or whether it needs an AI agent, this is a good moment to sort things out.
Both announcements say the same thing underneath: what you built in these tools (instructions, reference documents, a few connections) has value, but the container belongs to the vendor, who renames it or retires it on their own schedule.
The short answer, for the busy
A custom GPT (OpenAI), a Gem (Google) and a ChatGPT or Claude project all run on the same recipe: standing instructions, reference files and sometimes connected apps, shared across a team so everyone gets consistent answers. They are genuinely useful for a house style, a reference library or a prompt that comes back every week. In fall 2026 the container changes: OpenAI plans to retire custom GPTs on December 11, 2026 in favour of plugins, and Google will replace Gems with skills (from November 2026 for personal accounts, March 2027 for Workspace business accounts). At OpenAI, instructions and files migrate, but custom actions and sharing settings do not. A configured assistant works inside one person's account, with that person's access. An agent becomes the better choice when the task must start from an event, write into your systems, follow versioned rules and leave a record of every decision.
Four names, one recipe
A custom GPT, a Gem, a ChatGPT project and a Claude project are built from the same ingredients: standing instructions applied to every conversation (tone, structure, what to avoid), reference files the assistant consults, and sometimes tools such as web search or connected apps.
GPTs and Gems are assistants you call by name for a specific job. A project is a workspace that also keeps the conversation history, which suits a piece of work that runs for weeks. The volumes are modest: a GPT takes 20 files, a ChatGPT project between 5 and 40 depending on the plan.
One detail matters more than those limits: where the file comes from. A Gem that draws on Google Drive uses the most recent version of the file, and a ChatGPT project accepts links to Drive files or Slack channels. An uploaded file stays frozen on the day it was uploaded: the price grid you corrected the following month does not make its way in. Why those documents weigh more than the model is the subject of our article on assistants that answer from your own documents.
What changes this fall, and what survives
At OpenAI. Personal accounts (Free, Go, Plus and Pro) can no longer create new GPTs, and affected Enterprise workspaces lose that ability on October 26, 2026 (planned date). OpenAI plans to retire custom GPTs on December 11, 2026 (February 11, 2027 for Enterprise workspaces with an approved deferral) and offers to migrate them to plugins, which bundle skills (reusable instructions) with connected apps. The GPT's instructions become a skill; its files and connected apps come along.
Three things do not. Custom actions, the API connections someone configured, have to be rebuilt, sometimes with a custom MCP server. Sharing settings are lost: the migrated plugin starts private. And the selected model does not carry over either; OpenAI itself warns that a migrated plugin may respond differently.
At Google. Gems will go away in November 2026 for personal accounts, March 2027 for Workspace business, enterprise and non-profit accounts, and June 2027 for education. Google will convert them to skills automatically, along with supported files. A skill can be called from any conversation, and Gemini applies it on its own when it looks relevant. Sharing and Drive files, both things Gems could do, are announced for skills soon.
At Anthropic. Projects stay, but a new version in beta, limited for now to some Pro and Max subscribers, turns a project into one conversation that splits work into parallel threads running in the cloud. Anthropic also offers skills on every plan, which Team and Enterprise owners can push to the whole organization.
All three vendors are converging on the same object, a folder of instructions and resources the AI loads when a task calls for it, and Google even imports skills created on other platforms. Your instructions and documents are the asset; the GPT or the Gem is packaging. Keep the master copy of your instructions in a document you own, with a date and a named owner.
What a configured assistant does well for a team
Three uses justify the setup time. House style: proposals, replies to Google reviews and follow-up emails come out in the same voice. A reference library: the refund policy, the sheet for each service, the client onboarding procedure, all queryable in plain language. A repeatable prompt: the job posting written to the company template, every time you hire.
Without a configured assistant, every employee writes their own prompt and quality varies from one person to the next. With one, the team's best prompt becomes everyone's prompt.
Sharing is what makes that work. A ChatGPT project can be shared with chat or edit access, up to 100 collaborators on Business. At Anthropic, sharing a project requires a Team or Enterprise plan, with view or edit rights. A Gem can be shared with viewers or editors, with an expiry date if needed. Keep edit rights to one or two people: whoever can change the instructions changes the whole team's answers.
The data settings to check before you share
The account type. The general gap between personal accounts and business plans is covered in our comparison of Copilot, ChatGPT and Claude. Sharing adds a layer: on a personal ChatGPT account, a shared project can be used for training if the owner and every contributor have left "Improve the model for everyone" on, while content in a Business, Enterprise or Edu workspace is excluded by default. At Google it depends on your Workspace edition: where the Gemini app is a core service, chats and uploaded files are not reviewed by humans or used to improve models. Where it is only an additional service, consumer terms apply and human reviewers may read conversations.
Who can get in. On ChatGPT's personal plans, a project opened to "Anyone with a link" lets in any signed-in ChatGPT user who has the link. A Gem made with a personal account can even be made public.
What the assistant reveals. Google says plainly that a shared Gem's instructions and uploaded files are visible to anyone with access, and a Claude project shared with view rights shows its contents too. Put nothing in there you would not show every user, and certainly no personal information about your clients: that is shadow AI inside a tool you thought was under control.
Where a configured assistant stops
The line has moved. A ChatGPT plugin can now prepare a daily summary or check for updates every hour, and some respond to events such as a document changing. Gemini offers scheduled actions on some Workspace editions. Saying a configured assistant never works on its own is no longer true, and desktop agents go further still. The limits that remain are structural.
It runs on one person's access. A plugin uses connected apps under each user's own authorization, and every app keeps its own approval prompts. Good for security, but the process now depends on how each person is wired up.
It only writes where a connector exists. Your management software or in-house scheduling tool usually has no app in the vendor's directory. OpenAI says as much about the actions that will not migrate: rebuilding them may take a custom MCP server and technical setup.
Its behaviour changes without you. The vendor swaps the model, the migration drops the one you picked, and the same prompt returns a different answer. If nobody asks the same questions again after each change, the drift goes unseen.
It leaves no process record. A chat history does not tell you which version of the rules produced which decision. At Anthropic, audit logs are an Enterprise feature, absent from Team.
Where an agent becomes the better call
The useful question is what happens after the answer. If someone copies the assistant's output into another piece of software, that copy-paste is the real work, and it is an agent's job.
Four signs confirm it. The task should start from an event (an order, a form, a supplier email) rather than from someone opening a window. The result has to land in a system of record, such as the CRM or the calendar. The rules must be the same for everyone and changed by one person, with a record of the change. And you need to be able to show later what was done, under which rule.
An AI agent meets those needs by design. Its rules are written down, versioned and yours. It connects to your systems, including those with no off-the-shelf connector, and logs every step. The model is a component you can replace without rebuilding everything. If your business runs on Microsoft 365, the same line is drawn in our article on Copilot Studio.
The two coexist well: a configured assistant for the whole team's writing, an agent for the process that earns one.
Where to start
Before December 11, list the GPTs and Gems your team actually uses and who created each one: at OpenAI, only the creator (or, in an Enterprise workspace, an admin) can migrate a GPT. Copy each set of instructions into a master document, publish any pending edits (only the latest published version migrates), note the custom actions, since they will not migrate, and keep a few familiar prompts plus one hard case to compare before and after.
Then sort each assistant into one of three piles: migrate, drop, or turn into a process. The third pile is often the shortest, but it is the one that frees up real time, and it is the one we look at together in a first conversation.
→ First conversation, no commitment
Product names, retirement dates, limits and terms read on OpenAI, Google and Anthropic help pages on October 15, 2026. The vendors present them as subject to change: check the notice in your own account.
Written byXavier Peich