The short answer.

Custom GPTs in ChatGPT, Projects in ChatGPT and Claude, Gems in Gemini and agents in Microsoft 365 Copilot all do the same job: they hold standing instructions and reference files, so everyone who opens one starts from the same method. The differences that matter to a company are how each is shared across a team, which files it holds and which business systems it can reach. Use the version inside the assistant your company already pays for; the idea transfers, so switching assistants for a different version rarely pays.

Makers rename and move these features often. Everything below was read on the makers’ own help pages in November 2026.

One idea, four names.

Every major AI assistant lets you save a set-up version of itself: a name, instructions it follows every time, and files it reads before it answers. Most people never open these features.

  • ChatGPT: GPTs and Projects.A GPT is a custom version of ChatGPT with instructions, knowledge files and a choice of built-in tools. A Project in ChatGPT groups related chats, files and instructions so ChatGPT keeps the same context across them, and colleagues can be invited into it.
  • Claude: Projects.A Project in Claude is a workspace with its own chat history, project knowledge (the files you upload) and project instructions that shape every answer inside it.
  • Gemini: Gems.A Gem is a customized version of Gemini with a name, instructions and knowledge files, which can include files from Google Drive.
  • Microsoft 365 Copilot: agents made in Agent Builder.An agent built in Agent Builder customizes Copilot with instructions and chosen knowledge sources for one task. Microsoft’s own page says it adds no new abilities beyond Copilot itself; it points Copilot at the right material.

A project is a room where chats and context build up, suited to ongoing work on one account or matter. A GPT, Gem or Copilot agent is a tool you open for one task that repeats the same way for everyone.

How each one shares across a team.

Sharing matters most: a set-up assistant only one person can open is a personal shortcut.

  • ChatGPT.Projects can be shared with people or groups, each given edit or chat access, or by link. GPTs are shared through the options your workspace allows.
  • Claude.Projects on a company account can be shared with named colleagues, with lists of people or with the whole organization, as view-only or edit.
  • Gemini.A Gem, or a conversation that used one, can be shared.
  • Copilot.An agent can be shared with named people as an owner who can edit or a user who can only chat, or with everyone in the organization. Publishing it to the organization’s Agent Store needs an administrator’s approval.

The advantage of Copilot’s model is control: a company already on Microsoft 365 gets owners, chat users and an approval step it can govern. The trade-off is ceremony: a shared Claude or ChatGPT Project is quicker to set up and change.

Files and connections to your systems.

Each set-up assistant holds reference files, and each reaches live business data differently.

  • ChatGPT.A GPT can use apps a person has connected, or actions a builder defines to call another system, though not both at once. A Project can take linked files from Google Drive and other apps as sources.
  • Claude.Connectors let Claude read from and act in the apps you use, and Claude inherits each person’s own permissions in the connected service. An administrator turns a connector on for the company and can allow reading while blocking changes.
  • Gemini.A Gem’s Drive files stay current: when the file changes, the Gem reads the newest version.
  • Copilot.An agent can draw on SharePoint and OneDrive, Teams chats and meetings, Outlook email, uploaded files and Copilot connectors to outside systems, which an administrator enables. Agent Builder itself does not add actions that change records elsewhere; for that, Microsoft sends you to Copilot Studio.

The trade-off to watch is uploaded files. In Copilot, anyone who can use an agent can get answers grounded in the files uploaded to it. Everywhere, an uploaded file is a copy that goes stale; a linked file stays current and keeps its own permissions. How AI workflows connect to your business systems covers the connections in more depth.

All four makers are moving toward instructions that load when needed.

The set-up assistant you switch to is giving way to written instructions the assistant picks up on its own when a task calls for them. All four makers now use the word skills for some version of this.

  • ChatGPT.OpenAI says custom GPTs are scheduled to retire on 11 December 2026, replaced by plugins. A GPT’s instructions become a skill inside a plugin, its knowledge files become reference files and its connected apps come across; custom actions and sharing settings do not, and a migrated plugin starts private.
  • Gemini.Google is moving Gems to skills, starting with personal accounts this month and business accounts later. Skills work inside your current chat, and several can be combined in one.
  • Claude.Skills are written instructions, with optional reference files, that Claude loads when the task needs them. Plugins bundle skills and connectors into one install, and a company’s owners can install plugins for everyone.
  • Copilot.Agent Builder now lets you add skills to an agent.

The practical lesson: keep the method in plain written instructions you own, in a document outside any one product. When a maker retires a feature, the instructions move with you, because the method is the part that took the work.

The advantages and trade-offs of a set-up assistant.

A set-up assistant is the first step past the blank chat box.

  • Advantage: the same start for everyone.The instructions and files are there before anyone types, so a new hire and a long-serving colleague begin from the same method.
  • Advantage: no new software.It lives in the assistant people already open, under the company’s account and its rules.
  • Advantage: easy to change.One owner edits the instructions and every user gets the change.
  • Trade-off: someone has to own it.A shared project nobody keeps current drifts, and people stop trusting it quietly.
  • Trade-off: it waits to be asked.A set-up assistant answers when someone opens it. Work that runs on a schedule, or across several steps without a person, needs the makers’ newer features.

Should we standardize on one AI assistant? explains why one company assistant makes all of this simpler.

What lives in the developer tools, not the chat.

Some of the most advanced features sit in the makers’ developer and coding tools, not in the everyday chat your team uses. In Claude, a plugin can include agents that Claude hands parts of a task to, and those run in Claude Code and Claude’s desktop work mode, not in ordinary chat. Some ChatGPT plugins work only in Codex, OpenAI’s coding tool, or only in the desktop app. In Microsoft’s world, connections that take actions in other systems are built in Copilot Studio, a separate builder. Gemini has its own version of this in Google’s developer tools. For most teams the first step is still a well-written set-up assistant in the ordinary chat.

Questions to answer before you build one.

Copy these and answer them for each one.

  1. What one task is it for?Name the piece of work in a sentence. One assistant per task.
  2. Who owns it?One person, by name, who edits the instructions and answers for them.
  3. Who uses it, and who can change it?List the people or team, and who gets edit access.
  4. Which files does it read?Link live files where you can; list any uploaded copies and when each is refreshed.
  5. Which systems does it reach, and can it change anything there?Read only, or read and write, with a person approving.
  6. Where do the instructions live outside the product?A document the company owns, so they survive a renamed or retired feature.
  7. Who checks the output before it leaves the team?One name.

This is what FullyPowered AI helps companies do: get far more from the AI they already pay for, by planning it, developing the workflows and deploying them across teams. The plan lives in FullyPowered Planner, and How organizations plan their use of AI sets out the method.