An IT administrator at an engineering firm gets one question from the board: "ChatGPT or Copilot? Pick one and we switch the rest off." She has looked at both, knows that half her colleagues already use a personal ChatGPT account and that the Microsoft 365 licences already have Copilot Chat switched on, and she suspects the question is the wrong one.
It is. The biggest difference is not between the two tools but between the accounts they run under. Only once that is settled does a real difference remain: what each assistant can see, who manages that, and where the work of keeping it safe sits.
Which account decides the data protection?
Both services come in variants that share a name and have very different terms. ChatGPT exists as a personal account (Free, Plus, Pro), as a Business workspace and as an Enterprise or Edu environment. Copilot exists as a consumer version without a work account, as Microsoft 365 Copilot Chat with a work account, and as the paid Microsoft 365 Copilot.
The dividing line does not run between the brands; it runs straight through both. Personal ChatGPT accounts can use chats to improve the model unless the user turns that off, and the organisation has no say in it. For Business, Enterprise and Edu, OpenAI states that workspace data is not used for training by default. The consumer Copilot has no business data protection; the signed-in variants do, and Microsoft states that prompts, responses and Graph data there are not used to train the underlying models.
So whoever asks "ChatGPT or Copilot" is usually comparing a personal account with a purchased licence, and that is not a comparison between tools. Why that difference works the same way for every AI service is in free versus business AI accounts.
What does each assistant see?
This is the difference that remains once the accounts are equal, and it is a big one.
ChatGPT sees what someone hands it: the pasted text, the uploaded file, the screenshot. Memory can carry context over from earlier conversations, and a shared link makes a conversation readable for anyone who has the link. The reach is therefore as wide as the prompt, and the risk sits exactly there: which customer details, case files or contract clauses ended up in that text box. The six prompts that go wrong most often are in ChatGPT and work data.
Microsoft 365 Copilot sees what the user is allowed to open. It searches email, calendar, Teams, SharePoint and OneDrive via Microsoft Graph, with exactly the permissions of the person asking. It grants no new access, but it makes years of oversharing findable with one question in plain language. Copilot Chat, the variant without the paid licence, does not have that reach and works with uploads and what is on screen. Why inheriting permissions is the whole Copilot privacy story is in Microsoft Copilot inherits your access.
What that means for the work: with ChatGPT the risk sits in behaviour at the moment of typing, with Copilot in a permission structure that grew over years. That is why Microsoft puts cleaning up those permissions first in its own deployment guidance.
How do connectors and administration differ?
| ChatGPT | Microsoft Copilot | |
|---|---|---|
| Business variants | Business, Enterprise, Edu | Copilot Chat (work account), Microsoft 365 Copilot |
| Training on input | Personal: possible unless turned off; business: excluded by default | Consumer: no business protection; signed in: no training of the underlying models |
| Reach | What is pasted, uploaded or handed over as a screenshot | Copilot Chat: uploads and screen; Microsoft 365 Copilot: everything the user may open via Graph |
| Connectors | GPTs, apps and connectors; off by default in Enterprise and Edu, managed by workspace owners | Connectors, agents from Copilot Studio, plug-ins; who switches them on has to be recorded somewhere |
| Administration | Business: members, apps, GPT settings; Enterprise and Edu: SSO, SCIM, roles, retention, residency | Licences per user, Purview sensitivity labels, DLP for Copilot, Restricted SharePoint Search as a temporary brake, flex routing |
| Processing in the EU | Depends on how the Enterprise environment is set up | Storage within the EU Data Boundary; flex routing can move processing outside the EU at peak load, can be turned off |
| Where the work sits | In the prompt | In the permissions |
Connectors widen what either assistant can read, and for both the question is not "on or off" but who switches them on and where that is recorded. With ChatGPT those are GPTs, apps and connectors; with Copilot they are connectors, agents and plug-ins that sometimes arrive with a monthly update nobody asked for.
The administration differs in kind. ChatGPT Enterprise gives control over the workspace: who is in it, which apps are on, how long conversations are kept. Administrators do not automatically get access to users' private chats. Copilot administration is mostly about the environment around it: sensitivity labels that keep encrypted content out of Copilot, DLP rules, and the permission reports in SharePoint Advanced Management. With Copilot you mainly manage what the assistant may find; with ChatGPT mainly who uses the workspace and how.
When does which one fit?
Microsoft 365 Copilot fits when the work lives in Microsoft 365 and the organisation wants the assistant to search there, and only if the permissions on SharePoint and OneDrive have been cleaned up before the licences go out. Without that step you are buying a search engine for your own overshared folders.
ChatGPT Business or Enterprise fits when the work consists of text and documents that someone supplies themselves, when there is no Microsoft 365 environment, or when people simply work better with it. The reach is then limited to the prompt, which makes the risk easier to oversee but not smaller: everything depends on what gets pasted.
Personal ChatGPT accounts and the consumer Copilot fit no work involving personal data at all. That is the one rule that reads the same in both columns.
What if both are in use?
For most organisations that is the reality. The engineering firm from the opening has Copilot Chat switched on and a group of colleagues used to ChatGPT, and neither disappears because of a choice on paper. The useful question is not which tool wins but which variant of each gets a status: allowed, limited use, no decision yet or not allowed. Personal ChatGPT on not allowed, with ChatGPT Business or Copilot Chat as the alternative, is a different decision from "ChatGPT banned", and it keeps the use visible.
BeeSensible lists the consumer and business variants of the same service as separate rows in the catalogue, each with its own risk score, and links them to the use observed in the organisation: only the domain, at most once a day per device, without a user ID. The administrator records a decision per variant, and that decision appears in the browser when someone opens the tool, with the allowed alternative as the first button. This is how the variants sit side by side:
