An IT administrator at a consultancy has three requests on his desk. Marketing wants Claude, because it rewrites longer texts more comfortably. Finance has had ChatGPT accounts for six months, paid on a personal card. And the directors have just bought Microsoft 365 Copilot licences and are asking whether the other two can now go.
None of the three wins. The account decides the data protection for all three, and only then does something differ that matters: what each assistant can see, and where the work sits to keep it safe. This piece uses the facts from the separate pieces on ChatGPT, Claude and Copilot.
Which account decides the data protection?
For all three the dividing line runs straight through the brand, not between brands.
ChatGPT exists as a personal account (Free, Plus, Pro) and as a Business, Enterprise or Edu workspace. Personal accounts can use chats for model improvement unless the user turns that off; for the business workspaces OpenAI states that data is not used for training by default.
Claude exists as a consumer account (Free, Pro, Max) and as Claude for Work and Enterprise. For consumer accounts training has been on by default since September 2025 until the user turns it off. Claude for Work and Enterprise are contractually excluded from training, and Enterprise can be set up with Zero Data Retention.
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 consumer version has no enterprise data protection; the two signed-in variants do, and Microsoft states that prompts, responses and Graph data there are not used to train the underlying models.
So the three requests concern six tools, not three. Why that holds at every vendor is covered in free vs business AI accounts.
What does each assistant see?
ChatGPT and Claude see what someone supplies: pasted text, uploads, screenshots. The reach is as large as the prompt, and that is where the risk sits. Both stretch that reach: ChatGPT through memory and shared links, Claude Enterprise through documents in projects that stay until someone removes them.
Microsoft 365 Copilot sees what the user can open. It searches email, calendar, Teams, SharePoint and OneDrive via Microsoft Graph with the permissions of the person asking. It grants no new access, but makes years of oversharing findable with one question.
So with ChatGPT and Claude the work sits in behaviour at the moment of typing. With Microsoft 365 Copilot it sits in the permission structure, and Microsoft puts cleaning that up first in its own deployment guidance.
How do the three compare?
| ChatGPT | Claude | Microsoft Copilot | |
|---|---|---|---|
| Business variants | Business, Enterprise, Edu | Claude for Work, Enterprise | Copilot Chat (work account), Microsoft 365 Copilot |
| Training on input | Personal: possible unless turned off; business: excluded by default | Consumer: on by default since September 2025 unless turned off; business: contractually excluded | Consumer: no enterprise protection; signed in: no training of the underlying models |
| Reach | What is pasted or uploaded, plus memory | What is pasted or uploaded, plus projects and knowledge bases in Enterprise | Copilot Chat: uploads and screen; Microsoft 365 Copilot: everything the user can open via Graph |
| Integrations | GPTs, apps and connectors; off by default in Enterprise and Edu | Code Interpreter with network access, projects; abused through prompt injection in October 2025 to exfiltrate conversation data | Connectors, agents and plug-ins; EchoLeak (2025) showed prompt injection reaches the whole Graph scope |
| Administration | Business: members, apps, GPT settings; Enterprise: SSO, SCIM, roles, retention, residency | Work: members and workspace settings; Enterprise: SSO, SCIM, audit log, Zero Data Retention | Licences per user, Purview labels, DLP for Copilot, Restricted SharePoint Search as a temporary brake, flex routing |
| EU processing | Depends on how the Enterprise environment is configured | Transfers covered by standard contractual clauses in the commercial data processing agreement | Storage within the EU Data Boundary; flex routing can move processing outside the EU at peaks, can be turned off |
| Where the work sits | In the prompt | In the prompt and in what stays in projects | In the permissions |
Two rows deserve a note. Integrations widen what the assistant can read, and for all three a crafted document or email has made the assistant do what the attacker wanted. The difference is what was reachable at that point: with Claude the session's conversation data, with Copilot everything the user could open via Graph.
EU processing is a question for all three. With Copilot it is a switch that is on by default for newer tenants, with ChatGPT a configuration choice in the Enterprise environment, with Claude a contractual arrangement on transfers. Anyone who cares about this checks it.
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 once SharePoint and OneDrive permissions have been cleaned up.
The business variants of ChatGPT and Claude fit when the work consists of text someone supplies themselves: rewriting, summarising, translating, code. The reach is then bounded by the prompt, which makes the risk clearer to see but not smaller. Which of the two is a matter of preference and contract: whoever needs Zero Data Retention looks at Claude Enterprise; whoever wants to configure residency in a workspace looks at ChatGPT Enterprise. With both, the question remains what goes into the prompt, and no account settles that.
Personal ChatGPT accounts, Claude consumer accounts and the consumer Copilot fit no work involving personal data. More on the weighing per tool and per use is in how do you assess the risk of an AI tool.
What if all three are in use?
That is the reality at the consultancy, and at most organisations: the licences are there, and the other accounts do not disappear because of a decision 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.
For the consultancy it becomes: Microsoft 365 Copilot allowed after the permissions clean-up, ChatGPT Business and Claude for Work on limited use with the condition "no client data in the prompt", and the personal accounts of all three not allowed, with the business variant as the alternative. The comparison of just ChatGPT and Copilot, with more on administration, is in ChatGPT versus Microsoft Copilot for business use.
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: only the domain, at most once per 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. Side by side:
