Can AI tools see my emails and files once connected?
Once connected, an AI tool can see what the connecting account is allowed to see, and nothing more. ChatGPT, Claude and Microsoft 365 Copilot all say they follow existing permissions. So the risk is mostly over-shared files, plan choice and what you let the tool change, not the tool itself.
This is general information, not legal advice. If you handle sensitive or regulated data, check your own obligations or ask a data protection adviser.
What can an AI tool see once I connect it to email and files?
It sees what the account you connect can see. That is the rule all three major vendors describe in their own documentation:
- Claude: Anthropic says Claude inherits each person's permissions from the connected service, so if someone cannot open a file, channel or record in the source system, the connector cannot reach it from Claude either. The exception is a custom connector that uses a shared credential, which reaches whatever that credential can access.
- ChatGPT: OpenAI says connected apps such as Google Drive or Slack let ChatGPT work with their information in a conversation, existing permissions still apply, and ChatGPT may ask for your approval before reading information or taking an action.
- Microsoft 365 Copilot: Microsoft says Copilot only surfaces organisational data that each person already has at least view permission to.
In plain terms: connecting an AI tool does not give it a master key. It gives it your key.
Is that actually safe?
That depends on what your key opens. The usual problem is not the AI, it is that files were shared more widely than anyone realised. Microsoft's own page makes the point: use the permission models in services like SharePoint to make sure the right people have the right access, because Copilot will show anything a person is allowed to see. If a folder of staff pay data is open to everyone, an AI assistant will find it for anyone who asks.
Before connecting anything:
- Check who can see what. Fix the sharing on sensitive folders first.
- Start read-only. Claude's Team and Enterprise owners can allow reading while blocking writing or sending, and ChatGPT asks for approval on actions.
- Connect one mailbox or folder first and see what it does.
- Keep an approval on anything that sends, deletes or shares.
Will my emails and files be used to train the AI?
It depends on the tool and, crucially, the plan:
| Tool and plan | Training, according to the vendor |
|---|---|
| ChatGPT Free, Plus, Pro | On by default; you can turn off "Improve the model for everyone" |
| ChatGPT Business, Enterprise, Edu | Not used for training by default |
| Claude Free, Pro, Max | You choose; if you allow it, retention rises to five years, otherwise 30 days |
| Claude Team, Enterprise, API | Not used for training by default |
| Microsoft 365 Copilot | Prompts, responses and data accessed through Microsoft Graph are not used to train foundation models |
One catch from OpenAI: giving thumbs up or down feedback can lead to the whole conversation being used for training, even after you opt out. These policies change, so check the linked pages before you rely on them.
For anything containing customers' personal data, my advice is to use a business plan where training is off by default, rather than relying on staff to change a personal setting.
What does UK data protection law say?
If the emails and files contain personal data, UK GDPR applies, whichever tool you use. The ICO publishes guidance on AI and data protection, covering lawfulness, transparency, accountability and security. Its page notes the guidance was last updated on 15 March 2023 and is under review because of changes made by the Data (Use and Access) Act. Treat it as a starting point and check for updates.
What does this look like when it is built properly?
One thing I have built logs every email, call and meeting against the right customer in a customer list, and prepares a briefing before every call (see the example). Everything is built on the business's own accounts and stays under their control, and nothing routes through my systems once it is handed over. Nothing sends without the owner's OK.
Questions people also ask
Can ChatGPT read my Gmail or Outlook?
Only if you connect it, and then only what that account can access. OpenAI describes connected apps that ask for approval before reading information or taking an action. Review the permissions on the sign-in screen, and disconnect it from settings when you no longer need it.
Can Microsoft 365 Copilot see files I am not allowed to open?
No. Microsoft says Copilot only surfaces organisational data that a person already has at least view permission to. The practical risk is files that are open to more people than you meant, so audit sharing before rolling it out.
Is it safe to connect AI to Google Workspace or Microsoft 365?
It can be, with the right plan and permissions. Fix over-shared folders, pick a plan where training is off by default, begin with read-only access, and keep approvals on sending and deleting. Then add more access gradually.
Do I need a policy for staff using AI with company data?
Yes. Say which tools are allowed, which plans, and what data must never go in. Without a policy, people use whatever personal account they already have, which is the setup where training is most likely to be on.
See it working
Things I've built that do this, and the time they give back.
A customer list that updates itself
Every email, call and meeting logged against the right customer, and a briefing ready before every call.
Typically 4 hrs a week per person
See how it works →