Data privacy
Where your company data goes
when you paste it into a chat window
The honest answer depends on which account your team is using and what they put into it. This page explains where the text sits afterwards and who can reach it.
Not by default. Whether ChatGPT is safe for your company data depends on which account your team uses and what they paste into it, and in most offices the account in use is a personal one that was never set up for company records.
Nothing here is a reason to panic. A chat window is a useful tool and I use one every day. The point is that most teams do not know which account they are on, what the retention setting is, or who at the provider can read a conversation afterwards. Those three answers decide whether the tool is safe for the material going into it.
The account decides what happens
A personal sign-up, a paid team plan and a company-managed workspace keep your text for different lengths of time and give different people access to it.
Training is a setting, and it moves
Whether an outside service may use your text to improve its own models is controlled by a setting and by the plan you are on. It can change when the plan changes.
The attachment is the risk
The material that causes trouble is rarely the question. It is the contract, the salary sheet, the client list or the drawing that gets pasted in so the answer makes sense.
A copy you cannot account for
Once company material has left your network, you can no longer say where every copy of it sits. That is the sentence an auditor is trying to avoid hearing.
There is a version that stays put
The same work can run on a machine in your own office, where the records and the conversations never leave the building. See what that is.
What happens to the text after you press send
The words you type are sent over the internet to a computer somewhere else, read by a model running there, and an answer comes back. The conversation is then stored against your account so you can open it again. How long it is kept, who inside the provider can look at it, and whether it is used to improve the model are three separate questions with three separate answers.
None of that is hidden. It is written into the terms of the account somebody signed up with. The problem in most offices is that nobody has read the terms of the account their staff actually use, because the staff signed up themselves.
The three kinds of account, and what each one does
Almost every worry about AI and company data comes down to which of these you are on. The table is deliberately generic, because the names and the settings change and the shape does not.
| Account | What it is | What your text may be used for | Who can reach the conversation |
|---|---|---|---|
| A personal sign-up | An account your staff opened themselves, on their own email | Improving the service, unless the person found the setting and turned it off | The person, the provider, and anyone who gets into that personal email |
| A company-managed plan | Seats bought by the company, with an administrator | Usually kept out of training under the business terms | The person, your administrator, and the provider |
| A model on your own machine | An open model running on a computer in your office | Nothing. It never leaves the machine | The people you give access to, on your own network |
What gets pasted in, in practice
This is the list I write down in an audit. None of it is anyone's fault. People paste the document in because the answer is useless without it.
- Contracts and letters of intent, so the model can summarise them
- Salary sheets and appraisal notes, so a review can be written faster
- Client lists and enquiry exports, so somebody can sort them
- Drawings, costings and vendor quotations, for a comparison
- Screenshots of internal dashboards, pasted in as an image
- A customer's phone number and address, inside a message being drafted
The question that is actually being asked
When an owner asks me whether a chat window is safe, the real question is usually narrower. It is whether they could sit in a room and say where every copy of a customer's details currently sits. With an outside service in the middle, the honest answer is that they could describe the terms they agreed to, and no more than that.
India's data protection law, the Digital Personal Data Protection Act 2023, is why that question now gets asked by boards rather than by IT. I am not a lawyer and nothing here is legal advice. What the law is about is a separate page, written the same way.
If you are going to keep using an outside service
Most businesses will, and should, for some of the work. These are the rules I give a team when the tool stays.
- Move everyone onto a company-managed account, so an administrator can see what exists
- Write down one line about what may never be pasted in, and name the documents
- Turn off the training setting where the plan allows it, and check it again after any plan change
- Keep customer names, numbers and addresses out of the window, even inside a draft
- Give the team a private place to do the same work, so the rule is one they can follow
The version where the records never leave the building
The work most teams use a chat window for, summarising, drafting, pulling an answer out of a long document, can run on a computer in your own office instead. The model sits on that machine, reads only the records you point it at, and the conversation is stored on your own network.
That is the whole of what I build. My own studio runs on one, and I built the people system at a forty-person studio the same way. See what sits in the room, or open the working systems.
Built, not described
Questions people ask
Is ChatGPT safe for company data?
It depends on the account and the material. A company-managed account with the training setting off is a reasonable place for ordinary drafting. Contracts, salary sheets and customer details belong somewhere you control, because once they leave your network you can no longer say where every copy sits.
Where is my ChatGPT data stored?
On the provider's own computers, outside your network, under the terms of the account somebody signed up with. The conversation is kept against that account so it can be reopened. The retention period and who inside the provider may view it are set by them, not by you.
Does AI train on my data?
On a personal account, usually yes unless somebody turned the setting off. On a business plan, usually no under the business terms. Both of those can change when the plan or the terms change, so it is worth checking again rather than assuming.
Can AI read my WhatsApp messages?
An outside assistant inside a messaging app reads what you send it. An Agent I build reads only the chats and folders you point it at, that list is visible, and it writes down what it looked at. Nothing is read because it happened to be nearby.
Is a paid business account safer?
A company-managed plan is better than staff on personal sign-ups, because an administrator can see which accounts exist and the business terms usually keep your text out of training. The text still leaves your network, which is the part a paid plan does not change.
What should we never paste into a chat window?
Anything that identifies a customer, anything to do with pay, and anything you signed. If somebody needs AI help with those, give them a place to do it inside your own office rather than a rule they will quietly ignore.
Can we check what our team has been pasting in?
Only on a company-managed account, where an administrator has visibility. On personal sign-ups there is nothing for you to look at. That gap is usually what moves an owner from writing a policy to buying a machine of their own.
Next step