Is it safe to put company data into ChatGPT or Claude?
It is the right first question. Before anyone pastes a customer list, a contract, or last quarter’s numbers into an AI tool, somebody should know where that information goes and who can see it.
The short answer: it can be safe, and whether it is depends far less on which tool you pick than on which kind of account your people are using. That one distinction covers most of the risk.
Business accounts and personal accounts are different products
ChatGPT, Claude, Gemini and Copilot all come in two flavors: a personal version that anyone can sign up for, and a business version that a company buys for its staff. They look almost identical on screen. The terms underneath are not.
On the business versions, all four vendors commit that your conversations are not used to train their models. On the personal versions, training is either switched on by default or something each user is asked to decide for themselves, and the conversations may be kept for longer.
| Tool | Business version (not used for training) | Personal version (check the setting) |
|---|---|---|
| ChatGPT | ChatGPT Business or Enterprise | Free, Plus, Pro: Settings → Data Controls → “Improve the model for everyone” |
| Claude | Claude Team or Enterprise | Free, Pro, Max: Settings → Privacy → model-training setting |
| Gemini | Gemini inside a Google Workspace account | The personal Gemini app: its activity and training settings |
| Copilot | Copilot signed in with a work account | Copilot signed in with a personal Microsoft account |
These terms change. The table reflects the vendors’ published policies as of October 2026. Check the vendor’s own privacy page before you rely on it.
The real risk is the account nobody approved
The risk most companies actually carry is not a breach at OpenAI or Google. It is an employee who found a useful tool, signed up with a personal email, and has been pasting work into it for months.
They are not doing anything wrong on purpose. They are trying to get through the day. But a personal account means the company has no say over the settings, no record of what went in, and nothing to switch off when that person leaves.
So the first step is not a policy. It is a question, asked without blame: who here is already using an AI tool, and for what? You will usually find two or three people who are, and they are the ones to build on.
What “used for training” means, and what it does not
“Your data may be used to train the model” sounds like your files are about to show up in somebody else’s answer. That is not quite it.
Training means your conversations become part of the material the next version of the model learns from. It is unlikely that a specific invoice gets repeated word for word to a stranger. The more practical concerns are plainer:
- Retention. The conversation is stored by the vendor, sometimes for years.
- Review. On some personal plans, conversations can be read by people at the vendor for quality and safety.
- Your own promises. Your contracts may already forbid it. An NDA, a client agreement, or a regulation like HIPAA may say that information cannot go to an outside service without specific terms in place, whatever that service does with it.
The third one catches people out most. The question is not only “is this tool safe?” It is also “have we promised anyone that their information stays with us?”
What never goes in, on any account
Some things stay out of AI tools no matter which plan you are on:
- Passwords, API keys, and logins. Ever.
- Payment card numbers and Social Security numbers.
- Health information, unless the vendor has signed a business associate agreement with you. Some enterprise plans offer one. Most plans do not.
- Anything a contract says stays inside the company. If you are not sure, ask whoever signed it.
Everything else (procedures, emails, drafts, spreadsheets of your own numbers) is generally fine on a business account. That is most of the work where AI actually saves time.
A one-page policy is enough
A small company does not need a forty-page AI governance framework. It needs a page people will actually read. Five lines cover it:
- Use company accounts only. We pay for them; personal accounts are not for work.
- Never paste in passwords, card numbers, Social Security numbers, health records, or anything under an NDA.
- You own the output. Read it before it goes to a customer. AI gets things wrong with total confidence.
- When in doubt, ask [a named person] before you paste.
- Tell us what is working, so the rest of the team can use it too.
Put a name in line four. A policy with no one to ask is a policy people work around.
What to do this week
- Ask who is already using AI tools, and for what. No penalties for honest answers.
- Buy business seats for the people who use it regularly. Compare the price with the personal plans they may already be paying for out of pocket. It is rarely a big line item.
- If anyone stays on a personal plan, have them check the training setting from the table above.
- Write the one-page policy, put a name on it, and send it out.
None of this needs a consultant. It needs an afternoon and a decision. Handled this way, data safety stops being a reason to wait and becomes part of doing it properly.
This is practical guidance, not legal advice. In a regulated industry such as healthcare, finance, or law, check your obligations with your counsel before you start.