What Is Safe to Put Into ChatGPT About Your Business

"Is it safe to put this in ChatGPT" is the wrong shape of question, and it is why the answers people get back are useless. Safe against what? A competitor reading it, a model learning from it, a regulator asking about it, and a customer suing over it are four different risks with four different answers.

The useful version is narrower. What happens to this specific thing after I paste it, and does that matter for this specific kind of information.

The tier is most of the answer

The same product behaves differently depending on what you are paying for, and this is the single largest factor in whether any of it is safe.

Consumer tiers are the free and personal-subscription versions. Historically these have defaulted to using conversations to improve the underlying models, with an opt-out available in settings. That default is the thing people are actually worried about when they ask this question.

Business and enterprise tiers are the team, business and enterprise plans. These generally commit contractually not to train on customer content, and that commitment sits in the terms rather than in a settings toggle. That distinction matters, because a setting can be reset and a contract cannot.

The gap between those two is wider than the price difference, and a business running client information through a personal account is making a decision it probably does not know it has made.

Check rather than trust this paragraph. Vendors change these terms, and a page describing them is stale the moment they do. Open your account's billing page, confirm which plan you are on, then open that vendor's current terms and find the section on customer content and training. Fifteen minutes, once, and repeat when you change plans.

What "trains on your data" actually means

The phrase carries more fear than it warrants and less nuance than it needs.

It does not mean your document is stored somewhere another user can retrieve. Models do not work as a searchable archive of what people typed.

It does mean your content may be used to adjust a model that other people use, and that memorisation of unusual, specific, repeated strings is a real if uncommon phenomenon. A product formula, a customer list or a distinctive contract clause is exactly the kind of unusual specific string that argument applies to. Generic business questions are not.

So the risk is not that someone reads your chat. It is that highly distinctive information becomes, in a small and hard-to-quantify way, part of something you no longer control.

Categories that stay out regardless of tier

Some things do not belong in a general-purpose tool no matter what the contract says, because the exposure is not really about the vendor.

• Customer personal information. Names with addresses, card details, health information, government identifiers. Your obligations to those customers do not transfer to a vendor because you agreed to some terms.
• Anything under someone else's confidentiality agreement. A client NDA usually names who you may share with. A tool vendor is not on that list.
• Credentials. Passwords, API keys, account numbers. This is a common real-world leak, and it is a security question more than a confidentiality one.
• Employee matters. Performance, medical, disciplinary, compensation.
• Anything you would not want quoted back at you. A useful last test when the category rules do not settle it.

The things that are genuinely fine

The corollary matters as much, because over-caution has a cost too, and a business that decides the entire category is unsafe gives up the benefit entirely.

Drafting from a blank page. Rewriting your own marketing copy. Explaining a concept. Summarising a document you wrote. Turning notes into a structured format. Generating options you will then judge. Checking your own writing for clarity.

None of that involves anyone else's information, and it is most of the day-to-day value.

The redaction habit

Between "obviously fine" and "obviously not" sits the bulk of real work, and the technique that resolves it is unglamorous.

Replace the identifying parts and keep the structure. A customer becomes "a customer". A dollar figure becomes "a mid-five-figure amount" when the exact number is not load-bearing. A company name becomes "a regional distributor".

The answer you get back is almost always as useful, because the model is reasoning about the shape of the situation rather than about who is in it. Ten seconds of redaction converts a question you should not ask into one you can.

What to do this week

Three things, and none takes long.

Find out which tier you are on, for every tool the business uses. This is the highest-value fifteen minutes available, and it shows whether anyone is using a personal account for business work.

Write down your own never list, on one page, in your categories rather than a generic one. A restaurant, a law firm and a retailer do not have the same list.

Tell whoever else uses these tools, because the policy that lives only in the owner's head is not a policy. One page and a short conversation is enough at this size, and it is the difference between a rule and an assumption.

When the terms change

They will, and the change will not be announced in a way you notice.

The practical habit is to re-check at two moments rather than on a schedule you will not keep. When you change plan or add a tool, and when a vendor emails about updated terms, which is the one email of that kind worth opening. Look for the same two things each time: whether customer content is used for training, and whether that is a contractual commitment or a settings default.

A commitment that lives in a settings toggle is one administrator away from being reversed, and on a shared business account that administrator may not be you.