People in your business asked this question weeks ago, and most of them answered it for themselves by signing up. The useful version of the question is narrower: safe on which account, with which data, under which rules. Once you can answer those three, you know whether it is safe in your business.
It depends on the account before it depends on the tool
Two people can open the same chat window and be in different positions, because the account decides where the conversation goes and who can see it afterwards. On a free personal account, the terms are written for consumers, and the business tiers sit on different terms, which usually say whether what you type can be used to improve the model and how long it is kept.
Read the terms for the account type your people have, because that is the version they are using. What changes between ChatGPT's plans sets out each one, and for a tool other than ChatGPT, checking a new AI app before anyone uses it covers the same questions about terms and accounts.
The account also decides what you can do later. On a personal subscription, paid on a personal card and registered to a work email address, you cannot see what went in, you cannot switch it off when the person leaves and it appears on no system you administer. Since nobody gave them rules, none of that is misconduct. It is still the largest exposure most businesses have, and the first thing to move onto a company account.
What should never go into a chat window
The rule holds for every tool, whichever account it sits on, because these are things that should not leave your business at all.
- Anything a client gave you in confidence: contracts, drafts, case notes, correspondence, anything under a non-disclosure agreement.
- Personal data about an identifiable person: a colleague, a client's customer, a member, a beneficiary, a candidate.
- Credentials of any kind, including API keys, passwords and the contents of a configuration file.
- Financial figures that have not been published, pricing and anything a board has not yet seen.
- Source code you do not own outright.
Your people will hit a borderline case every week, which is why the list belongs in a written policy where they can check it without asking.
If it has already happened, here is what to do next.
Safe to use for what
Most of the work people bring to a chat tool passes this test, because nothing confidential goes in and a person reads the output before it goes anywhere. That covers a first draft of an email from your own notes, a rewrite of a paragraph you wrote, a summary of a public document and a question about how a spreadsheet function works.
The risk arrives when the shortcut gets taken with something that mattered, and the output is trusted because it reads well. A summary of a contract is only as good as the model's reading of it, and a figure in an answer stays unchecked until someone compares it with the source. So the policy should name the categories of work where a person reviews the output before it leaves:
- Anything that becomes a commitment.
- Anything with a number in it.
- Anything about a named individual.
- Anything a client will read as a statement of fact from you.
Three questions that settle it for your business
- Which accounts are in use, and whose are they? Including the free ones, the browser extensions, the note-takers connected to calendars and the subscriptions on personal cards. Finding the AI tools nobody approved sets out where to look.
- What is the written rule on what may go in? If the answer is a message in a chat channel, you do not yet have one.
- Who reviews the output, and for which work? The answer is named roles against named categories of work.
If you can answer all three in writing, ChatGPT at work is as safe as any other tool your people use, and considerably safer than the version of it that is already running without rules. If you cannot, the gap you find is the same one a client, an insurer or an assessor will ask about.
Where to start
The quickest way to see where you stand is to answer ten questions, which takes about two minutes and sends nothing anywhere. If you already know the accounts are out there and want them found, on whose logins and reaching what, that is the readiness snapshot. The written rules themselves are set out in what an AI policy should contain.