Someone on your team has a 40-page contract in front of them and very little time. They open ChatGPT in the browser, paste the whole text and ask for a summary. They do it with the best intentions, from a personal account, and nobody finds out. In ten seconds, a confidential document has left your systems and landed in a third party's.

This happens every day in companies of every size, and almost never out of bad faith. The useful question is not whether your staff uses AI, but what information they are handing it.

◢What happens to what you type into ChatGPT?

That means whoever runs the tool, or their partners and contractors, may be able to read what you paste. The NCSC adds two more risks: stored queries can be hacked, leaked or accidentally made public, and the provider may later be bought by a company with a different approach to privacy than the one in place when you entered the data.

◢What shouldn't you paste into a public AI tool?

The NCSC's advice is clear: don't include sensitive information in queries to public models, and don't submit anything that would cause problems if it became public. For an ordinary company, that covers:

• ▹Personal data about customers, employees or candidates: names with ID numbers, addresses, records, payslips, CVs.
• ▹Contracts, quotes and commercial terms that are not public yet.
• ▹Credentials, access keys, tokens and passwords. It happens more often than you would think when someone pastes a code snippet or a config file to ask for help.
• ▹Your own source code and internal technical documentation.
• ▹Financial information, strategy, and anything you would not say out loud on a train.

◢Why do people do it if they know it's risky?

Because the tool works and nobody has given them another one. If the company doesn't provide an approved alternative, the work still gets done, but through personal accounts and outside any control. This is called shadow AI, and I touched on it when writing about the AI literacy duty in the AI Act.

Banning everything rarely works. The NCSC makes the same point at the end of its post: make sure people who want to experiment with these tools can do so, in a way that doesn't put the organisation's data at risk.

◢Is there a safe way to use AI with company data?

Yes, but it is not the free version in a browser tab. The NCSC distinguishes between public models, which it does not recommend for sensitive information, and private ones, which a cloud provider can offer or which you host yourself. In the first case, what the contract says is what counts:

• ▹Who inside the provider can see your queries, and under what conditions.
• ▹Whether your data is kept separate or mixed with that of other organisations.
• ▹Whether the provider and its partners can use it to research or improve the model.

The NCSC also notes that self-hosting a model is likely to be very expensive and should follow a security assessment. For many small and mid-sized companies the sensible choice sits in the middle: a business-grade tool with written terms, or an assistant connected only to the documents it really needs. If you connect it to your systems, don't skip the permissions question. I cover it in the article on AI agents with too many permissions.

◢How to build an AI use policy on one page

You don't need a forty-page document. One page is enough if it answers four questions:

1. ▹Which tools are approved for work and which are not.
2. ▹Which kinds of data must never go into any external AI, with concrete examples from your company's daily work.
3. ▹Who to ask when in doubt, with a name and an email address, not "the IT department".
4. ▹What to do if someone makes a mistake. Reporting it must carry no punishment, because a mistake that gets hidden is the one that becomes a big problem.

Then comes the hard part: explaining it. A policy that is only sent by email is read by nobody. It works far better with real examples from your own company and ten minutes of conversation in a team meeting.

◢What you can do this week

Ask openly which AI tools your team uses and what for. Don't make it an inspection: what you want to learn is the need behind it. With that list, decide what you offer as an approved alternative, write the policy and talk it through in a meeting.

If you want your staff to learn to use AI without giving information away, that is part of what I teach in my cybersecurity and AI training for companies, with examples designed for non-technical teams.