Staff Are Already Using ChatGPT at Work. What Should You Decide First?

If you employ more than a handful of people, some of them are already using public AI tools at work. A flat ban rarely stops it. A short, clear rule about what may be pasted in usually does more, and it keeps the useful habits while closing off the risky ones.

Your staff are probably not being reckless. They are trying to get through a to-do list faster, and a free tool that drafts an email in seconds is hard to resist. If you ban it without offering an alternative, the use goes out of sight instead of stopping, and you lose any chance to guide it.

Why does it matter?

Anything typed into a public tool leaves your control. Depending on the service and the plan, it may be stored, reviewed by staff at the provider, or used to improve the product. That is a concern for client details, staff records, contracts and pricing.

If personal data is involved, UK data protection rules apply to what happens to it. This is general guidance, not legal advice, and a data protection adviser should confirm what applies to your business. The practical point is simple: you should know what is going into these tools and be able to explain it if a client or regulator asks.

What should you decide?

What is never allowed. Personal data about clients or staff, anything covered by a confidentiality agreement, passwords and access keys, and unreleased financial information. Write the list on a single page so nobody has to interpret a long document.

What is fine. Drafting a generic email, summarising a public article, brainstorming headings, tidying the wording of text that contains nothing sensitive. Say so clearly. If the policy is all prohibitions, people will feel they are doing something wrong by using the tools at all, and they will stop asking.

Which tools are approved. Business plans from major providers usually come with firmer data terms than free consumer accounts, including commitments about whether your inputs are used for training. Pick one or two, read their terms, and name them in the policy. "Use the approved tool" is easier to follow than "use AI responsibly".

Who to ask. Name a person, not a department. People break rules when they do not know who to ask, or when asking feels like admitting a mistake.

What does a workable policy look like?

A one-page version might say:

• You may use the approved AI tool for drafting and summarising material that contains no client or staff personal data.
• You must not paste client names, contact details, contract terms, financial records or credentials into any AI tool.
• If you are unsure, ask the named person before you paste, not after.
• You remain responsible for anything you send out. Check AI-drafted text for accuracy before it leaves the building.

AI drafts can be fluent and wrong, and the person who sends the email owns the error.

When is a public tool the wrong answer?

When the work needs your actual documents. Answering staff questions from your own policies, or searching years of project files, means giving the system access to material you would not paste into a public chat window. Uploading files to a chat tool is awkward, hard to control and hard to audit. A private system that only reads approved documents, with access controls, source citations and logs, is the better fit. I covered the difference in another post on private knowledge bases versus chat uploads.

How do you roll it out?

1. Find out what is happening. Ask people, without blame, which tools they use and what for. You will learn which tasks are worth supporting properly.
2. Write the one-page policy and explain the reason in plain words. People follow rules they understand.
3. Walk the team through it in ten minutes, with two or three real examples of what is and is not fine.
4. Review it every six months. The tools and their data terms change often, and a policy that was sensible last year may now be out of date.

What happens if someone breaks it?

Treat the first slip as a sign that the policy was unclear, not as a disciplinary matter. If you punish honest mistakes, people will stop reporting them, and the reports are what let you fix a problem early. Reserve firmer action for deliberate, repeated disregard.