ChatGPT at work: what should actually be in your AI policy?
A lot of AI policies are a page of principles. They say the organisation will be responsible, human-centred and careful. Nobody can use that to decide whether a client brief can go into ChatGPT.
The policy needs to answer the decision people actually face.
Write the decisions, not the slogans
- Which tools are approved, and for which kind of work.
- What must not be entered: client material, personal data, credentials, unpublished financials, anything a contract restricts.
- Whether a personal account is acceptable. For most client work, it is not.
- Who can approve a new tool, and how quickly.
- What a person should do if they have already pasted something they should not have.
- Where a human still has to check the output before it goes to a client.
Name the tools. “Generative AI” is too vague when the company means ChatGPT, Copilot, Claude, Gemini, Firefly and a meeting assistant. People recognise product names. They do not recognise a category.
Keep it short enough to follow
Two pages is usually enough. Link to the register of approved tools rather than pasting a list that will be out of date next month. Say who owns the policy. Date it. Review it when the tools change, which will be sooner than the annual policy cycle.
If the policy cannot be explained in a five-minute team meeting, it will not change what gets pasted into a prompt. The governance around it is what keeps the page honest: a register, an owner, and a way of noticing new tools.
One test
Hand the draft to someone who is about to start a pitch. If they cannot tell, from the page, whether the client brief can go into ChatGPT, the policy is still a slogan. Add the missing sentence. Take one out if the page is getting long.
Do not write that AI use is monitored unless a tool is actually doing that. A policy that claims a control you do not have is worse than a short page that says what is allowed today.
If this is the question already sitting with you, a conversation is the useful next step.
