Your staff are already pasting company data into ChatGPT
Banning it does not work and blocking it rarely holds. The practical answer is to give people one approved tool and make that one the easiest to reach.
Every time this comes up in a client meeting, somebody says their policy prohibits it. Then somebody quieter mentions that the sales team has been using it to rewrite proposals since last year. Both statements are usually true.
This is not a discipline problem. The tools are genuinely useful, they are free, and the friction of not using them is felt immediately while the risk is abstract. Any approach that depends on people choosing the slower option will fail.
What the actual exposure is
Consumer accounts on most AI services may use what you type to improve the model, unless you turn that off. Business and enterprise tiers generally do not, and say so contractually. That difference is the whole ballgame, and it is invisible to the person doing the pasting, because both versions look identical in a browser.
Under the PDPA, personal data does not stop being your responsibility because an employee moved it somewhere convenient. If someone pastes a customer list into a free chatbot to tidy the formatting, you have disclosed personal data to a third party without a basis for doing so. Whether anyone ever finds out is a separate question from whether it happened.
The second exposure is quieter. Browser extensions and note-taking bots that join meetings often request very broad permissions, and they are installed by individuals rather than by IT. A meeting transcription tool with access to your calendar and mailbox is a data processor nobody has assessed.
What we suggest instead of a ban
Pick one tool and pay for it. For a business already on Microsoft 365 that is usually Copilot, because prompts stay inside your tenant boundary and inherit the permissions the user already has. If your team prefers something else, buy the business tier of that instead. The point is having a sanctioned option that is at least as convenient as the unsanctioned one.
Then narrow the rest. Entra ID lets you control which third-party applications can be granted access to company data, and turning off user consent for unverified publishers stops the browser extension problem at the source. It takes a few minutes and almost nobody notices.
A short written rule helps more than a long policy. Something on the order of: use the approved tool for work content; never paste customer records, contracts, credentials or anything you would not email to a stranger; and if a tool asks to connect to your mailbox or calendar, ask IT first. People remember three rules. They do not remember a nine-page acceptable use document.
A caution about Copilot specifically
Copilot respects your existing permissions, which sounds reassuring until you consider how permissions accumulate in a company that has been using SharePoint for six years. If a payroll folder was shared with everyone in 2021 and nobody noticed because it was buried, Copilot will happily surface its contents to anyone who asks a well-phrased question.
Run a permissions review before you roll it out, not after. Oversharing that was harmless when it required knowing where to look becomes visible the moment you add a search engine that reads everything on your behalf.
Where to start
Ask your team what they are already using, without any consequence attached to the answer. You will get a more accurate picture in one conversation than from any amount of monitoring, and it tells you which tool to sanction.
Worth knowing that CSA expanded the Cyber Essentials mark to cover AI security, so if certification is on your roadmap this is now in scope rather than optional.