What to keep out of ChatGPT when you work with clients

It is late on a Thursday. You have just finished a coaching session and you want the follow-up email out before you forget what was said. So you open ChatGPT and paste your notes straight in:

Session notes, Priya Shah, Head of Ops at Northgate Logistics. She's been signed off with anxiety for a fortnight, back Monday. Worried her MD, Tom Briggs, will push her out before the restructure. Mortgage stress, partner just lost his job. Wants to ask for a pay rise anyway. Write a warm follow-up email recapping actions.

The email that comes back is good. That is the problem. It took her name, her employer, her boss, her health and her household money, and put all of it into a tool you do not control, to write something that needed almost none of it.

Why it happens

It happens because the tool is so easy to talk to. A chat box feels like a notepad, and notes are where client detail lives, so the detail goes in with everything else. Nobody decides to share a client's health history. It just travels along with the request.

It also happens because we have been told, correctly, that AI needs context to do good work. That is true. But the context it needs is about the situation, not about the person. A follow-up email needs to know what was agreed and how the client felt. It does not need to know who they are.

AI needs the shape of the situation, not the identity of the person.

What to keep out

Before anything client related goes into a chat, run it past a short list. If a detail falls into one of these groups, it stays in your notes:

• Names of clients, their colleagues, their family and their employers
• Health details, including mental health, sick leave and medication
• Money details, such as salaries, debts, mortgages, fees and bank information
• Contracts and anything you are bound to keep confidential
• Anything that would let someone work out who it is, such as a job title in a small firm, a town, or an unusual event

That last group is the one people miss. "Head of Ops at a logistics firm in a small town, back from sick leave" can be enough to identify someone without a single name in it.

How to anonymise without losing the answer

Anonymisation means removing identifying details so the information can be used more safely. In practice, for a coach or consultant, it comes down to a few moves.

Swap names for roles. Priya becomes "my client". Tom becomes "her manager". Northgate becomes "her employer". The AI does not care what anyone is called. It only needs to know who is who.

Blur the specifics. "Signed off with anxiety for a fortnight" becomes "returning after a period of leave". "Mortgage stress, partner lost his job" becomes "under some financial pressure at home". You keep the emotional weight the email needs to respect, and drop the facts it does not need to mention.

Keep the job and the tone. This is the part that makes the answer good. Say what was agreed, what the client is nervous about, and how you want the email to feel. That is the context that actually changes the output.

If you use an AI Brief, the same rule applies. Your brief should describe your business, your offer and how you sound. It should never hold client records.

Before and after

Here is the raw paste again:

Session notes, Priya Shah, Head of Ops at Northgate Logistics. She's been signed off with anxiety for a fortnight, back Monday. Worried her MD, Tom Briggs, will push her out before the restructure. Mortgage stress, partner just lost his job. Wants to ask for a pay rise anyway. Write a warm follow-up email recapping actions.

And here is the anonymised version:

I'm a coach. Write a short, warm follow-up email to a client after today's session. Context: she's returning to work next week after some time away, and she's nervous about where she stands with her manager during a restructure. She's decided to ask for a pay review anyway. Actions we agreed: write down a few recent wins before Monday, book a one to one with her manager, and draft her opening line for the pay conversation. Tone: encouraging, practical, not gushing. Don't mention her leave or her personal circumstances. Sign off as me.

Here is the kind of reply the anonymised prompt gets back:

Hi [Name],

Thank you for today. I know the timing feels tricky, and I thought you were clear and brave about what you want.

Before Monday, jot down a few recent wins while they're fresh. Then book a one to one with your manager for early in the week, and have a go at the opening line for the pay conversation. It doesn't need to be perfect. We'll shape it together next time.

You've done the hard part, which is deciding. The rest is preparation.

Speak soon, ReadThinkAI

That is as good as the email the raw notes produced, and arguably better, because you told it what to leave out. You add her name when you paste it into your own email, where it already belongs.

The anonymised prompt is also more useful in a quieter way. Writing it forced you to say what was agreed and how the email should feel. The raw notes left the AI to guess both.

What it won't fix

Anonymising lowers the risk. It does not remove it. If the story is unusual enough, a determined reader could still piece together who it is, and a careless detail can slip through when you are tired. Treat anonymised text as less sensitive, not as safe.

It also won't make you compliant. If your contract with a client promises confidentiality, or you work in a regulated area, your obligations do not change because you removed a name. When in doubt, the honest answer is to leave it out of the chat entirely and write that one yourself.

And a tool's settings are not a contract. Both providers publish pages on how they use conversations. OpenAI's page on how your data is used to improve model performance describes a setting that controls whether your conversations help improve its models, and notes that giving feedback on a response can bring the whole conversation in. Anthropic's page, Is my data used for model training?, describes a similar choice and a similar point about feedback. Read them, set them the way you want, and check them now and then, because the pages are updated. But a toggle you switched on is not a promise you made to your client, and it is not a promise the provider made to them either.

One thing to do today

Open your last few AI chats and look for a client name. If you find one, write yourself a one line rule and stick it above your desk: roles, not names; shape, not specifics. Then use it on the next follow-up email you write.

If you want the wider picture of what is safe to put into an AI tool and what is not, the AI Foundations course covers it from the beginning.