Why this matters
A new hire pastes a client’s payroll file into a chatbot. She wasn’t careless: nobody had ever told her where the line was. In a small office that one paste can mean a breach notice, an unhappy client and a week of cleanup.
The fix is one page. Four levels (Public, Internal, Confidential and Restricted) and one AI rule for each. Once it’s on the wall, nobody has to guess, and every other AI rule your office writes can point back to it.
Practice
Do every step yourself in a practice copy of the real tool. It checks your work, and “Do it for me” finishes any step you get stuck on.
What you’ll build
You are Priya, office manager at Harbor & Pine Bookkeeping, a made-up six-person firm. You sort four everyday items into the four levels: a services brochure, a staff memo, a client’s payroll summary and a client’s Social Security number.
Then you use the levels on a real job. A client email has a Social Security number in it, so you take that out by hand, turn on Temporary chat, and have ChatGPT swap the rest for placeholders. You catch the mix-up ChatGPT really made when we tested it (two people, one placeholder) and fix it with one follow-up.
What you need
- Nothing to start the practiceThe sheet and ChatGPT are practice copies inside the lesson. No account is needed and every name and number is made up.
- The data levels sheetDownload it below (CSV), then in Google Sheets choose File, Import, Upload.
- A ChatGPT account, to do it for realA free account at chatgpt.com is enough. Temporary chat is the dashed circle at the top right of a new chat.
- A list of your common documentsTen minutes with whoever handles client files. Twenty items is plenty to start.
Concepts
Four levels are enough
Big companies have long classification policies. A small office needs four words everyone remembers: Public (fine anywhere), Internal (approved tools, training off), Confidential (approved tools, placeholders first) and Restricted (never goes in).
When someone isn’t sure, the rule is to treat it as one level higher until the sheet’s owner decides.
Redacting with AI means sending it to AI
To swap a name for [CLIENT], ChatGPT has to read the name first. That’s acceptable for Confidential details in an approved tool. It’s never acceptable for Restricted data such as Social Security, bank or card numbers.
So Restricted items come out by hand before you paste. Temporary chat lowers the risk for the rest: it isn’t saved to your history or used for training, though OpenAI may keep a copy for up to 30 days for safety.
Check the list of replacements
In our real test, ChatGPT redacted the email neatly but gave the office manager and the client the same placeholder, [CLIENT]. The clean email hid the mistake; the list of replacements showed it.
Always ask for that list, read it, and fix anything wrong with one short follow-up.
Do it for real
Now repeat it in your own account. Tick each step as you go.
Downloads
- Data levels sheet (CSV)The four levels with examples and the AI rule, ready to import into Google Sheets.
- Four-level cheat sheet (.txt)One printable page: the levels, a 30-second before-you-paste check, the redaction prompt and space for your own list.
Quick check
3 questions to check you’ve got it.
1. A staff memo about new holiday office hours. Which level?
- Public
- Internal
- Restricted
Show the answer
Internal. It isn’t secret, but it’s meant for staff only. Approved tools with training off are fine.
2. A client email contains a Social Security number. What do you do before asking ChatGPT to redact it?
- Nothing, ChatGPT will remove it
- Delete the number yourself first
- Turn on Temporary chat, then paste it all
Show the answer
Delete the number yourself first. Restricted data never goes into AI, not even to be redacted. Temporary chat lowers risk but doesn’t remove it.
3. ChatGPT’s redacted email looks clean. What else should you check?
- Nothing, it’s done
- The list of what it replaced
- The word count
Show the answer
The list of what it replaced. In our real test, two different people got the same placeholder. Only the list showed it.
Choose the AI Assistant Your Team Should Use, Tested on ChatGPT, Claude, Gemini and Copilot
You know which data can go into AI. Next, decide which assistant your team is allowed to use: the same office task, tried in ChatGPT, Claude, Gemini and Copilot.
Start module 4 →