Four data classes, a name-swap habit and a tool card
A manager types an employee's name, a bus-schedule excuse and three late arrivals into a free chat tool. The tool now holds a record about a real person, under rules the manager never read.
18 min · You'll sort what to paste, swap names for labels and fill a tool cardYou'll swap a real name for a placeholder, sort what you handle into four classes, and fill a tool card for the tool you use.
Step 1 of 6: The idea
Step 1 of 6
The one idea this lesson is about, and why it matters.
Complete Where AI helps in HR work, and where it must not decide first.
Quick recall · Where AI helps in HR work, and where it must not decide
Open the sort you made last lesson. Which of your amber tasks start from real people's details and so need placeholders? Was any red task red because of the data and not the decision?
Amber tasks that start from notes or comments about real people (a review rewrite, a survey summary) need the name swap. A red task such as reading an employee's doctor's note is red because of the data. It is protected, so it stays out of a chat tool. If you named your own tasks and said why, you have the habit.
What you can paste depends on two checks: what the information is, and which plan you are on. The tool matters less than those two. Each takes under a minute once it is a habit.
First, the four data classes. A data class is a label for how careful to be with a piece of information. This is our own simple scheme for training, not a legal term.
Public is anything already posted for everyone, such as your job ads. Internal is yours but not secret, such as a blank form or a draft checklist.
Confidential is about a real, named person: a resume, a manager's note, a pay rate. Protected is the most sensitive: medical or leave details, accommodation requests, ID, bank and tax numbers, and any protected trait of a named person.
The rule of thumb. Public and internal can go into any plan. Confidential goes in only as placeholders or on a work plan your company has approved.
In a small company, the owner decides and writes it down.
Protected does not go into any general chat tool, on a personal plan or a work plan. A placeholder does not fix it, because the detail itself is the problem. That means medical details, leave reasons, accommodation requests and complaints.
Handle those by hand in the locked HR file, shared with only the short list of people who may see them. US disability rules treat medical information as confidential and keep it in separate files. A chat tool is not on that list (our reading, so ask your counsel, an employment lawyer: see lesson 1).
Second, the plan. A personal account is a free account, or one you pay for yourself (vendors call it a consumer account). A work account is a plan a company buys for its staff, such as Team, Business, Enterprise or Workspace.
They are different deals, even in the same app. Personal accounts have a setting on whether your chats help train the model (the AI behind the app learns from them). Some also let vendor staff read a sample of chats.
Work accounts say they do not train on your content by default. An admin (the person who manages your company's accounts) controls their settings. Plans change, so check the vendor page and write the date on it.
The facts below are as of Nov 19, 2026. If your plan is not listed, the page for it will say.
Claude (Anthropic). On Free, Pro, Max and other personal plans, you choose whether chats help improve Claude. Open your name, then Settings, then Privacy, and look for the "Help improve our AI models" switch.
A chat started in incognito mode (the ghost icon on a new chat) is not used to improve Claude. Team, Enterprise and the API (how programmers connect software to the model, not the chat app) do not use your content for training by default.
ChatGPT (OpenAI). On Free, Go, Plus, Pro and other personal plans, open Settings, then Data controls. Turn off "Improve the model for everyone" if you do not want your chats used for training.
OpenAI's help page would not open for us on Nov 19, 2026. If the names differ, search Settings for "improve the model".
A Temporary Chat is a chat the app does not save to your history or memory. To start one, open a new chat and look for "Temporary Chat" (names change; checked Nov 19, 2026). It is not used for training but may still be kept for a short time.
Business and Enterprise do not train on your content by default.
Gemini (Google). In the personal app, open myactivity.google.com/product/gemini (your Gemini activity page) and look for the control that lets Gemini use your chats to improve Google's AI. Turn it off if you do not want that.
A sample of chats is read by human reviewers and kept up to three years, even if you delete them. Work and school accounts on Workspace are not used for training without your admin's permission.
To start a temporary chat, open a new chat and look for "Temporary chat" (names change; checked Nov 19, 2026). It is not used for training, but Google's page says it is still kept for about two months.
A temporary chat is a safer place for a one-off question, but it is not a vault. Keep the same classes rule. Google also says not to connect Gmail or Drive to the personal Gemini app if they hold confidential information.
No work plan? You can follow this whole track on a free account. Use it for public and internal text and for placeholder drafts, and keep confidential and protected details out.
The name-swap habit. Replace every real name with a label such as Employee A. Keep the list that matches labels to names in your own file, not in the chat.
Remove other details that point to one person, such as an exact start date or a phone number.
A placeholder is not anonymity. In a 12-person office, "the hygienist who started in March" still points at one person. Swap the name, then ask whether anyone could still guess who it is.
The tool card. A tool card is a one-page record of what a tool does with what you paste. It lists the plan, whether it trains on your chats, how long it keeps them, who can read them, and what you will never paste.
Fill it from the vendor's own pages, write "not stated" where the page is silent, and date it. This is training, not legal advice, so your company's policy and your counsel come first.
Check the class, check the plan, swap the names, and date the card.
That is the idea. Next you watch it work.
Step 2 of 6
See one real prompt do one real job, then see which part of it made that work.
You type this
Tidy this manager's note into three clear sentences for a file. Keep the facts and dates. Do not add your own opinion of the employee. Note: Employee A (front desk) was late on three Mondays in October, 10 to 20 minutes each time. I spoke with Employee A on Oct 28. Employee A said the bus schedule changed. We agreed Employee A will text me by 7:45 if running late.
Your AI tool does this
Your AI tool goes next.
Step 3 of 6
Write one part of the prompt yourself, then compare with one good answer.
Your first step before pasting
Think about what in the note points to one person.
Your answer is saved with your progress, so leave out anything private.
Try it before you go on.
Getting it wrong first is what makes it stick.
Step 4 of 6
Do the task on your own work, in your own tool.
Pick one real note, email or draft with a name in it (no medical details, no ID numbers). Swap the names, ask for a tidy version, then swap the names back in your own file. Then fill in the tool card for the tool you use.
Go do it for real.
Come back with a result.
Leave out anything private. This uses one of your Course Council messages.
Step 5 of 6
Fill in your own card, then save it to your workbook.
Your takeaway
Loading your card…
Save your card before you go on. It goes in your workbook.
Step 6 of 6
Say it from memory first, then check it.
Pulling this out of your own head now is the part that actually moves it into long-term memory. Re-reading only feels like it does.
Without scrolling up, name the four data classes in order, and say which one never goes into a general chat tool. Then name one thing from your own week in each class.
Public, internal, confidential, protected. Protected never goes into a general chat tool. A good answer for your week: public is a posted job ad, and internal is a blank onboarding checklist. Confidential is a manager's note about a named person, and protected is an employee's medical note.
Sort by class. You can sort what you handle into public, internal, confidential and protected, and say what goes where.
Before you paste, check the data class and the plan. Swap names for placeholders, keep protected details out of general chat tools, and date a tool card from the vendor's own pages.
Copy the text of the vendor's public data page and ask: answer these five questions only from this text, and quote the words. Tell it to write 'not stated' where the page is silent. The page is public, so this is safe to paste. You still read the quoted passages, because the point of the card is that a person read them.
That is the lesson. Mark it complete, and its question comes back to you in three days.