It's the end of a long day and you want one last thing off your plate. A customer needs their invoice emailed tonight, so you open a chat with an AI tool, and to save time you paste in the whole file you've got on them. Name, home address, the card number they read you over the phone, a note about the surgery that's keeping them home while you work. Ten seconds later you've got a tidy invoice, and you never think about that paste again.
That text just left your computer. It traveled to a company's servers, and depending on your settings, some of it may sit in a log for a while or even help train the next version of the tool. Most of the time nothing bad comes of it. But "most of the time" isn't a plan, and some of what you pasted was never yours to hand out.
You don't need to be paranoid to use these tools, and you don't need to quit them either. What you need is a simple rule for what's fine to share, what to clean up first, and what should never leave your own files. That's what this is.
Green light: fine to paste as-is
A lot of your daily work carries nothing private at all, and the AI can chew on it freely. This is where you should be using these tools without a second thought.
- Marketing and general writing. A draft blog post, a social caption, a rewrite of your "about" page, an email going out to your whole list.
- How-to questions. "What's a fair rate to tune up a boiler in New England?" or "Write me a checklist for a pre-winter furnace service." No customer attached.
- Your own templates and language. Standard estimate wording, a warranty paragraph, your cancellation policy, a phone script.
- Public information. Anything already on your website, a product spec sheet, a town's permit process.
If you'd be comfortable reading it aloud at a Rotary breakfast, it's green. Paste away.
Yellow light: clean it up, then paste
Most of your real work lands here. You want the AI's help with an actual customer situation, and that's fine, as long as you pull out the parts that point to a specific person first. The tool does just as good a job on a de-identified version, which is a fancy way of saying you swapped the real details for blanks.
Say a customer sent a rambling three-paragraph complaint and you want to write a calm reply. Before you paste, change "Mrs. Sullivan at 14 Maple in Wellesley" to "the customer." The AI writes the same good reply either way, and nothing traceable left the building.
Strip these before pasting:
- Names, and swap in "the customer" or "the client."
- Street addresses, which become "the job site."
- Phone numbers and email addresses.
- Account numbers, invoice numbers, anything that links back to one file.
It takes about ten seconds and it becomes a habit fast. Highlight, delete, type "the customer," done.
Red light: this never leaves your own files
Some things should never go into a chat window, no matter how convenient it would be. An AI tool is built to be helpful and chatty, not to be a locked filing cabinet. It's the wrong place for a secret, full stop.
- Full credit card numbers, security codes, or card expiration details.
- Social Security numbers and tax ID numbers.
- Bank account and routing numbers.
- Medical details tied to a name. If you run a clinic or send out home health aides, that surgery note stays in your own system.
- Driver's license and passport numbers.
- Passwords and login codes, yours or anyone else's.
- Anything under a signed confidentiality or non-disclosure agreement.
Here's the honest limit. Even with training turned off, your text still travels to a company's servers and may sit in a log for a while for safety and abuse checks. That's fine for a de-identified question. It is not fine for a card number. When you truly need AI to work with sensitive records, that's a job for a proper business setup with a signed data agreement, not the free chat box on your phone.
How to check that your chats aren't training the AI
The major tools each let you turn off using your conversations to improve their models. The exact menu moves around as these companies update things, so the trick is knowing what to look for rather than memorizing a path.
Open the tool's Settings, then hunt for a section called Data Controls, Privacy, or Data & privacy. You're looking for a toggle worded something like "improve the model for everyone" or "use my chats for training." Switch it off.
- ChatGPT: Settings, then Data Controls. Turn off the option about improving the model, and consider a "temporary chat" for one-off sensitive questions.
- Claude and Gemini: look under settings for a privacy or data section with the same kind of training toggle.
- Paid business and team plans usually don't train on your data by default, which is one real reason to pay the roughly $20 to $30 a month per seat once you're using these daily.
One caveat worth saying out loud: turning off training is not the same as the data never leaving. It just means your words aren't feeding the next version. That yellow-light habit of stripping names out still matters.
Worth doing this week
You don't need a policy binder. You need a couple of habits that take an afternoon to set and then run on their own.
- Flip the training toggle off in every AI tool you use. Ten minutes, once.
- Practice the strip. Next time you paste a customer message, swap the name and address for "the customer" and "the job site." Do it three times and it's muscle memory.
- Post the red list where your crew can see it. Card numbers, Social Security numbers, bank details, medical notes, passwords. Those never go in a chat box.
- Pick your yellow default. Decide that customer-specific work always gets cleaned before it's pasted, and tell whoever else touches these tools.
- If you handle truly sensitive records daily, get a business setup with a signed data agreement instead of the free app. We help Massachusetts shops sort out which tools are safe for which jobs, so give us a shout if you're not sure where a task falls.
Use these tools every day. Just know which door your data walks through, and keep the private stuff on your side of it.