Three office recipes
and the postcard rule

The pile is on every desk in the trades. The email thread that's forty messages deep, with an answer you owe somebody buried around message twenty-two. The job notes scribbled on the back of a materials receipt. The estimates you sent last week that nobody answered, and the follow-ups you keep meaning to write after dinner and never do.

That pile is where an AI chatbot — ChatGPT, Copilot, Gemini, Claude — earns a seat in your business. Not the marketing everyone keeps pitching you — the boring office work that eats your evenings.

And the first thing most owners say next is the right instinct: "I'm not pasting my customers' information into some company's chat box." That's the caution this whole page is built around. Call it the postcard rule, and it covers all three recipes: nothing goes into the box that you wouldn't put on a postcard. Names, addresses, and phone numbers come out before you paste and go back in after. The full rule, and what these companies actually do with what you type, is at the bottom. First, the work.

Recipe one: the forty-message thread

Open your AI chatbot — ChatGPT, Copilot, whichever you picked — and start a new chat. Copy the whole email thread. Before you paste it, swap the customer's name, address, and phone number for blanks — [customer], [address], [phone]. It takes a minute. Then paste it under this:

Type this, then paste the thread under it

Here is a long email thread about a [type of job] job. Give me the short version: what got decided, what is still open, and what I owe a reply on.

Then the part that stays yours: read the summary against the thread. You're checking two things — did it miss a decision, and did it say anything nobody actually said. The machine only knows what you pasted, so any line in the summary you can't find in the thread gets cut. Nothing gets acted on until you've signed off.

Recipe two: scribbled notes into a clean paragraph

The site-visit notes on the back of the receipt are real information in a form nobody else can use. Blank out the homeowner's name and the address, then:

Type this, then paste your notes

These are my rough notes from a [type of job] site visit. Turn them into one clean paragraph for the job file. Do not add anything that is not in the notes.

That last line matters. Read the result once against your own scribbles: if it states something your notes don't, delete the sentence. Then put the real name and address back in your copy — the one on your machine, not theirs.

Recipe three: the estimate follow-up

The follow-up you keep not writing is a five-minute job when the blank page isn't yours. Your facts go in; a plain ask comes out:

Type this with your own facts filled in

I sent an estimate for [the work] on [day]. I have not heard back. Write a short, plain follow-up: no pressure, just checking they got it and asking if they have questions. Sound like a person, not a sales letter.

Read the draft line by line before it goes anywhere. Every fact in it must be a fact you typed — if it invents a deadline, a discount, or an eagerness you don't feel, cut it. Add the customer's name in your own email program, not in the chat box. It's your name at the bottom. The deeper version of this ask-then-check method is on the prompt method.

The postcard rule

Nothing goes into someone else's machine that you wouldn't put on a postcard. Customer names. Addresses. Phone numbers. Card numbers. Which job site has the lockbox and where the key hides. A chat box feels like a notepad, but it's a message handed to a company's computers.

The working method is the one you just used three times: swap the private parts for blanks before you paste, put them back after the text comes out. The machine doesn't need to know who the customer is to shorten a thread. The blanks cost you nothing.

What these tools do with what you type

Here's why that rule isn't paranoia. "Training" means using what you type to teach the company's future models. This is each company's own published position on it, rendered live from their own pages on July 20, 2026. These pages change, so before you trust any switch, open your own settings and look. The four common tools are in the table below — find your row and skip the rest; Grok and Meta AI aren't covered here.

Tool	The company's own position (checked live July 20, 2026)	Where the switch is
ChatGPT (consumer)	Trains on your conversations "unless you opt out." Temporary Chat excluded.	Settings > Data Controls — the toggle is literally named "Improve the model for everyone."
Copilot (consumer)	Trains on conversation activity; "You can opt out of use of your conversation activity for model training at any time."	Profile icon > Privacy — "Training on conversation activity" and "Training on voice conversations."
Gemini	A subset of chats can be read by human reviewers. Turning off "Keep Activity" stops future chats from being used to train their models — unless you choose to send Google feedback.	myactivity.google.com/product/gemini — check whether Keep Activity is on in YOUR account; don't assume either way.
Claude	The "Help Improve Claude" toggle controls whether new chats feed future training. The vendor doesn't publish what a new account starts at.	Settings > Privacy — open it and check the toggle yourself.

Two footnotes worth knowing. If you're on a paid business account, the rules differ — as of July 2026, OpenAI states that Business, Enterprise, and API conversations are not trained on by default, and Microsoft states that business Microsoft 365 Copilot conversations are not used for training.

And here's the honest line the settings menus won't volunteer: turning training off is not the same as being invisible. As of July 2026, per the vendors' own pages:

• OpenAI says opted-out chats stay in your history, and giving a thumbs up or down can put that whole conversation back into training.
• Google says that even with Keep Activity off, it still processes your chats to answer you and for safety — including possible human review.
• Anthropic's July 2026 privacy policy says a conversation flagged for safety review can still be used even if you opted out.

Two more limits worth knowing, as of July 2026. ChatGPT's Temporary Chat skips training, history, and memory, but OpenAI keeps a safety copy for up to 30 days. Copilot has a similar temporary mode — but on a work or school Microsoft 365 account, the IT admin can still see what you typed through the audit log, the activity record a company account keeps.

So put the tool on the pile. Let it shorten the thread, clean up the notes, draft the follow-up — with blanks where the people are, and your own read on every line before anything goes out under your name. The machine gets the postcard version. Your customers' real details stay where they belong: in your files, in your hands.