Your customers' data isn't yours to give away.
To get useful drafts out of an AI you have to paste real things into it — real messages, real prices, real situations. That's the deal, and it's a fine deal. But there's a line, most people find it by crossing it, and nobody selling you AI tools is in a hurry to point at it.
This is the line. Five minutes, and it's the same advice I follow.
The short version
Paste the context. Withhold the identity and the credentials.
An AI needs to know what happened to draft a reply. It almost never needs to know exactly who it happened to, or anything that could be used against that person if it ended up somewhere else.
Keep it out — always
- Credentials. Passwords, API keys, tokens, recovery codes. Anything that grants access. If it's in a password manager, it doesn't go in a chat.
- Payment details. Card numbers, bank accounts, anything you'd cover with your hand at a café.
- Government identifiers. ID numbers, passport numbers, tax IDs.
- Health information. Yours, your staff's, your customers'. Different rules apply to it in most countries, and none of them contemplate a chat window.
- Other people's private contact details at scale. One customer's message with their name in it is one thing. Your whole contact list pasted in for "analysis" is another.
Think twice
- Names, phone numbers, addresses. Usually strip them. "A customer asked about a wedding cake for Saturday" drafts exactly as well as the version with her full name and number in it. Test: if this conversation appeared somewhere you didn't choose, would this person mind?
- Anything under an NDA or a client contract. You may have signed something that already answers this question. Go and read it.
- Your unreleased work. Not a privacy issue — a judgment one. Decide deliberately rather than by habit.
- Whole documents when a paragraph would do. Most people paste the entire file because it's easier. Paste the part that matters.
Fine to paste
- The message you're replying to, with the identifying details stripped.
- Your prices, policies, lead times, refund terms — you publish these anyway.
- How you talk. Examples of your own writing. That's the good stuff, and it's what makes drafts sound like you.
- The situation, described plainly.
What "trains on your data" actually means
This is where the confusion lives, so here's the plain version:
It varies by provider, by plan, and it changes. Consumer plans and business/API plans usually have different defaults. Some providers train on chat content by default and let you turn it off. Some don't train on business-tier data at all. Some retain conversations for a period even when they're not training on them.
I'm deliberately not listing "provider X does Y" here, because that would be out of date by the time you read it, and a confidently wrong answer about your customers' data is worse than no answer. That's the whole argument of this site, applied to itself.
Do this instead — once, today, five minutes:
- Open your AI provider's settings and find the data controls. Read what the default actually is for your plan.
- Turn training off if it's on and you'd rather it wasn't.
- Find the retention policy. Know how long conversations stick around.
- If you're handling other people's data seriously — health, finance, anything regulated — that's a conversation with someone qualified, not with me and not with a chat window.
Do it once, and then you're deciding instead of hoping.
If you already pasted something you shouldn't have
Don't panic, and don't pretend it didn't happen.
- Delete the conversation if the provider lets you. It doesn't unring the bell, but it stops it being sitting there.
- If it was a credential — a password, a key, a token — rotate it. Today. That's the one that actually matters, and it's about ten minutes of annoyance versus a real problem later.
- Check your data settings so it doesn't happen by default again.
- If it was someone else's data and there's a legal obligation attached, get proper advice. I'm not a lawyer and this page isn't legal advice.
The bit that generalises
The reason this is a five-minute page and not a policy document: you are the control. Not a setting, not a toggle, not a promise on a pricing page. You decide what goes in, the same way you decide what goes out.
That's the same discipline as the gate, pointed the other direction.
→ The 20-Second Judge — the same discipline on what comes out.
→ Try the Reply Drafter, free — a memory document lives on your machine, not in a chat.