Most managers I talk to aren't worried about AI taking their jobs. They're worried about the email they'll get from Legal after someone pastes a client's full name and project budget into a ChatGPT prompt.

It happens more than you think. And the problem isn't that people are careless — it's that nobody told them where the line is.

After working with dozens of teams on safe AI adoption, I've found that the managers who avoid problems all do the same thing on day one: they define three categories of information that never go into a prompt.

Here they are:

1. Real names

Client names. Employee names. Vendor contacts. Anyone identifiable. Instead, use role-based labels: "the project lead," "Client A," "the new hire." The AI doesn't need to know who — it needs to know the role and the situation.

2. Raw notes

Meeting notes, 1:1 feedback, performance reviews, Slack threads copy-pasted wholesale. These always contain more than you realize — offhand comments, internal tensions, sensitive decisions. Summarize first, then prompt. Three sentences of summary beats three pages of raw notes every time.

3. Private context

Internal strategy, unreleased product plans, revenue numbers, board discussions. If it would cause a problem on a screenshot, it shouldn't be in a prompt.

---

The pattern is simple: before you type anything into a chat tool, ask yourself — does this contain a name, a raw note, or private context? If yes, strip it or rephrase it first.

This takes about 15 seconds per prompt. It prevents 90% of the problems I've seen teams run into.

I wrote a free playbook that walks through this entire framework step by step — how to define your boundaries, how to rewrite prompts that are too revealing, and how to train your team to do the same. It's called Safe AI: The Ultimate Playbook and you can grab it right here for free.