Aegis Boardroom
Tier 0 · Module 0.3 · 30 minutes

What AI should not touch

What you'll get out of thisLearner rejects three sample prompts that would expose customer or financial data, and explains why.

Lesson

The rule is short, hard, and matters more than anything else in this course:

What you put into a public AI tool may be used to train that tool, may be visible to the vendor, and may surface in someone else's output later. Treat anything you type into an AI tool the way you'd treat saying it out loud at a competitor's office.

That rule has nuance. Different vendors have different terms. Enterprise plans behave differently than free plans. Your company may have a specific approved tool. But until you know your company's policy, default to the rule above.

The five categories that don't go into AI

1. Customer personal data. Names, emails, phone numbers, addresses, social security numbers, customer account numbers, credit card numbers. If you're working on a customer email, never paste their full record into AI to "give context."

2. Customer financial data. Account balances, transaction history, payment status, credit information, tax documents. Even if you can't see a real consequence, treat this category as off-limits unless your company has explicitly approved an enterprise AI tool with appropriate contracts.

3. Employee data. Salaries, performance reviews, disciplinary actions, medical or disability information, immigration status, anything from HR.

4. Trade secrets and proprietary IP. Pricing models, vendor contracts, formulas, recipes, manufacturing processes, code, customer lists, contract terms with major customers.

5. Regulated information. Patient health information (HIPAA), student records (FERPA), financial advice (FINRA), classified or government-restricted information, anything covered by NDA or specific contractual confidentiality.

What about an enterprise AI tool?

Many companies are moving to enterprise versions of AI (ChatGPT Enterprise, Microsoft Copilot Enterprise, Claude for Enterprise, Google Gemini Workspace) that contractually do not train on your input. If your company has approved one, the rules are different. Your IT or operations leader should tell you:

  • ·Which AI tool is approved
  • ·What data categories are still off-limits even on the approved tool
  • ·Where to ask if you're not sure

If your company has not yet made that policy clear, ask. A 10-minute conversation prevents an avoidable incident.

The "would I email this to a stranger?" test

Before you paste anything into an AI tool, read what you're about to paste and ask: would I email this to a random person I just met at an industry event?

If the answer is no, edit it down before you paste, or use the approved enterprise tool, or don't paste it.

In-module exercise (15 minutes)

Below are six sample prompts. Three are safe, three are risky. Identify which are which and explain why.

Prompt 1

"I'm writing a follow-up to a customer who didn't sign the proposal we sent last week. The proposal was for tree removal services at their home. Help me draft a friendly check-in email."

Safe? Risky? Why?

Prompt 2

"Here's the full transcript of my call with John Smith at Acme Corp, including his cell number and what he said about being unhappy with our pricing. Summarize the key points for my manager."

Safe? Risky? Why?

Prompt 3

"We're updating our employee handbook policy on remote work. Help me rewrite this paragraph in plain language: [paragraph from current handbook]."

Safe? Risky? Why?

Prompt 4

"I have a list of 200 customers and their account balances. Help me identify which ones are overdue and draft outreach emails."

Safe? Risky? Why?

Prompt 5

"I'm writing a job description for a new sales coordinator role. Here's the rough description I've drafted. Help me make it more compelling and inclusive."

Safe? Risky? Why?

Prompt 6

"Here's our pricing model for service contracts, including margin assumptions. Help me think through whether to raise prices for new customers next quarter."

Safe? Risky? Why?

Answer key

#VerdictWhy
1SafeNo customer-identifying data shared. Generic context.
2RiskyCustomer name, employer, cell number, content of conversation. Customer PII plus business confidential.
3SafeInternal policy text, no PII, no proprietary data. Common, low-risk use.
4RiskyCustomer financial data, named accounts. Hard no.
5SafeInternal job description, no PII. Solid AI use case.
6RiskyTrade-secret pricing model and margin data. Use an approved enterprise tool or do this offline.

Knowledge check

  1. ·The default rule is: (a) AI is private; (b) treat AI input like saying it out loud at a competitor's office)
  2. ·Which is safe to paste into a public AI tool: (a) customer email thread with full names and account numbers; (b) draft job description with no employee data)
  3. ·If your company has approved an enterprise AI tool: (a) all rules are off; (b) some categories are still off-limits. Check the policy)

Answers: 1. (b). 2. (b). 3. (b).

What's next

Module 0.4 teaches you how to ask AI for useful work. Context, specifics, and the "give it a job description" pattern.

Run this with your team

Tier 0 is the activation layer for companies that do not have a CTO, CIO, CISO, or internal AI owner. Plain English. Practical. Safe.