AI policy

What is safe to paste into an AI tool

August 15, 2026 · 5 min read

The safest AI policy is not "never paste anything into it." Nobody follows that for long, because it turns a useful tool into a guessing game.

A better rule is to sort information into three buckets: public, internal, and restricted. The bucket tells you how cautious to be before you paste.

This is the individual version of the owner-side policy question. For the company setup, see keeping company data out of AI training sets. This article is the rule I want an employee to remember while they are actually working.

Bucket one: public or already shareable

This is the easy bucket.

• Public website copy
• Product descriptions already online
• Marketing drafts that do not include customer information
• Generic job descriptions
• Publicly available competitor information

If you could put it on your website, send it to a vendor, or hand it to a freelancer without concern, it is usually safe to use in a normal AI workflow.

Bucket two: internal working material

This is where most useful work lives, and where most confusion happens.

Internal working material includes things like process notes, draft SOPs, meeting outlines, anonymized examples, spreadsheet formulas, and non-sensitive templates. It is not public, but it also is not automatically dangerous.

The question is not just "is this internal?" The question is "would this cause harm if it left the company, and is this tool approved for that kind of material?"

For example, a blank invoice template is usually fine. A real invoice with customer names, addresses, prices, and account numbers belongs in the next bucket.

Bucket three: restricted material

This bucket needs a much higher bar. Do not paste it unless your company has explicitly approved the tool and the workflow for that kind of data.

• Customer names, emails, phone numbers, addresses, account numbers, and order history
• Private employee information, payroll, medical, disciplinary, or HR material
• Passwords, API keys, tokens, credentials, or screenshots that show them
• Contract-controlled information, especially if a customer or vendor agreement limits sharing
• Regulated information in healthcare, finance, legal, education, or similar contexts
• Trade secrets, unreleased financials, acquisition plans, and anything you would not put in an ordinary email

This does not mean AI can never touch restricted material. It means the decision belongs at the system level, not in a one-off chat box.

The rewrite trick

Most people do not need to paste the sensitive version. They need help thinking about the situation.

Instead of pasting:

Here is the full customer email thread with names, addresses, order numbers, and the refund dispute. Write a reply.

Rewrite it as:

A customer ordered the wrong item because our product note was unclear. They want a refund and are frustrated. Draft a polite reply that apologizes, offers a return label, and says we are updating the product note. Do not invent policy details.

You removed the identifying data and made the prompt better at the same time.

When in doubt, ask these four questions

1. Does this identify a customer, employee, vendor, or account?
2. Would I be comfortable emailing this to an outside contractor?
3. Is this governed by a contract, regulation, or confidentiality promise?
4. Can I rewrite the example so the useful context stays and the sensitive details leave?

If the answer to the first three is yes, slow down. If the answer to the fourth is yes, rewrite it first.