Using AI at work safely: what never to paste into ChatGPT

Using AI at work saves hours. But everything you paste into an assistant leaves your computer and goes to another company’s servers. Depending on the tool and the plan, that content may be stored, reviewed by people or used to train new models.

None of this stops you from using AI. It just calls for the same care you already take when emailing someone outside your company.

What never to paste into an AI assistant

• Personal data about clients or coworkers, such as ID or social security numbers, home addresses, phone numbers, emails and health information.
• Financial data, such as account and card numbers, balances, invoices and banking details.
• Passwords, API keys and tokens, including inside code snippets or config files.
• Contracts and confidential documents that are not public yet.
• Strategy and internal numbers that would cause harm if leaked, such as launch plans, targets and results not yet announced.

A simple rule: if you would not paste the text into a group chat with people outside your company, do not paste it into the AI either.

How to get help without exposing data: swap in examples

For most tasks, the AI does not need the real data to help. It only needs to understand the shape of the problem. Replace the data with generic versions before pasting.

Instead of:

Write an email chasing Alpha Construction Ltd., account 4471,
for invoice #4471 of $48,300.00, due on October 10.

Use:

Write a polite but firm email chasing a client for an invoice that is
20 days overdue. Include the invoice number, amount and a payment deadline
as fields for me to fill in: [number], [amount], [deadline].

The email comes out the same, and you fill in the real details on your own computer at the end.

Other swaps that work well:

• replace people’s names with “Person A”, “the client” or “the manager”;
• replace the company name with a description, such as “a mid-sized construction company”;
• replace exact figures with rounded or made-up numbers of the same magnitude;
• instead of the whole spreadsheet, send only the column names and 3 made-up example rows.

Company account or personal account?

Whenever your company provides a corporate AI tool, use it for work. Business plans usually have stricter rules on how data is stored and used, and the company can control who has access to what.

Using your personal account for company tasks mixes the two and, in many companies, breaks internal policy. If in doubt, ask your IT team which tools are approved.

Be careful connecting email, calendar and files

Many assistants can connect to your email, calendar, Google Drive or OneDrive. That is powerful, but it means giving the tool access to everything in there.

• Grant only the permissions the task needs.
• Prefer connecting work accounts approved by your company.
• Review what is connected every so often and disconnect what you no longer use.

AI makes mistakes, and makes them confidently

AI assistants sometimes invent information that sounds true, such as numbers, names of laws, quotes and references. This mistake is called hallucination.

• Check any figure that will feed a decision against a reliable source.
• For legal, financial or health topics, confirm with an expert.
• Ask the AI to flag uncertainty. A rule like “if you do not know, say so” in your prompt helps a lot.
• Reread everything before sending it to clients, managers or anyone outside the company.

Checklist before you send a prompt

Ask these 6 questions before hitting Enter:

1. Is there any personal data, financial data or password in the text? If so, swap it for an example.
2. Would I show this to someone outside the company? If not, do not send it.
3. Am I using the approved tool, with the right account?
4. Do connected tools have only the access they need?
5. Will I review the answer before using it?
6. Does any fact in the answer need to be checked elsewhere?

Wrapping up

AI helps a lot with writing, summarizing and organizing, as long as you decide what it sees. Swapping real data for examples almost never makes the answer worse and removes most of the risk.

To build clear prompts without exposing anything, use the prompt builder. In the context and limits pillars, describe the situation in general terms and tell the AI what not to make up.