SafePrompt Add to Chrome

Using ChatGPT at work without leaking client data

Updated , 7-minute read

Using AI at work is different from using it at home for one reason: most of what you'd paste isn't your data. It belongs to clients, employees, patients and customers who never agreed to have it sent to an AI company. The good news is that almost every useful work task can be done without the real details.

Why work use is riskier

  • It's someone else's data. One pasted contract, payroll line or patient note can expose people who have no idea it happened.
  • It's common. A 2025 report by LayerX found that employees who paste into AI tools paste sensitive company data almost four times a day on average, and that most AI use at work happens on personal accounts outside company controls.
  • It has consequences. In 2023, Samsung banned generative AI tools on company devices after employees reportedly uploaded internal source code to ChatGPT.
  • Professional rules apply. For lawyers, the American Bar Association's Formal Opinion 512 (July 2024) says lawyers should get a client's informed consent before entering information about the representation into a generative AI tool that can learn from it.

This isn't legal advice. Your company's policy and your profession's rules come first, so check them before using any AI tool with work material.

Start with the basics

  1. Know your company's policy. Some employers allow specific AI tools, some provide a business plan, some ban AI for certain data.
  2. Prefer a business account if your company has one. Business plans of ChatGPT, Claude and Gemini are generally not used for training by default, but your employer can still access what you type.
  3. Turn off training on any personal account you use (how).
  4. Replace real details with placeholders before you send (how). This is the step that actually keeps the data out.

Examples by role

Lawyers and legal teams

Mask: client and party names, case numbers, addresses, ID numbers, settlement amounts that could identify the matter.

Summarize the attached case notes and list next steps for [CLIENT] in the dispute with [OPPOSING_PARTY], case [CASE_NUMBER].

AI is helpful for structure, drafting and summarizing. The legal reasoning doesn't need the client's identity.

HR and recruiting

Mask: employee and candidate names, salaries, ID numbers, performance and health details, home addresses.

Draft a respectful letter to [EMPLOYEE] confirming the change in role from [OLD_TITLE] to [NEW_TITLE], effective [DATE], with a salary of [SALARY].

Finance and accounting

Mask: client names, account numbers, IBANs, card numbers, tax IDs, invoice numbers tied to a client.

Write a polite payment reminder to [CLIENT] for invoice [INVOICE], due [DATE], payable to account [IBAN].

Healthcare and care work

Mask: patient names, dates of birth, record numbers, addresses, and anything that links a condition to a person. If your organization is covered by health-privacy rules such as HIPAA, consumer AI apps are usually not approved for patient information at all, so follow your organization's rules.

Rewrite these visit notes for [PATIENT] into a clear summary for a family member, in plain language.

Customer support and sales

Mask: customer names, emails, phone numbers, order and account numbers.

Reply to [CUSTOMER] apologizing that order [ORDER] was charged twice, and confirm the refund to card ending [CARD].

Developers

Mask: API keys and secrets (rotate any key that leaks), customer emails and IPs in logs, internal hostnames.

Why does this query fail? SELECT * FROM users WHERE email = '[EMAIL]' AND ip = '[IP]'

A workflow that sticks

The method only works if it happens every time, including on busy days. That's why a tool that checks at the moment you press Send helps more than a policy document.

SafePrompt is a Chrome extension that does this on ChatGPT, Claude, Gemini, Grok, Copilot and DeepSeek. It warns before a prompt with personal data goes out, swaps names, emails, phone numbers, IDs, card numbers and bank details for placeholders in one click, and lets you mask anything else by highlighting it, such as a case or project name. When the answer comes back, the real values are restored on your screen only. Detection runs entirely in the browser, so client data never reaches SafePrompt either.

Rolling it out to a team? Email us and we'll help you set it up.

Sources