Is It Safe to Put Company Data in ChatGPT? What to Strip Out First

TL;DR: It's safe to put company data in ChatGPT or any other AI assistant only when three things are true: you're using the tool and account your company approved, you've sorted the text and kept out anything that must never leave your systems, and you've replaced the personal details and secrets that are left with placeholders. Sort first, strip second, paste last, and put the real values back in after the answer comes out.
If you're asking whether it's safe to put company data in ChatGPT, you may have already done it. You had a long email thread, a messy spreadsheet or a customer complaint, and the assistant made the job faster. This lesson shows you how to decide, before you paste, which work text can go into an AI tool, what to strip out of it, and what to do if something slipped through. It works the same way for any AI assistant, and you don't need any technical background.
Our lesson on AI for administrative tasks has a quick strip-out checklist. This one adds how to sort a document you're unsure about, how to keep text useful after redaction, and how to recover from a mistake.
Why this question matters now
Pasting work into AI tools is common, and much of it happens outside the tools companies have approved. Verizon's 2026 Data Breach Investigations Report found that 45% of employees "are now considered regular users of AI (authorized or not) on their corporate devices, up from 15% in the previous year", and that, among people using AI services on their corporate devices, 67% were using non-corporate accounts to access them (Verizon 2026 DBIR, key findings as published in its Healthcare snapshot; these are findings across the whole report, and Verizon's release on the 2026 DBIR reports the 15% to 45% rise too). The same report calls this use of unapproved tools "Shadow AI" and found it was "the third most common non-malicious insider action detected in our data loss prevention (DLP) dataset in 2025."
The report also shows what people sent. In untrusted data loss prevention events involving generative AI tools, source code was the most common type (28%), followed by images (16%), structured data (14%), documents (13%) and PDFs (10%). The report classes this as a non-malicious action, so the fix starts with a habit people can follow before they paste. The glossary entry on shadow AI explains the term.
What can happen to text after you paste it
When you paste text into a hosted AI assistant such as ChatGPT, it leaves your computer and goes to the company that runs the tool. What happens next depends on the tool, your plan, your settings and your company's agreement with the vendor. A provider may keep the conversation for a period, people with access to the account may be able to see it, and some services may use conversations to improve their models unless that's switched off. You can't tell which applies by looking at the chat window, which is why the approved tool matters.
There's a second risk that has nothing to do with storage. The US National Institute of Standards and Technology lists data privacy as one of the main risks of generative AI and notes that "Models may leak, generate, or correctly infer sensitive information about individuals" (NIST AI 600-1, Generative AI Profile). In plain terms, a model can piece together details about a person from clues, even if you never typed their name.
Data protection law gives you a useful rule of thumb. The UK GDPR says personal data must be "adequate, relevant and limited to what is necessary in relation to the purposes for which they are processed", a principle known as data minimisation (ICO, Principle (c): Data minimisation). For your prompt, that means a simple question: does the assistant need this detail to do the job? If it doesn't, take it out.
Sort it before you paste: green, amber and red
Before you strip anything, decide which of three groups the text belongs to. It takes seconds and stops you redacting something that should never go in at all.
Green: paste as it is
- Text that's already public, such as your company's website copy or a published press release
- Your own wording with no names, numbers or client details in it, such as a draft job ad or a general how-to
- Questions about how to do something ("how do I write a polite reminder about an overdue form?")
Amber: strip it first, then paste
- Emails and meeting notes that mention customers, colleagues or suppliers by name
- Spreadsheets with names, emails, phone numbers or account numbers in some columns
- Internal documents that are fine to share inside the company but name people or clients
Red: keep it out entirely
- Passwords, login codes, access keys, and links that open private files
- Health, pay, disciplinary or HR records
- Card numbers, bank details and full ID numbers
- Contracts, prices or plans your company has promised to keep confidential
- Anything your company's AI policy or a client agreement says can't go into an AI tool
If text is red, don't try to redact your way to green. Do that task by hand, or ask your manager whether there's an approved tool built for it. If your team doesn't have a written policy yet, the AI acceptable use policy template is a good place to start, and the glossary entry on AI acceptable use policies explains what one covers.
What to strip out, and what to put in its place
For amber text, swap each sensitive detail for a placeholder such as [CUSTOMER_1] or [EMAIL_2] instead of deleting it. Placeholders keep the shape of the text, so the assistant can still write "Dear [CUSTOMER_1]", and you can put the real values back afterwards.
Strip these from amber text:
- Names of people and companies, when the text also says something about them
- Contact details: email addresses, phone numbers, street addresses
- Identifiers: account, order, policy, invoice and employee numbers
- Dates that identify someone, such as a date of birth or the date of a medical appointment
- Secrets that crept in: a password in a pasted config file, an access key in an error message, a private link in an email signature
Our free PII redactor does the first pass for you. It runs in your browser tab and sends nothing anywhere. It finds email addresses, phone numbers written with separators, card numbers, IBANs, IP addresses, US Social Security numbers and common key and token formats, swaps each one for a numbered placeholder, and swaps the real values back into the assistant's answer afterwards. It matches patterns, so it can't spot a person's name or a street address on its own. Type the names you care about into its "Your own terms to hide" box, and always read the result before you paste. The glossary entry on PII redaction explains the wider practice.
The context test
Removing names isn't always enough. "The only finance manager in the Leeds office, who came back from leave in March" identifies one person as clearly as a name does. After you've redacted the text, read it once and ask: could someone who knows the company work out who this is? If the answer is yes, make the details vaguer ("a manager in one of our offices") or leave them out.
A worked example: redact, ask, restore
Say a customer has sent a long complaint and you want help with a calm, clear reply. You type the customer's name and the order number into the redactor's "Your own terms to hide" box and run the email through it. It swaps the email address and phone number it finds, plus the terms you typed, for placeholders such as [EMAIL_1], [PHONE_1], [TERM_1] and [TERM_2]. Your terms are numbered in the order they appear in the email, so in this one the customer's name, which comes first, becomes [TERM_1] and the order number becomes [TERM_2]. Then you paste this:
The Prompt:
I work in customer service. Below is a complaint from a customer.
Personal details have been replaced with placeholders like
[TERM_1] and [EMAIL_1].
Write a reply under 200 words that:
1. Thanks the customer and says sorry for the delay.
2. Explains the next step: we will check [TERM_2] and reply within
two working days.
3. Keeps every placeholder exactly as written. Do not invent names,
numbers, dates or promises that aren't in the email.
If you need a detail I haven't given you, write [NEED: what] instead.
Complaint:
[paste the redacted email]
Why This Works: Telling the assistant what the placeholders are stops it "fixing" them into made-up names. The rule against inventing promises protects you from a reply that offers a refund nobody approved, and [NEED: what] turns a gap into a visible question.
What you get back: A short reply with the placeholders still in place. Paste it into the redactor's second box to swap the real values back, read it once more, and send it from your own email.
Use the right tool and the right account
A personal account and your company's managed account can look identical and still run under different terms for storage, training and who can see the history. That's the gap the Verizon figure on non-corporate accounts points to. Use the login your company approved for work. If it hasn't approved any tool, ask before you paste anything amber, and keep to green work until then.
If you already pasted something you shouldn't have
- Stop and don't paste more. Close the conversation.
- Delete the conversation from the tool's history if it lets you. This may not remove every copy, so keep going.
- Change any password or key that was in the text, straight away. Treat a secret that has left your control as exposed.
- Tell your manager or IT what went in, which tool and which account. If personal data was involved, your company may have reporting duties, and it can only meet them if it knows.
For more, see the office and admin track and the first-week guide to using AI at work.
Frequently asked questions
Is it safe to put company data in ChatGPT?
It can be, if you use the tool and account your company approved and strip out personal data, secrets and confidential terms first. Passwords, health and pay records and anything under a confidentiality agreement shouldn't go in at all.
What should you never put into an AI chatbot at work?
Passwords, login codes and access keys, card and bank details, full ID numbers, health, pay and HR records, and anything your company or a client agreement says must stay confidential. If you'd hesitate to email it to a stranger, keep it out.
Is removing names enough to make data anonymous?
Not always. Job titles, locations, dates and unusual details can still point to one person, and NIST notes that models can infer sensitive information from clues. Read the redacted text and ask whether someone who knows the company could tell who it's about.
Can I use my personal ChatGPT account for work?
Only if your company's policy allows it. Personal and company accounts can have different settings for storage, model training and who sees the history, and your company can only protect data in the tools it manages.
What should I do if I pasted confidential information into AI by mistake?
Stop, delete the conversation if the tool allows it, change any password or key that was in the text, and tell your manager or IT the same day. Quick reporting lets your company contain the problem.
Terms in this article
From the library
Rather have someone set it up for you?
That's a fair call. LTFI, the agentic business operating system from Kief Studio, can set up the work these guides describe and run it with you. No website yet? Kief Studio will build you one for free.

