How to Use ChatGPT at Work Without Sharing Sensitive Information
A practical privacy checklist for using ChatGPT on work tasks, from redacting documents to choosing the right workspace settings.
An assistant can save you twenty minutes on a routine email and create an afternoon of trouble if you paste the wrong thing into it. That is the trade-off many teams are working through. The issue is rarely whether AI is useful. It is whether a useful shortcut sends information somewhere it was never meant to go.
This guide is for people who use ChatGPT to draft emails, organize notes, think through spreadsheets or prepare customer-facing material. It isn't a substitute for your employer's security policy. It is a way to make everyday decisions with fewer surprises.
Start with the question nobody asks: whose data is this?
A document can look harmless because it has no dramatic secrets. Yet a quotation may contain a client's budget. A support ticket may include an address. A project update may expose an unreleased product. Even a plain email chain can contain private phone numbers and negotiation details.
Before copying text into an AI tool, separate what you own or are authorised to use from what you merely have access to. Those are not the same thing. Access to a customer spreadsheet does not automatically authorise uploading it to a third-party service.
Use a three-bucket rule:
- Public: published product documentation, your own public blog copy, publicly released announcements. Normally the least sensitive material, though copyright still matters.
- Internal: meeting notes, team plans, unpublished drafts and non-public performance data. Obtain approval and use the organisation's approved tools.
- Restricted: passwords, API keys, customer records, personal information, legal advice, financial account details and confidential contracts. Do not paste these into a personal AI chat as a convenience.
Some organisations approve specific enterprise AI workspaces for certain internal or confidential material. Follow that approved scope rather than assuming every ChatGPT account has the same rules.
Know what the privacy controls do—and what they don't
OpenAI documents separate controls for personal accounts, including model-improvement settings and Temporary Chat. In supported personal ChatGPT settings, you can find Data controls and review whether new conversations may be used to improve models. Temporary chats are handled differently from regular chats: they do not appear in normal chat history and are not used to improve models while temporary, but may still be retained for safety reasons.
Those controls matter. They are not a magic confidentiality agreement. An organisation may need contractual protections, access controls, retention requirements and an approved workspace. Read the current provider documentation rather than making a decision based on a screenshot from last year. See OpenAI's consumer-data explanation and its data controls guide.
A 60-second redaction routine
Imagine a sales manager wants help writing a follow-up about a delayed delivery. Their original draft mentions a named customer, purchase order number, shipping address, quoted price and a dispute about payment. A better approach is to remove the identifying facts before asking for writing help.
Before: 'Draft a reply to Greene Labs about PO 48391, the $47,850 order, and their overdue invoice.'
Safer: 'Draft a calm, concise business email about a delayed shipment and an outstanding invoice. Apologise for the delay, explain that an updated delivery date will follow, and ask to schedule a call. Do not invent dates or promises.'
The second version preserves the writing task but not the confidential record. Add your verified names and numbers later, inside the approved work system.
A simple checklist:
- Remove names, email addresses, phone numbers and physical addresses unless there is clear approval to share them.
- Remove customer IDs, account numbers, contract values, access tokens and unique project codes.
- Replace sensitive events with neutral descriptions without changing the nature of the task.
- Ask for a structure, options or editorial improvement—not a final factual claim about your customer.
- Review and finish the output in the system where the real information belongs.
Choose tasks where private data adds little value
Good uses often rely on reusable patterns. Ask for three versions of a follow-up email. Ask for a troubleshooting checklist for a generic API timeout. Ask for a meeting-agenda template or a formula explanation using invented rows.
Higher-risk uses involve raw personnel files, medical records, legal disputes, source code with secrets, private customer conversations and highly sensitive business plans. In those cases, stop and check what your organisation permits. For UK teams, personal data may trigger data-protection obligations; US organisations may face contractual, sector-specific or state-level duties. This is an operational caution, not legal advice.
A prompt you can use without copying the document
I need help improving a business message. I will describe the situation without identifying information. The goal is [goal]. The recipient is [role]. Our confirmed facts are [non-sensitive facts]. Draft two options: one direct and one warmer. Use placeholders for any dates, prices or names. Do not invent commitments. After drafting, list what I need to verify before sending.
The last line is important. Good AI usage is not only producing a draft; it is catching what the draft might have made up.
If you have already shared something sensitive
Do not rely on simply deleting visible chat text as a complete response. Follow your organisation's incident-reporting procedure. Depending on the situation, the response may include revoking exposed credentials, informing a security lead, assessing what was uploaded and checking provider retention controls. Speed and accurate disclosure are more useful than trying to hide an honest mistake.
Frequently asked questions
Is Temporary Chat the same as a private company workspace?
No. Temporary Chat is a conversation mode with documented handling rules. A managed business workspace may have different organisational controls, agreements and policies. Check the exact product and plan before making sensitive-data decisions.
Can I paste an NDA if I remove the company name?
Not necessarily. The details can still identify the parties, and the NDA may restrict third-party disclosure. Seek permission or work from a generic summary that contains none of the protected terms.
Is it safe to use AI for code reviews?
It depends on the code and the authorised environment. Never include live credentials, and consider proprietary code, security configurations and customer data before sharing a snippet.
The practical goal is simple: let AI help with the shape of the work while keeping the actual confidential facts where they belong. That habit takes less time than repairing a data leak.