Key takeaways
- Start with the task rather than the entire document.
- Redaction reduces exposure but does not guarantee anonymity.
- Check the specific account, retention settings, and connected services.
- Review generated outputs before sharing or acting on them.
The email is ordinary: a customer wants an explanation, the wording is awkward, and an AI assistant could tidy it in seconds. Copy, paste, rewrite. The problem is what travels with the paragraph: a phone number, an invoice reference, an internal comment, perhaps a price that was never meant to leave the company.
AI privacy decisions often happen in moments like this. A shorter, more deliberate request can preserve the useful work while avoiding unnecessary disclosure.
1. Decide what the assistant actually needs
Write the task in one sentence before adding the source material. “Help me explain a delayed delivery politely” needs less information than “analyze this complete customer history.” Start with the smallest useful excerpt. Add context only when you can explain why it matters.
For example, a fictional request could say: “Draft a calm reply to a customer whose order is three days late. A replacement is available. Do not promise a delivery date.” That provides the writing problem without exposing a real account.
2. Confirm permission for this tool and account
A tool you are allowed to use for public marketing copy may not be approved for employee files or customer contracts. Check the organization’s rules and the service arrangement that applies to your particular account. If the information belongs to someone else, your ability to access it does not automatically mean you may share it with another service.
Do this before uploading. An impressive answer does not resolve a permission problem.
3. Remove details that do not help the task
Replace real people and organizations with consistent placeholders such as Customer A and Supplier B. Remove unnecessary addresses, contact details, account identifiers, credentials, and internal references. Then read the remaining text as a stranger would. Could the combination of facts still reveal the person or business?
OWASP identifies sensitive-information disclosure as a risk in AI applications and discusses access restrictions and sanitization as mitigations. These are layers of protection, not a promise that a rewritten document is anonymous. See OWASP’s guidance.
4. Check where the information can go
Review retention, training use, sharing controls, connected applications, and who can access the workspace. Treat those as separate questions. A setting that addresses model training is not, by itself, an answer about storage or access.
For a simple writing exercise, avoid attaching a whole folder or connecting a business system. Give the assistant the passage it needs. AWS’s data-security guidance provides further context on sensitive inputs and governance in generative AI systems.
5. Review what comes back
Read the output against the original task. Has the assistant invented a refund policy, repeated a confidential detail, or changed a tentative statement into a promise? A polished sentence can still be wrong. Restore real names only in the approved final working environment, after the wording is checked.
A small habit worth keeping
Keep a reusable prompt with placeholders for routine tasks. A support team, for example, could agree on a template containing the issue, the permitted resolution, the desired tone, and facts that must not be changed. That makes careful handling easier under pressure.
The YNIA tools directory includes a Prompt Privacy Cleaner for recognizing common sensitive patterns. Use it as a review aid. It cannot identify every business secret or decide whether you have permission to disclose something.
Before pressing send, ask: “Would the task still work if I removed one more detail?” Often, it will.
Frequently asked questions
Is removing a customer name enough?
Not necessarily. A location, unusual order, job title, or combination of details can still identify someone.
Can I use public information in an AI prompt?
Public information is often a lower-risk starting point, but account rules, licensing, and the intended use still matter.
Does a privacy cleaner make a prompt safe?
It can flag recognizable patterns. It cannot reliably understand every confidential fact or authorize disclosure.