How to Use AI Without Sharing Sensitive Data
AI tools can save time, but convenience is not a reason to expose confidential information. The safest workflow begins before the prompt: classify the data, remove what the system does not need, and choose an approved account and tool.
Start with data classification
Use a simple classification scheme:
- Public: already approved for anyone to see.
- Internal: ordinary business information not intended for publication.
- Confidential: customer records, contracts, pricing, credentials, investigations, or non-public strategy.
- Restricted: highly sensitive personal, legal, financial, health, security, or regulated information.
Public content is usually low risk. Internal content may require an approved business account. Confidential or restricted content should not be submitted unless the organization has explicitly approved the tool, configuration, purpose, and controls.
Use the minimum necessary information
An AI system rarely needs a full document to help with one paragraph. Extract the relevant section. Replace names, IDs, addresses, account numbers, vehicle plates, phone numbers, and exact commercial figures with neutral placeholders.
For example:
“Customer A rented Vehicle B for [DAILY RATE]. The agreement includes a [TYPE OF CLAUSE]. Draft a neutral list of questions for legal review.”
Redaction is more than hiding a name
A person can sometimes be identified from a combination of details. Remove unnecessary dates, locations, job titles, unique events, and document metadata. Screenshots may contain browser tabs, email addresses, or file paths. PDFs may include comments, hidden layers, or author information.
Separate drafting from facts
Ask the AI to improve structure using placeholders. Insert the real facts yourself in the approved system after the draft is complete. This reduces exposure and makes the review step explicit.
Check account and retention settings
Consumer, business, and enterprise accounts may have different controls. Before using a tool for work, confirm who owns the account, whether prompts may be used to improve services, how long data is retained, which connectors are enabled, and whether shared links can expose content.
Do not paste secrets
- Passwords, recovery codes, private keys, API keys, or access tokens
- Complete payment-card or bank information
- Unredacted identity documents
- Private medical or legal records
- Security vulnerabilities that have not been safely disclosed
A pre-submit checklist
- Is this tool approved for this type of data?
- Can the task be completed with less information?
- Have identifiers and metadata been removed?
- Does the prompt contain credentials or secrets?
- Could the output reveal or reconstruct sensitive information?
- Will a qualified person review the result?
Example workflow: summarize a complaint
Copy only the relevant complaint text. Replace the customer’s name with “Customer A,” the booking number with “[BOOKING ID],” and exact dates with relative timing if exact dates are not needed. Ask for a five-point summary, unresolved questions, and a neutral response outline. Review the output against the original and compose the final reply inside the company’s approved system.
When not to proceed
If the task requires the AI to see restricted data to be useful, stop and use an approved internal system or a qualified professional. Productivity does not override confidentiality, contractual obligations, or law.