Use AI with sensitive data
A practical method for deciding what data to share, reducing exposure, and choosing safer workflows.
AI tools are often most useful when they receive real context, but real context may contain personal, confidential, or regulated information. Before sharing it, decide what the task actually requires and what exposure the workflow creates.
Classify the information first
Look for personal identifiers, credentials, private communications, customer records, internal financial data, source code, legal material, and information covered by contracts or regulation. Do not assume that removing a person’s name makes a detailed record anonymous.
If your organization has a data policy, it takes priority over convenience.
Minimize before you send
Share the smallest useful excerpt. Replace names and identifiers with neutral labels. Remove metadata, unrelated rows, hidden spreadsheet columns, document comments, and revision history.
For many tasks, the model needs the structure rather than the real values. A synthetic example can be enough to develop a formula, prompt, or processing workflow.
Understand the service boundary
Check which account and product you are using, where data is processed, how long it is retained, whether it may be used to improve the service, and who can access conversation history. Consumer and managed business offerings may have different controls.
Do not paste secrets such as passwords, private keys, session tokens, or recovery codes. If a secret is exposed, treat it as compromised and rotate it.
Prefer controlled retrieval
For repeated work, use a system that retrieves only the documents relevant to the current request and enforces the user’s existing access rights. This is safer than copying a whole shared drive into one conversation.
Review outputs for leakage
Generated summaries and drafts can repeat sensitive details. Apply the same classification and sharing rules to the output as to the input. Before forwarding or publishing, check names, embedded quotations, links, and file metadata.
The core rule is simple: reduce the data, narrow the audience, and keep an approval step before information leaves its intended boundary.