A chatbot is a convenient workspace, but not every document belongs in it. The safe default is to share only the minimum information needed for the task and to confirm the vendor’s controls before using sensitive material.
This guide is general educational information, not legal advice. Organizations handling regulated or confidential data should consult their internal policies and qualified advisers.
Keep these out of casual AI prompts
- Passwords, private keys, access tokens and recovery codes.
- Unpublished financial results or confidential contracts without authorization.
- Customer identity records, payment information and personal contact details.
- Private medical, legal, employment or student records without an approved process.
- Proprietary source code or business plans subject to confidentiality obligations.
- Material you lack permission to upload or process.
Even if a product offers privacy controls, your organization’s rules may impose additional restrictions. Verify them before use.
Use synthetic examples during evaluation
To test a customer-support prompt, invent a fictional customer, order number and issue. To test summarization, use a public report or your own non-sensitive writing sample. Synthetic data can reveal whether a workflow works without exposing information you do not need to share.
Be careful with “redaction”: removing a name does not automatically make a document anonymous. Combinations of dates, locations and unusual details may still identify someone.
Read the settings, not just the homepage
Inspect a tool’s policies for content retention, use of prompts for model improvement, administrative access, regional processing, file deletion and connected app permissions. Look for differences across free, individual and business plans. Features can change, so rely on the current official documentation.
Connected apps create another layer of risk
When connecting an inbox, calendar, drive or customer database, review exactly what the integration can read or change. Grant the narrowest scope practical, remove unused integrations and periodically audit connected access.
An automation that drafts responses is usually lower risk than one that can send, refund, delete or publish without review. Keep consequential actions behind a human approval step until the process has proven reliable.
A five-minute preflight checklist
- Purpose: Is AI appropriate for this task?
- Data: Can I replace the input with synthetic or public information?
- Permission: Am I authorized to share the remaining data?
- Vendor: Have I verified retention, training and access settings?
- Output: Who checks accuracy before the answer is used?
If sensitive information was already submitted
Follow your organization’s incident-response process. Identify the data involved, review the vendor’s available controls, revoke exposed credentials where applicable, and seek advice from the responsible security or privacy contact. Avoid assuming that deleting a chat removes every copy of the information.
Responsible AI begins with good judgment about what a tool should see. Pair this checklist with our Beginner Automation Guide to design a safer workflow from the start.