Privacy and data safety when using AI tools
Which data must never enter a chatbot, how personal accounts differ from enterprise ones, and how to control tool access — a checklist for individuals and teams.

What never to type into a chatbot
Chats may be retained or reviewed to improve products and can surface through account compromise. A simple rule: if seeing it published would hurt you, do not type it. Industry reports estimate close to 40% of workplace AI interactions contain sensitive data — much of it from employees’ personal accounts.
- Never: passwords and keys, confidential customer documents, users’ personal data, sensitive financials, NDA-covered code.
- Give synthetic or anonymized versions instead; the model does not need real names to analyze patterns.
- Turn on “do not train on my data” in every service’s settings where available.
Personal or enterprise account?
“Shadow AI” — employees using personal accounts for official work — has been the biggest persistent risk of recent years. An enterprise account changes three things: a data contract (no training on your content), admin control (who may use which tool with which data) and auditable logs. If you manage a team, these differences are worth every dollar of the enterprise tier.
- Write a three-line policy: what is allowed, what is forbidden, who answers questions.
- For highly sensitive data, self-hosted open-weight models are the middle path.
Keep tool access on a leash
Browser extensions, calendar/email connectors and MCP servers gain real power over your data. Treat every connection like lending your house key: grant only what is needed, review periodically, and revoke what is no longer used.
- Before installing any extension or MCP server, answer: what data does it access and where does it send it?
- Purge unused connections quarterly.
- For sensitive work, keep a separate browser profile with minimal extensions.