Privacy & data

Using AI well at work means being thoughtful about what you share and where it goes. This page covers the principles; always defer to your organization's policies and the official, current terms.

Know what you're sharing

Anything you type or upload becomes part of the conversation Claude processes. Before pasting, ask:

• Is this confidential (customer data, secrets, unreleased plans)?
• Is this personal data about real people?
• Am I allowed to share it with a third-party tool under our policies?

When in doubt, anonymize or omit; replace names and identifiers with placeholders.

How data is handled (general principles)

Data handling differs by plan and surface, so verify the specifics for yours:

• Consumer vs. business plans can have different defaults for data retention and whether content may be used to improve models.
• Team/Enterprise plans typically offer stronger guarantees, admin controls, and data governance.
• The API has its own data-handling terms, and some configurations support zero or limited data retention for sensitive workloads.

Don't assume; check

Defaults and options change. Confirm the current data-use, retention, and training settings for your plan in Anthropic's official documentation and your account settings.

Controls you have

• Privacy/training settings. Review whether your content can be used to improve models and adjust if available.
• Temporary chats. Use a non-persistent/incognito chat for sensitive one-offs so nothing is retained or remembered.
• Memory management. Memory can be turned off, viewed, and cleared.
• Connector permissions. Connectors/MCP should be granted the least access necessary, and reviewed regularly.

Practical guidelines for work

1. Follow your company's AI policy first. If there isn't one, ask before sharing sensitive material.
2. Minimize. Share only the portion needed for the task.
3. Mask identifiers. Swap real names/numbers for placeholders when the specifics don't matter.
4. Prefer business plans for business data. They're built for it.
5. Be careful with actions. Agents and connectors that can do things deserve extra scrutiny.

A simple rule of thumb

If you'd hesitate to paste it into an email to an outside vendor, hesitate here too, and check the policy first.

Next: understanding where Claude can go wrong in limitations & trust.