Security, privacy and compliance
Claude only works with what you put in reach: the folder you selected, the files you attach, and the connectors you approved. That is the boundary. It is not a promise that nothing leaves your laptop.
Think of it like giving a smart contractor their own desk, not the keys to the building. The folder is the desk. You hand over what the job needs. The rest of the laptop stays out of reach.
A normal chat is sent to Anthropic so the model can answer. A scheduled task set to run in the cloud runs on their servers. Treat both as you would treat any other work system, not as a private notebook.
This is practical guidance from sessions. It is not legal advice. If your policy is stricter, the policy wins.
How to access the settings
Desktop or web, then Settings. On a Team plan, some switches are owned by the admin, not by you.
The switch people miss: thumbs up and thumbs down. Rating a chat can send the whole conversation for review, and it may be used for training. Don't use them.
The second one: on Free, Pro, and Max, model training is on until you turn it off. Team and Enterprise are off by default. "I'm on a paid plan" is not the same as "they don't train on this".
What happens to a prompt
A plain version, so you can explain it to someone else.
- The message is encrypted on the way.
- The chat is stored as a record. Deleting it in the app may not wipe every copy, including memory and safety logs. Don't type something on the assumption you can unsend it.
- The model does not remember you by itself. Memory is a separate file you can turn on, read, and delete. If the organisation turns memory off later, people's saved memories can go with it. Don't flip that casually.
- Safety checks scan messages. Content that gets flagged can be kept for longer, and a person may read a sample.
- Rating a response (the thumbs) is a separate decision. Leave it alone.
- On Free, Pro, and Max, training is opt-out. Turn it off for work. On Team and Enterprise, training is off by default. Confirm that in your admin settings rather than assuming.
- Other doors exist. A browser extension can see the page you're on. On a Team plan, assume an owner can see chats, including one you thought was private.
Check Settings > Usage while you're in there. Limits and plan are in the same neighbourhood as the privacy switches. The labels move. If you can't find training or memory, look under privacy and capabilities before you conclude it isn't there.
Settings checklist
Do this once, on the account you use for work.
- Work happens on a Team or Enterprise account if the company has one. A personal Pro account full of client files is the wrong place.
- Training off, if you're on Free, Pro, or Max.
- Don't rate chats.
- Memory: decide on purpose. Useful for preferences. A bad place for secrets. You can delete memories. Read them once so you know what's stored.
- Connectors: read allow, write ask, delete never. See Connectors and permissions.
- Folder: your personal OneDrive or Drive, starred. Not your entire documents directory.
- Global instructions on, so the phone isn't a blank stranger. Keep them short and non-sensitive.
What can go in
Fine. Everyday drafts, meeting notes that aren't restricted, briefs, research, the context files from the plugin setup.
Think first. Client contracts, financials, HR cases, anything a client would expect you to be careful with. Check the company policy. Anonymise if you can still get a useful answer. "Would I be comfortable if this text appeared in the company channel with my name on it?" If not, don't paste it into a personal account.
Never. Passwords and API keys. Bank details. Government contract material. Anything your policy already forbids in other tools.
The folder rule still helps. No folder selected, and Claude is in plain chat. It isn't reading your disk. A folder selected, and it only sees what's inside that folder. That doesn't make a bad paste safe. You pasted it.
Recording and transcripts
Granola and Wispr Flow store what was said. That can be personal data.
- Company policy first. Some clients, and some sectors, don't allow it.
- Tell people at the start. Out loud. A note in a calendar invite that nobody read is not the same thing.
- If someone says no, don't record them.
- A person can ask for a copy of what you hold about them. Don't record a meeting you would refuse to hand over.
- Know how long your plan keeps transcripts before you promise the team a year of searchable history. See Granola.
Review before you send
Claude will write a confident sentence that isn't quite right. You are the last check on anything that leaves the building: an email, a number in a board pack, a clause, a name.
Don't turn on Always allow. Don't let a scheduled task send.
If something matters, open the source. A cited link is not the same as a link you've read.
Pro tips
- One work account. Don't mix client material into a personal login because the limit was kinder.
- Read the memory list once a month. Delete anything you wouldn't put in the folder.
- If IT asks what to approve, send them connectors, Granola, Wispr Flow, and Claude in Chrome. Named requests get further than "AI tools".
Limits
Anthropic's terms and the switches in Settings are the source of truth, and they change. This page is the habit. It isn't a certificate.
We can't see your admin console from here. If Team training, retention, or who can read chats matters for a contract, your admin needs to confirm it on your tenant. Don't take this page as that confirmation.
A local folder is not a private machine if the folder is synced, or if the task runs in the cloud. That's a reason to choose what you put in, not a reason to avoid the backup.
Where to start in 30 minutes
- Open Settings on the account you'll use for work.
- Turn training off if this is Free, Pro, or Max. On Team, confirm it's already off.
- Leave rating off. Look at memory and delete anything that shouldn't be there.
- Write five lines in claude.md: no invented facts, no sending, no deleting, no passwords, flag gaps.
- Pick one real document you were nervous about. Decide which bucket it's in: fine, think, or never. Don't upload it until you know.