Data protection in daily work
What belongs in a chat, what your workspace's protection level does, when Strict or security mode fits and how to handle attachments, images and sharing.
Custodos replaces detected personal data with placeholders before sending, once your admin has turned this on, and offers security mode for sensitive conversations. The rest is a matter of habits.
Know what your workspace protects
Which protection level applies is shown in your profile under How your data is handled here.
| Protection level | Replaced before sending | Readable for the model |
|---|---|---|
| Off | Nothing. Every new workspace starts this way. | Everything |
| Standard | AHV, tax ID and social security numbers, IBAN, card number, UID/VAT ID | Names, email addresses, phone numbers, addresses, images |
| Strict | Every category, including names and contact details. Hard identifiers and uploaded images are blocked. | Whatever detection misses |
| Custom | Set per category | Depends on the settings |
The most important point: under Standard, names, email addresses and phone numbers stay readable. That includes your own name, which goes along in chats and projects. Under Strict and in security mode it is replaced. See How pseudonymisation works.
What belongs in a chat
| Content | Example | Recommendation |
|---|---|---|
| General content without personal data | Draft text, an Excel formula | Ordinary chat |
| Business content with a few names | Email to a client, minutes | Ordinary chat if your company allows names in clear text, otherwise Strict |
| Client and personnel files, health and pay data | A reference letter, a payroll list | Security mode, if your admin offers it |
| Credentials | Passwords, e-banking codes, API keys | Never |
Only include what the task needs. For an email about missing documents, the model does not need an AHV number.
Strict, security mode or a group
| Situation | Fits |
|---|---|
| Customer data day to day, files only rarely | Standard |
| Names should never reach a model in clear text | Strict |
| A single conversation is particularly sensitive | Security mode: a fixed model, everything pseudonymised, no web search, no connectors |
| Almost every conversation concerns a client engagement | An admin makes new chats start in security mode |
| Only one team is sensitive, such as payroll | A group (on request) with Strict data protection, Enforce security mode or Switzerland only |
Turn security mode on with Turn security mode on before the first message. Once the chat has messages, it cannot be switched off there. Such a chat processes in Switzerland only if your admin chose a Swiss model. See Security mode and Groups.
Using pseudonymisation well
- Watch the note below the message box. It shows how many values were detected; a click opens the list.
- Expect gaps. A surname without a first name or title can be missed, and company names usually are. Write "Ms Keller" rather than "Keller", or add the value in the list with Add. It is then replaced throughout the chat, attachments included.
- Recurring values as a rule. An admin can add client names or customer numbers under Your rules.
- Send original only for a reason. If a product name is replaced by mistake, send it unchanged with Send original. This is not possible in security mode and never for blocked values.
- Check when in doubt. Show what the model received under your message shows what arrived.
The flow with pictures is in Protecting personal data in chat.
Attachments, images, web search
- Documents such as PDF, Word or Excel are checked like your message, scanned pages after text recognition.
- Images cannot be checked. Under Off and Standard the model sees everything on them: black out names and account numbers first. Under Strict and in security mode you cannot attach images.
- Generating images: names in an image description go to the image model as typed, even under Strict. No images are generated in security mode.
- Web search: only the search query goes to the search service, never the chat. Custodos does not send a query with a hard identifier, but under Standard names do go out. See Web search.
Sharing and brains
- Chats with client data are best shared only with Selected people. Attached files stay with you. Heed the warning when a chat contains answers from a brain the recipients cannot read. See My shares.
- Prompts and agents saved with Whole workspace can be read by everyone. Remove personal data from the text, instructions and files.
- A brain has its own level. A brain set to Open passes its passages on unchanged, even if the workspace pseudonymises. See Building a good brain.
Frequently asked questions
No. Admins see your credit usage, not the content of your chats, unless you share a chat with them.
That is intended. Standard replaces only the hard identifiers. For names to be replaced too, you need Strict, a matching setting under Custom or security mode. Talk to your admin.
No. "The managing director of our biggest joinery is ill" contains no name and still makes a person identifiable. That is what security mode and your own judgement are for.
Next steps
Building a good brain
One brain per audience, only current documents, clear instructions, the right access and real test questions, then answers and sources are right.
Rolling out AI
A 30-day plan for teams of 5 to 50 people, with a named owner, first tasks, an invitation link, shared prompts, training and measuring success.