Sensitive data and AI privacy
Control what information AI may receive without obscuring the normal human view of your files.
Sensitive-data controls are a protection layer between AI and your data. They complement permissions; they do not replace workspace access, source permissions or careful choice of provider.
Human view versus AI access
| Situation | What the control means |
|---|---|
| You browse a file, author name or preview in CogLake | Your normal access rights apply; the AI privacy layer is not intended to replace your visible data with redacted text |
| An AI reads knowledge through tools or MCP | Applicable sensitive-data rules protect the information made available to that AI |
| Content is prepared for an AI request | The applicable protection rules can redact or block sensitive information |
| An AI requests an original file | Original-file restrictions can prevent passing an unprotected original to the AI |
| You lack access to the source or workspace | Neither a human view nor AI can use the privacy setting to bypass that denial |
- 1Select permitted information
- 2Apply the AI privacy rules
- 3Use the allowed model or tool
- 4Review the result
A placeholder means content was withheld for that AI interaction. It does not mean the stored source document was edited. If ordinary file browsing unexpectedly displays a protection placeholder, report the affected view rather than changing the source file.
Configure role policies
Open Administration → Workspace → Sensitive data. Choose a role, find the relevant data classes and select the desired action.
- Allow: permitted values may be used in the AI interaction.
- Redact: detected values are replaced with protection placeholders.
- Block: content that violates the rule is withheld; the task can be stopped or require different input.
When several roles apply, the stricter matching sensitive-data action wins. Being allowed to read a document as a person does not imply permission to send every part of it to an AI.
Use the policy preview with representative, non-sensitive sample text. Review false positives and missed matches before relying on a policy. Custom classes let you add terms such as internal project identifiers. Detection is a safeguard, not a guarantee that every possible secret is recognized.
Initial privacy setup
The initial privacy choice provides starting defaults. The privacy-focused option enables protective defaults and more restrictive personal resource controls; the open option starts more permissively. Administrators can later review and change the individual settings.
Completing a new setup is not a reason to overwrite unrelated existing users' settings. Check the actual role and organization policy when troubleshooting.
Models and document processing are separate
Your organization controls which CogLake-provided models, BYOK connections and personal model connections members may use.
Installation-level document processing, such as OCR and transcription, is configured by the service operator. There is no tenant setting called AI providers & data egress for moving those processing decisions to a company administrator. Contact the operator if a required processing capability is unavailable.
See models and BYOK and supported files.