Privacy and retention
Which content Holistic AI EndLayer collects by default, the four checks in front of conversation content, how long data is kept, and who can see it.
By default, Holistic AI EndLayer, the Guardian Agent Platform, collects metadata: what is installed, how it is configured, which models are used and how many tokens. A few content settings go further. They are what decides whether rolling out EndLayer is a configuration change or a conversation with your works council, so this page sets them out plainly.
Content settings
Each setting is a row in the console's policy library. The defaults below are a decision your organisation can change, and every change is recorded.
| Setting | What it adds | Default |
|---|---|---|
| Session messages | The prompts people typed and the model's replies | Off |
| Session titles | The model's one-line description of each conversation | Off |
| Skill contents | The full text of skills and commands | Off |
| Skill authors | The author names a plugin declares | Off |
| Repository files | The file paths in repositories agents work in, never their contents | On |
| DLP evidence | The prompt a detection was found in, with every match masked on the device | On |
Switching a setting off tells devices to stop sending it, and EndLayer strips it from anything a device sends before it has heard.
Four checks in front of conversation content
- The device must be collecting it
Conversation content is off by default in the agent's own configuration. The console can switch it off for the fleet, but never on, so no single fleet-wide setting can start reading conversations.
- The organisation must have enabled it
With a stated lawful basis, recorded against the person who gave it. While it is off, the content is removed before anything is stored, which also keeps it out of backups and exports.
- The reader must be an administrator
Members cannot read conversation content.
- Every read is logged first
An access record naming the reader, the session and a required reason is written before anything is shown.
Conversation content is kept only with the raw report it arrived in, so it becomes unreachable when that report's retention window closes.
How long data is kept
| Data | Default |
|---|---|
| Raw device reports | 14 days |
| Report metadata | 30 days |
| History: control changes, detections, alerts and token usage | 400 days |
Administrators can change each window within set limits. Data past its window is deleted every day, so retention is something that happens rather than a policy on paper. The audit trail and policy history are kept separately on purpose, so the record of decisions outlives the data they were about.
Erasure is available per device or for the whole organisation, from the console. Each erasure needs a stated reason, which becomes its audit record. Data copied in from a connection follows that connection: disconnecting deletes it.
Who can see what
Three roles, each holding everything the one before it does.
| Role | What it adds |
|---|---|
| Member | The fleet: posture, coverage, overall token cost, evidence and evidence packs |
| Admin | Details about people: identity, cost per person, session titles and conversation content, plus settings, enrolment, devices, roles and the audit log |
| Owner | Billing, and deleting the organisation |
A member sees the fleet; an administrator sees who. Evidence packs are built from a view with no conversation content, titles, serial numbers or per-device browsing in it, which is why any member can generate one and hand it to an auditor.
For your employee notice
When you tell employees about EndLayer, these points usually matter most:
- What is collected: AI tools, their settings and usage counts, AI website visits by domain, and device security posture.
- About people: login, display name and email address, visible only to administrators.
- Content: conversation content is off by default and needs a lawful basis, an administrator and a logged reason to read.
- On the device: secret values and matched sensitive text are discarded or masked before anything leaves.
- Visible to the person: every warning and block comes with a reason, and where the menu-bar or tray app runs it lists the recent ones.
For where data is hosted and Holistic AI's data processing terms, see the Trust page.
Updated 2026-11-20