Appearance
AI Protection
The AI Protection panel allows organization administrators to manage your organization's data protection policy for PII detection. You can monitor private model health, set org-wide detection defaults, and view protection metrics across your team.
Overview
The AI Protection tab provides:
- Protection Status — Health and configuration of your private (in-environment) detection model
- Detection Policy — Org-wide defaults for detection model and verification mode, with optional locks
- Protection Metrics — Sessions protected, entities detected, and model usage over time
- User Detection Overview — Per-user compliance with the organization's policy
Accessing AI Protection
- Open Admin Dashboard from the sidebar
- Select AI Protection in the sidebar
Protection Status
The Protection Status section shows the health and configuration of your private detection model.
Private Model Health
- Operational (green) — The private model is available and ready for use
- Unavailable (red) — The private model is not configured or cannot be reached
When available, the card displays:
| Field | Description |
|---|---|
| Model | The model identifier (e.g., phi-3-mini) |
| Runtime | Execution environment (e.g., ollama, vllm) |
| Deploy Mode | How the model is deployed |
| Status | Loaded (ready) or Loading |
Inference Target
The Inference Target card shows where detection inference runs:
- Local — Model runs in your environment (data never leaves your network)
- Remote Service — Model runs on a remote endpoint
Additional details include:
- Active Model — Model currently in use
- Endpoint — URL of the inference service
- GPU — GPU device (if applicable)
- GPU Memory — Allocated GPU memory (if applicable)
If the probe reports an error, it is shown in the card.
Detection Policy
The Detection Policy section lets you set org-wide defaults for all users. These defaults apply unless a user has overridden them (when locks are off) or when the user has no preference.
Detection Model
Choose the default detection model for the organization:
| Option | Description |
|---|---|
| Private Model | In-environment LLM — data never leaves your network |
| Cloud API | Uses the configured cloud LLM provider |
| Custom Endpoint | OpenAI-compatible self-hosted endpoint |
Lock for all users — When enabled, users cannot override the org default. The detection model is fixed for everyone in the organization.
Verification Mode
Choose how detections are verified:
| Option | Description |
|---|---|
| Verify All | Every detection is verified by a second pass |
| Suspicious Only | Verify only low-confidence detections |
| Off | Skip verification (faster, lower accuracy) |
Lock for all users — When enabled, users cannot change the verification mode.
When any lock is enabled, a notice appears: Locked settings prevent users from overriding the org default. Users will see a notice explaining the policy.
PII Detection Model
PII detection uses Dymium NER (dymium/Dymium-NER-v1), a DeBERTa-v3-large model fine-tuned for 13 PII entity types. The model is pre-downloaded during Docker image build and loaded automatically at runtime. Contact your administrator for custom model configuration.
Always-On PII Protection
When the force_pii_protection org policy is enabled, PII detection is mandatory for every message sent by every user in the organization. Users cannot disable or bypass detection.
| Setting | Description |
|---|---|
| Status banner | A prominent indicator at the top of the AI Protection tab shows whether always-on protection is active (green) or inactive (gray) |
| Toggle | Admins click the toggle to enable or disable the policy instantly |
| User impact | End users see an indicator that always-on protection is enforced; individual detection toggles are overridden |
This policy is stored as force_pii_protection in organization policies and takes effect immediately.
Protection Metrics
The Protection Metrics section provides aggregate statistics for the selected time period.
Time Period
Use the 7d, 30d, or 90d buttons to change the reporting window.
Top-Line Stats
| Metric | Description |
|---|---|
| Sessions Protected | Total chat sessions that used AI-powered detection |
| Entities Detected | Total PII entities detected across all sessions |
| Active Users | Number of users with detection activity in the period |
Detection Model Usage
A horizontal bar chart shows the split of detection traffic by model:
- Private — Sessions using the in-environment model
- Cloud — Sessions using the cloud API
- Custom — Sessions using a custom endpoint
Each bar shows count and percentage.
Entities by Type
A breakdown of detected entities by type (e.g., PERSON, EMAIL_ADDRESS, PHONE_NUMBER). Up to 10 entity types are shown with relative counts.
User Detection Overview
The User Detection Overview table lists each user who had detection activity in the selected period.
| Column | Description |
|---|---|
| User | Name and email |
| Detection Model | Model the user is using (private, cloud, or custom) |
| Entities Detected | Count of entities detected for that user |
| Policy | Compliant — matches org default; Override — user has chosen a different model |
Use this table to identify users who have overridden the org default and assess compliance with your data protection policy.
FAQ
How do locks work?
When you enable Lock for all users on Detection Model or Verification Mode:
- The org default becomes mandatory for all users
- User preferences for that setting are ignored
- The setting appears in the user's Settings → PII Detection panel with a notice: Your organization has locked the detection model. Contact your admin to change this setting.
- The control is disabled; users cannot change it
Locks apply immediately. Users see the locked state on their next settings load.
What do users see when a setting is locked?
In Settings → PII Detection, locked fields display:
- A notice banner: Org policy: Your organization has locked the [detection model / verification model]. Contact your admin to change this setting.
- The control grayed out or disabled
- The org-mandated value shown
Users can still view the setting but cannot modify it.
How does the 5-tier config resolver prioritize settings?
GhostAI uses a 5-tier configuration hierarchy for detection settings. Lower tiers override higher tiers, subject to permission constraints (such as org locks):
| Tier | Scope | Description |
|---|---|---|
| 1. Cloud | System | Platform defaults (e.g., from environment) |
| 2. Tenant | Organization | Org-level settings from Admin (AI Protection, Team Settings, etc.) |
| 3. User | Individual | User preferences from Settings |
| 4. Conversation | Session | Per-chat overrides |
| 5. Subject | Topic | Adjustments for detected subject (e.g., medical, legal) |
Example: If the org sets Detection Model to Private and locks it, the User tier cannot override it. The resolved value is Private for all users. If the org does not lock, a user who prefers Cloud can override; their preference (User tier) wins over the org default (Tenant tier).
The AI Protection tab controls the Tenant tier for default_ai_model, default_verify_mode, lock_ai_model, and lock_verify_mode. Locks prevent the User and Conversation tiers from overriding those settings.