The workspace AI policy is the control center for every AI feature in Doxbrix. From one place, admins decide which models are enabled, how much AI can spend, whi...
The workspace AI policy is the control center for every AI feature in Doxbrix. From one place, admins decide which models are enabled, how much AI can spend, whi...
The workspace AI policy is the control center for every AI feature in Doxbrix. From one place, admins decide which models are enabled, how much AI can spend, which features are on, how autonomous the agent may be, what style AI must follow, and where your data is allowed to go.
/settings/ai-policy). Only workspace Owners and Admins can edit it; changes apply across the workspace.The form is organized into six sections.
Models & Tiers
Choose which models and response tiers AI features use:
| Setting | Controls |
|---|---|
| **Enabled models** | Which underlying models AI features may use. |
| **Default tier** | **Spark** (fast responses, lower cost) or **Atlas** (deeper analysis, higher quality). |
| **Copilot tier override** | Pin a tier for the [Writing Copilot](/ai/copilot). |
| **Scans tier override** | Pin a tier for [quality scans](/ai/scan-and-content-quality). |
| **Agent tier override** | Pin a tier for the [AI agent](/ai/agent). |
Each override can also be set to Use workspace default. Restricting to a single approved model is common in regulated environments.
Token Budgets
Keep AI spend predictable:
- Monthly workspace budget (tokens) — a ceiling on total AI consumption per period (or No limit).
- Alert threshold (%) — notify when usage crosses this share of the budget (e.g. 80%).
- Alert recipients — the email addresses that receive budget alerts.
Track real consumption in Usage.
Feature Controls
Turn each AI surface on or off independently:
- Workspace assistant — the chat assistant at
/assistant. - Writing Copilot — inline AI writing and edits in the editor.
- Improve with AI — the page-header rewrite button.
- Scan with AI — automatic quality scoring.
- Site assistant — the reader-facing Ask AI on published sites.
- AI agent — scheduled, automated tasks.
Agent Autonomy
Set hard ceilings on how autonomous the agent may be:
| Setting | Controls |
|---|---|
| **Maximum autonomy** | **Suggest only** (AI proposes, humans decide), **Draft for review** (AI drafts, humans approve), or **Autonomous** (AI acts without review). |
| **Workspace-level approval policy** | The default approval requirement AI changes must satisfy. |
| **Default AI prompt context** | Background context automatically provided to AI runs. |
| **Escalation rule** | When AI should hand off to a human. |
The Maximum autonomy ceiling applies even if an individual task is configured more aggressively — the policy always wins.
Style & Terminology
Codify your editorial standards so all AI-generated and AI-improved content stays consistent:
| Setting | Example |
|---|---|
| **Voice** | *Professional but approachable* |
| **Grammatical person** | Second person ("you") |
| **Reading level** | *Grade 9–11* |
| **Preferred terms** | "sign in" (not "log in") |
| **Banned terms** | "simply", "just", "obviously" |
The Writing Copilot and Improve with AI follow these rules automatically, keeping AI-assisted writing aligned with your hand-written content.
Data Boundaries & Privacy
Set the guarantees your organization requires:
| Control | Effect |
|---|---|
| **No training on your content** | Your content is never used to train foundation models. |
| **Zero data retention** | Use model endpoints that retain no request data. *(Enterprise)* |
| **Data residency** | Pin AI processing to a region — **Global**, **United States only**, **European Union only**, or **Asia Pacific only**. *(Enterprise)* |
A recommended baseline
A safe starting policy for most teams:
Turn on the Writing Copilot, Improve with AI, Scan with AI, and the Workspace assistant.
Let the agent draft and submit for review, but not act autonomously.
Pick a token ceiling, set the alert threshold to 80%, and add alert recipients.
Add a voice, reading level, and preferred and banned terms.