How to manage AI access#

Audience: admin — the master switch and per-surface grants are settings-owner decisions.

Marula treats AI as an optional tool that you control down to each role and screen. Out of the box, the organization-wide AI features master switch is OFF. When you turn it on, every individual AI helper (such as ticket note drafting or voice transcription) still starts disabled until you grant it to specific roles under Settings → AI.

Prerequisites#

• Admin access to Settings → AI.
• A quick read of Why AI is deny-by-default if you want the full background on how permissions, rate limits, and model credentials fit together.

1. Check the organization-wide master switch#

At the top of Settings → AI, the AI features master switch defaults to OFF:

"OFF keeps the exact pre-AI product — every AI surface answers quietly with zero ledger rows; flipping ON lets the per-surface grants and configs below decide what is actually audible."

When the master switch is OFF, no AI buttons appear in the workspace, no prompts are sent to any model, and zero AI ledger rows are written. Flipping this switch OFF at any time acts as an instant kill switch across the entire organization.

2. Grant specific AI helpers to roles or users#

Once the master switch is ON, use the Surfaces table to choose which tools are active and who can use them. Every surface defaults to denied until you add a role or user grant:

• Built-in writing and chat helpers: doc_drafting, email_drafting, general_chat, and ticket_note_assist.
• Plugin and workflow helpers: automation_writing, dashboard_generation, data_analysis, and transcription.

To turn a helper on for your service desk:

1. Find the surface row (for example, ticket_note_assist).
2. Add the role or user grant under Role/User grants.
3. Save. Users without a grant for that surface will never see its button in the UI.

3. Configure decision models (routing and triage)#

Lower on Settings → AI, the Decision harness section configures structured evaluation models used for ticket triage and routing. If no decision model is configured, those features stay cleanly off and report "decision model not configured" rather than failing with an error.

4. Set usage caps and model credentials#

• Usage caps (ai_caps): Set token or request limits per time window so a busy week never runs up an unexpected model bill. When a cap is reached, the surface pauses gracefully with a clear status message.
• Usage ledger: Every AI call records its surface and actor alongside the token count in the usage ledger.
• Model credentials: Store your API keys in the encrypted credential slot (at the organization level or per user) or point Marula at your own local hardware using How to run transcription and embeddings on your own hardware.

What you learned#

• Leaving the master switch OFF gives you a zero-AI workspace with no hidden background calls.
• Turning the master switch ON still requires granting each individual AI helper (ticket_note_assist, transcription, etc.) to the roles that should use it.