# Approvals and trust tiers

> How Unifuse decides who must approve an action the AI assistant proposes, what the approval cards show, and what the assistant is never allowed to do.

Source: https://docs.unifuse.ai/ai-assistant/approvals/

Every action the assistant takes through a tool is recorded in the audit log. The tool's
trust tier decides who has to say yes before it runs.

## The tiers

| Tier | Who decides | What you see |
|---|---|---|
| **Tier 1, autonomous** | Nobody; the call is logged. | The result. Read-only tools such as knowledge-base search, student lookups, summaries and metrics. |
| **Tier 2, requester confirms** | The person who asked. | A card in the chat: the tool it wants to run and its exact parameters, with **Confirm** and **Don't run**. The request expires after 24 hours. |
| **Tier 3, staff approval** | A holder of *approve agent actions* other than the requester (Advisor by default). | "Sent to staff reviewers for approval" with a link to the Approvals page. Expires after 72 hours. |
| **Tier 4, administrator approval** | A holder of *approve escalated agent actions* other than the requester (Admin by default); approvers are emailed. | "Sent to an administrator for approval". Expires after 72 hours. |

A tier 2 request from someone who does not hold the tool's own permission is raised to tier 3.
Sending a nudge to a student is tier 2 and cannot be set lower.

## What the assistant can never do

Some actions are not available to the assistant at any tier: deleting records, changing
permissions or trust tiers, the kill switch, exports, bulk or campaign sends, holds, and
anything to do with financial aid or billing. People do those.

## The Approvals page

**Approvals** appears in the sidebar for people who can approve. It has two sections:
**Needs your confirmation** (your own tier 2 requests) and **Review queue**. The page checks
for new requests every 30 seconds.

Each card shows the tool and its tier, where the request came from (web chat, a channel or an
automation), who requested it and when, when it expires, the decision basis, and **Exact
parameters that will execute**, verbatim. Add an optional **Decision note**, then **Approve**
or **Deny**. Both are recorded in the audit log with the note.

Administrators also see an **Oversight review** tab for rating past decisions on scope
accuracy, utility and veracity.

## The kill switch

An administrator can pause every agent at the institution from the Orchestration Hub's
global configuration. While it is on, the assistant, the web chat widget and the SMS line
stop responding.

## Related

- [The AI Front Door](/ai-assistant/front-door/)
- [Audit log](/administration/audit-log/)