# Governance

The **Governance** tab is where you make a Coworker safe to run in production. It holds guardrail selections, a human-in-the-loop (HITL) approval workflow, per-skill gating, and business-rule approvals.

<details class="watch">
<summary><span class="watch-thumb"><img src="https://i.ytimg.com/vi/1LkoWLb0md0/mqdefault.jpg" alt=""><span class="video-length">3:05</span></span><span class="watch-text"><span class="watch-label">Watch: Governance</span><span class="watch-meta">3 min · Episode 7 of 9</span></span></summary>
<div class="video"><iframe src="https://www.youtube-nocookie.com/embed/1LkoWLb0md0?rel=0&autoplay=1" title="Governance · Neo: Coworker in depth, Episode 7" allow="accelerometer; clipboard-write; encrypted-media; gyroscope; picture-in-picture" referrerpolicy="strict-origin-when-cross-origin" allowfullscreen loading="lazy"></iframe></div>
<p class="video-caption">Part of <a href="https://www.youtube.com/playlist?list=PLdwdGIspgPs0">Neo: Coworker in depth</a>. <a href="../videos/coworkers.html#ep-7">All episodes</a>.</p>
</details>

[![Governance tab — input and output guardrails and the HITL settings](../images/coworker-governance-v2.png)](../images/coworker-governance-v2.png)

For how approvals work in general — what pauses, what the reviewer sees, and what happens after a decision — see [Human in the Loop](../9_agentic_flow_builder/human_in_the_loop.html). This page covers the setup specific to Coworkers.

## Input and Output Guardrails

Two dropdowns let you apply an **input guardrail** and an **output guardrail**. The options are the guardrails that already exist in your workspace — you select from that list rather than defining a new policy here.

> **Important**
>
> Guardrails are **not inherited automatically**. Create a guardrail in the [Foundry](../12_foundry/guardrails/index.html), then select it here at the Coworker level. Once selected and published, it is enforced on the live Coworker. If the dropdown is empty, no guardrails have been created in the workspace yet.

## Human in the Loop (HITL)

HITL routes selected actions through a human approval step before they run.

1. Choose an **Approval delivery target** — a **Microsoft Teams** route. Approvals are delivered on Teams only, so you need an enabled Teams channel on the Coworker first; publish it at least once so that channel is available to select.
2. Turn on **Enable HITL approval workflow**.

Each approval request is sent to that Teams route, where an approver can **approve** or **reject** the action before it proceeds. The request arrives as an interactive Adaptive Card with **Approve** and **Reject** buttons that the system adds for you.

Approval requests are not raised when the Coworker runs in NeoGPT. Test approvals through Microsoft Teams.

> **Note**
>
> Only one approval delivery target is supported here. Advanced approval settings — approval title and description, and post-approval instructions — live on the orchestrating agent’s **Human In The Loop** node, which you can open with **View your orchestrating agent** after publishing.

## Skill Gate Policy

When HITL is enabled, the **Skill gate policy** section lists every active skill with a two-option toggle:

- **Allow** (green) — the skill runs without pausing.
- **Ask** (yellow) — the Coworker pauses and sends an approval request before the skill runs.

New skills default to **Allow**. The skill gate is the most dependable approval control: an **Ask** setting makes that skill pause for approval every time it runs. Use it for any action that must always be approved — for example a password reset or an access change.

## Business-Rule Approvals

When HITL is enabled, you can add business rules. Each has a **Rule name**, a **Condition** (a natural-language description of when to require approval), and an **Approver** (a recipient on a Teams channel). Rules can be toggled on or off.

> **Note**
>
> **How business-rule matching works.** The Coworker decides when a rule applies by interpreting your plain-language condition — it is not exact logic. Treat business rules as a soft control, and for actions that must *always* pause, rely on the per-skill **Ask** gate above.