> ## Documentation Index
> Fetch the complete documentation index at: https://docs.typewise.app/llms.txt
> Use this file to discover all available pages before exploring further.

# Manual actions

> Actions that Human Agents perform when the AI Agent requests

Manual actions let the AI Agent request a Human Agent to perform tasks that
can't be automated, either because no API exists or the task requires human
judgment.

**Best for:**

* Systems without API or integration support
* Sensitive tasks requiring human judgment
* Bridging gaps while implementing full integration

## How manual actions work

1. **AI Agent identifies need**: During a conversation, it determines a manual
   action is required
2. **AI Agent collects information**: Gathers required data from the customer
3. **Request sent to Human Agent**: Conversation hands off with a pending action
4. **Human Agent executes**: Performs the action in the external system
5. **Human Agent confirms**: Marks the action complete or rejected

## Configure a manual action

<img src="https://mintcdn.com/typewise/ht7a5v6EXYgQrv7K/images/documentation/actions/manual-actions/manual-action-form.webp?fit=max&auto=format&n=ht7a5v6EXYgQrv7K&q=85&s=760a01050e89a96c2a05880c4435a2ea" alt="Manual action configuration form" width="1974" height="2438" data-path="images/documentation/actions/manual-actions/manual-action-form.webp" />

| Field                    | Description                                                       |
| :----------------------- | :---------------------------------------------------------------- |
| **Action name**          | Descriptive name (e.g., "Refund Order", "Update Billing Address") |
| **Action description**   | When the AI Agent should suggest this action                      |
| **Required information** | Data the AI Agent must collect before requesting                  |

### Required information fields

Each field has:

* **Name**: Identifier (e.g., `order_id`, `refund_reason`)
* **Description**: What to ask the customer
* **Type**: Text, Number, or Yes/No

## Human Agent workflow

When the AI Agent requests a manual action, Human Agents see it in their inbox.

### View pending actions

The ticket shows:

* Action requested (e.g., "Refund Order")
* Collected information from the customer
* Conversation context

### Accept or reject

**To accept:**

<Steps>
  <Step title="Review information">Review the collected information.</Step>

  <Step title="Perform action">
    Perform the action in your external system.
  </Step>

  <Step title="Confirm">
    Click **Accept** to confirm. Optionally add a note describing what you did.
  </Step>
</Steps>

**To reject:**

<Steps>
  <Step title="Click Reject">Click **Reject**.</Step>

  <Step title="Provide reason">
    Provide a reason (helps the AI Agent respond appropriately). The
    conversation returns to AI Agent handling with your feedback.
  </Step>
</Steps>

## Example: Refund order

**Configuration:**

| Field       | Value                                                                     |
| :---------- | :------------------------------------------------------------------------ |
| Action name | Refund Order                                                              |
| Description | Use when a customer requests a refund. Collect order ID and reason first. |

**Required information:**

| Name            | Description                     | Type |
| :-------------- | :------------------------------ | :--- |
| `order_id`      | Order number to refund          | Text |
| `refund_reason` | Why the customer wants a refund | Text |
| `refund_amount` | Amount to refund (or "full")    | Text |

**Human Agent sees:**

```
Action Requested: Refund Order

Collected Information:
- Order ID: ORD-12345
- Refund Reason: Item arrived damaged
- Refund Amount: full

Customer Context:
"I received my order yesterday but the package was crushed
and the item inside is broken. I'd like a full refund please."
```

The Human Agent processes the refund, then clicks **Accept**.

## See also

* [AI actions](/documentation/actions/ai-actions): For automated execution