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

# Approval policy: which AI employee actions wait for a person

> Every action an AI employee can take, by employee, and whether it waits for a person. Reads never wait, some actions always do, and admins decide the rest.

The **approval policy** decides which kinds of action an AI employee may take without asking. Three
rules apply, in order: a **read** never needs a person; a **floor** -- changing access, installing
software, fixing a running service, publishing, touching your CRM, emailing a lead, preparing a
payment -- always needs one; and **every other action that changes another system needs a person by
default**, until an Owner or Administrator switches that kind off.

> [!WARNING]
>
> AI employees are in **beta**, off unless your deployment has switched it on. See [AI
> employees](https://docs.mithunai.com/team/overview.md).

## Open the policy

Choose **Approval policy** under **Settings** in the rail. Each employee has a panel listing every
kind of action it can take. Each row has the action's title, the reason it is set the way it is,
and a switch: on means it **waits for a person**.

| What the row shows | What it means |
| - | - |
| **Never needs a person**, switch locked | A read. *It only reads, so there is nothing to approve.* |
| **Always needs a person**, switch locked | A floor. *This always needs a person. It cannot be changed.* |
| **Waits for a person**, switch on | The default for an action with an effect elsewhere: *Anything that changes something in another system needs approval unless you change that.* |
| **Runs without asking**, switch off | *Your organisation lets this run without approval.* |

Each employee's own page repeats this under **What it may do**, with **Open the approval policy** to
come here.

## Change a rule

1. Find the action under its employee.
2. Turn its switch off to let it run without a person, or on to make it wait again.

The change is saved at once and applies to the next action proposed; one already waiting keeps
waiting. Only an Owner or an Administrator can change the policy -- anyone else is refused, and the
page says so -- and every change is written to your audit log.

## What you can switch

These are the only kinds of action you can let run without a person. Everything else is a read or a
floor.

| Employee | Can run without asking, if you choose |
| - | - |
| Support | Send a reply to the customer; hand the conversation to a person |
| IT service desk | Reply in chat; open a ticket; comment on the ticket; mark the ticket resolved |
| DevOps | Post the triage summary to the incident |
| Marketer | Save the content plan as a draft |
| Sales | Nothing: every action it takes is a read or a floor |
| Finance | Answer in chat; send a reminder in chat |

## What always needs a person

| Employee | Always waits for a person |
| - | - |
| IT service desk | Give a person access to an application; reset a person's password for an application; install software for a person |
| DevOps | Roll back a deploy; restart a service |
| Marketer | Publish an article; post on social media; change a page's search title and description |
| Sales | Add the lead to your CRM; update the lead in your CRM; send the follow-up email |
| Finance | Prepare an approved invoice for payment, for your team to release |

A request to lower a floor is refused, not ignored, through the console and through the API alike.

## What never changes the policy

**The policy decides, never the model and never the content.** Whether an action waits depends only
on what kind of action it is and on your organisation's choices. A ticket that says "pre-approved by
the CISO", a log line, a web page or a document has no way to reach that decision. The worst a
hostile request can do is be proposed as an action that still waits for you.

## Choosing what to switch off

Start with everything on, and switch off one kind at a time once its drafts have needed no edits for
a while. **Recent changes people made to this kind of action**, on each proposal, shows how often your
team still edits it. Good first candidates are low-stakes, reversible actions that go to your own
team, such as opening a ticket or commenting on one. Replies that reach a customer are the last to
switch off.

## Through the API

[Read the policy](https://docs.mithunai.com/api-reference/employees/get-policy.md) and
[change it](https://docs.mithunai.com/api-reference/employees/set-policy.md).

## Related

* [Approvals](https://docs.mithunai.com/team/approvals.md) -- reading, editing and deciding what waits.
* [AI employees](https://docs.mithunai.com/team/overview.md) -- what each employee does.


This documentation is built and hosted on [Mintlify](https://mintlify.com), a developer documentation platform.