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

# Procedures: how recurring work is done

> Write down how your business handles a refund, a new lead or a delivery, so every AI agent does it the same way — and still asks you where it matters.

**A procedure is a short guide for one recurring thing: what usually happens, where the AI agent uses judgment, what always
asks you, and what never happens.** It is a guide, not a script: it shapes the work, while your rules and approvals still
apply.

## Find it

Open **AI workforce** and choose the **Procedures** tab (the page's own sub-line is "How your AI agents handle recurring
work"). Each procedure shows how many AI agents follow it and a line such as "4 steps · 2 always ask · 1 never".

Empty: **"No procedures yet"** — "Write down how you handle a refund or a new lead. Your AI agents follow it and still ask
where it matters."

## Write one

**New procedure** asks for:

| Field | What goes there |
| - | - |
| **Title** | The thing, like "Refund a customer" |
| **Goal** | What a good result looks like |
| **Usually** | What normally happens |
| **Use judgment when** | Where the AI agent decides |
| **Always ask me** | Anything here waits for your yes |
| **Never** | Never done, whatever the setting |
| **Steps** (optional) | In order, each marked **AI agent** or **A person** |
| **AI agents who follow it** | Who reads it whole; each starts at **Asks first** |

The four lists are the heart of it. "Usually: Reply the same day." "Use judgment when: The customer sounds upset." "Always
ask me: A refund over \$50." "Never: Promise a delivery date."

**Save** makes it active. **Archive** takes it out of the list, after asking; you can write it again any time.

Limits: title 80 characters, goal 240, each line 240, 12 lines per list, 20 steps, 20 AI agents, 30 procedures per project.

## How AI agents use it

* Every AI agent in the project sees the active procedures as a title and a goal, and reads one whole when the Task fits.
* The AI agents named on it always follow it whole.
* Where it says **Use judgment when**, the AI agent logs the call for you to rate: **Good** or **Not quite**.
* **Never** lines are hard stops, and money, refunds, deleting, card spending and your Never rules always ask first anyway.

## Suggestions from AI agents

**Suggested by your AI agents** lists procedures they noticed while working, with why. **Keep** makes one active; **Not now**
sets it aside (the same title is not suggested again for a while). AI agents never write an active procedure themselves.

## Who can change one

Anyone who can open the project can write, edit or archive a procedure. Only an owner or admin can set an AI agent to
**Sends and tells you** or **Auto approve** on it; everyone else keeps the level at **Asks first**.

## If something is wrong

| Symptom | Likely cause | Fix |
| - | - | - |
| **Procedures aren't on yet** | The server is older | Check back; AI agents work as before |
| **This is the most a project can hold** | 30 active procedures | Archive one |
| An AI agent ignores it | It is archived, or not relevant to the Task | Check the list; keep it active |
| The level buttons are locked | You are not an owner or admin | Ask one to change it |
| A suggestion keeps coming back | It was kept and then archived | Write it yourself, or dismiss again |

## Related

* [Company knowledge](https://docs.workser.ai/ai-workforce/company-knowledge.md): the facts and rules behind every Task
* [Autonomy and probation](https://docs.workser.ai/using/autonomy-and-probation.md): how much an AI agent does without asking
* [Front desk](https://docs.workser.ai/using/front-desk.md): the team behind one line
* [The AI agent panel](https://docs.workser.ai/using/agent-panel.md): one AI agent's own settings


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