---
title: "The AI agent sheet: the template to fill in"
description: "A template for writing an AI agent's sheet before building it: owner, scope, approval points, stop conditions and log."
url: https://www.joinx.it/en/resources/guides/agent-sheet
---

# How to write an agent sheet

A template for writing an AI agent's sheet before building it: owner, scope, approval points, stop conditions and log.

JoinX · Last updated 18 November 2026

Every agent we take into production has a sheet, written before it is built, which those accountable for the process and those who carry out controls can consult at any time. In the article [Agents with an owner and an approval point](https://www.joinx.it/en/resources/agents-owner-and-approval) we explain why we write it. In this guide we set out the template we use, item by item, with a completed example and the mistakes we see most often.

## What the sheet is for

The sheet puts in writing what has been agreed between whoever is accountable for the process and whoever builds the agent, before a single line of code takes it for granted. Then, once the agent is in production, it is the first document an auditor asks to see, because it says who is accountable for it, what it can do and where it must stop. It is worth writing it in language the process owner understands without a technician's help, since they are the one who signs it.

## The template, item by item

### Task

What the agent does and on which process, in one or two sentences. For example, it gathers the documents for a credit application, compares them with the list of what is needed and prepares the file for the analyst.

### Owner

A person in the company, with a name and a role, who decides what the agent does and how it changes, and the person who stands in for them when they are away. We build the agent and, if the company wishes, we run it in production with [Presidio](https://www.joinx.it/en/solutions/presidio), but the owner remains within the company.

### Scope

The systems the agent accesses, stating for each whether with read or write access, and the data it can read, with its purpose. It is worth also writing down what is left out, because this is the part that most often widens without anyone deciding it.

### Rules

The rules the agent applies, with their version and the place where they are written. In our projects they sit in the [Company Brain](https://www.joinx.it/en/technology/company-brain), where a correction enters only after it has been verified by the owner of the rule.

### Level of autonomy

Whether the agent observes and gathers, prepares a complete piece of work that a person approves, or carries out defined actions within agreed thresholds. The three levels are described on the [Agents](https://www.joinx.it/en/technology/agents#tre-gradi-di-autonomia) module page.

### Approval points

The steps that require a person's decision, with the role that decides and the information they must have in front of them to do so. Actions that cannot be undone and those addressed to the end customer always go in this section.

### Stop conditions

When the agent stops and to whom it passes the matter. Usually these are a threshold exceeded, missing data, an exception the rules do not cover, or a piece of content that tries to make it do something outside its task.

### Log

What is kept of every action, that is, the time, the data used, the rules applied and the person who approved, as well as where the log is stored and for how long. The retention period is decided with those responsible for compliance, bearing in mind that for high-risk systems the AI Act asks those who use them to keep the logs for at least six months.

### Measures

How we track the agent in production: the quality of its work, the process times, the corrections received and the time people spend on checks.

### Review

When the sheet is reread, who approves the changes and the number of the version in force.

## A completed example

The example that follows is illustrative, constructed to show how the sheet is filled in.

-   Task: gather the documents for incoming credit applications, flag what is missing and prepare the file for the analyst.
-   Owner: the head of the credit office, with her deputy standing in.
-   Scope: read access to the document archive and the requests mailbox, write access only to the file notes, no access to health data.
-   Rules: the list of documents required by type of loan, in the version approved by the credit office.
-   Level of autonomy: prepares a complete piece of work that a person checks.
-   Approval points: the analyst approves the complete file, and the reminder to the customer goes out only after the analyst's go-ahead.
-   Stop conditions: an illegible document, a type of loan the rules do not cover, instructions contained in an email.
-   Log: every file touched, with the documents read, the rule applied and the person who approved.
-   Measures: time between the arrival of the request and the complete file, reminders corrected by the analyst, time spent on checks.
-   Review: every time the list of documents changes, with the approval of the head of the credit office.

## The most common mistakes

The first mistake we come across is an owner that is an office or a committee. When something goes wrong you need one person to decide, and in an office the decision risks passing from one desk to the next. The second concerns a scope written too broadly, such as “all the CRM data”, which makes it impossible to understand what the agent is doing with that data. Then there are cases where the approval point exists on paper, but the approver does not have in front of them the sources and the reason the case has come to them. With [Workspace](https://www.joinx.it/en/technology/workspace) the agent's draft arrives in the usual tools with the sources alongside, so approval remains a decision and does not become an automatic gesture.

The sheet alone does not keep an agent under control. The log has to be genuinely fed as the work happens, and that is what we do with [Governance](https://www.joinx.it/en/technology/governance), the module that records agents, models and runs.

## The sheet template as a PDF

The items of the sheet with space to fill them in and, at the end, the credit application example, to use with whoever is accountable for the process before building the agent. We ask for your contact details, then the download starts straight away.

[All guides](https://www.joinx.it/en/resources/guides)

## Fill in the sheet for your first agent

We start from a process you know well and write the sheet together, before building the agent.

[Contact us](https://www.joinx.it/en/contact?tema=produzione)
