---
title: "AI agents with an owner and an approval"
description: "Mandate, owner, approval point and log: how we take an agent into production while leaving decisions with those accountable for them."
url: https://www.joinx.it/en/resources/agents-owner-and-approval
---

# Agents with an owner and an approval point

Mandate, owner, approval point and log: how we take an agent into production while leaving decisions with those accountable for them.

JoinX · Last updated 18 November 2026

At an insurance company, a claim can sit idle because documents are missing, more than because of the settlement decision. Someone has to read the notification and the loss adjuster's report, compare them with the policy wording of the right edition, chase what is missing and put the file in order for the claims handler. It is the kind of work we are happy to hand to an agent, while the settlement decision stays with the claims handler.

Put like that, it sounds simple. In production, though, an agent works on many files, in the systems the insurer already uses and with real data, so before putting it to work you need to know who is accountable for it, what it can do and where it must stop. These are the questions every agent we build with the [Agents](https://www.joinx.it/en/technology/agents) module starts from.

## The mandate, written before building

Before it goes into production every agent has a sheet, which those accountable for the process and those who carry out controls can consult at any time. The sheet says what the agent must do and on which process, which person in the company is accountable for it, which systems it accesses and whether with read or write access, which rules it applies and in which version. It also sets out the steps that require a person's decision, with the name of the role that decides, the conditions under which the agent stops and the measures we use to track it.

The owner is a person in the company, with a name and a role. We build the agent and, if the company wishes, we run it in production with [Presidio](https://www.joinx.it/en/solutions/presidio), but what the agent does and how it changes is decided by whoever is accountable for the process.

## How much the agent does and how much the person decides

We distinguish three levels of autonomy. At the first, the agent observes, gathering and ordering the information needed to decide. At the second, it prepares a complete piece of work, such as a draft reply or a proposed classification, which a person checks, corrects and approves. At the third, it carries out defined actions within agreed thresholds, and the person approves the critical steps and sample-checks the rest.

We usually start at the first or second level, and move up only when the measures justify it. Whatever the level, an action that cannot be undone or that reaches the end customer always goes through a person, as we describe in [Deployment](https://www.joinx.it/en/solutions/deployment#gradi-di-autonomia). We describe this in detail on the [Agents](https://www.joinx.it/en/technology/agents#tre-gradi-di-autonomia) module page.

## Where the approval arrives

An approval point works if the approver has in front of them what they need to decide. With [Workspace](https://www.joinx.it/en/technology/workspace) the agent's draft arrives in email, in the company chat or on a process page, with the sources alongside and the reason the case has come to that particular person, so they can approve or correct it without rebuilding the case from scratch.

The risk to watch is approval that becomes an automatic gesture. When the proposals are almost always right, attention drops, and the rare error may slip through. That is why we also measure the time people spend on checks, keep up sample checks when results are good and treat corrections as valuable information. When a person corrects the agent, the correction enters the [Company Brain](https://www.joinx.it/en/technology/company-brain) only after it has been verified by the owner of the rule, and from then on it applies to every agent that uses it.

## When the agent stops

An agent stops when a threshold is exceeded, when data is missing or when it meets an exception the rules do not cover, and in all these cases the matter passes to the person named in the sheet. It must also stop when what it reads tries to make it do something else, because an email can contain instructions written precisely to push it into forwarding a document. With [Security](https://www.joinx.it/en/technology/security) every agent has an identity and only the permissions its task requires, and any action suggested by a piece of content is checked against its mandate before it is carried out.

## The evidence of what happened

Faced with an agent in production, an auditor asks who is accountable for it, which rule it applied and who approved the decision. With [Governance](https://www.joinx.it/en/technology/governance) we record agents, models and runs as the work happens, and for every action there remains the time, the data used, the rules applied and the person who approved it. The log belongs to the company, like the rules and the documentation, and remains its own even when it decides to run the agents in-house or with another supplier.

[All articles](https://www.joinx.it/en/resources#articoli)

## Write the mandate for your first agent

We start from a process you know well and look together at where the approval points are needed.

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