In short
- Your AI agent prepares the work. Your team reviews it and chooses: approve, correct or reject.
- A case that keeps coming back is reported instead of being corrected by hand every day. The agent will also report its own problems to our support, in preparation.
- Onboarding, included in every set-up, teaches these actions to your team and shows them how to track their progress.
An AI agent is a program that reads your emails, your invoices or your quote requests, then prepares an action following rules you have approved. How does an administrative team learn to work with it? As with a new colleague: it reviews what the agent prepares, takes over when it makes a mistake and says what is wrong. This guide covers the onboarding programme included in every Kendu Intelligence set-up.
The approval screen: what the agent suggests, and why
The agent's suggestions will wait for your decision on an approval screen, in preparation. Each one will show three things:
- what the agent read: the email received, the invoice, the bank transaction;
- what it suggests: a draft reminder, a reply, an accounting classification, a quote;
- what will happen if you approve: sending, recording, a transfer to a colleague.
Why this screen? Because the agent works by autonomy levels. A level sets what the agent is allowed to do on its own. At level 0, it suggests without acting: your team does the work and we compare. At level 1, nothing goes out without your agreement. The next levels only come on the basis of the measurements.
Some actions never go beyond level 1: a payment, sending a document that commits you to a third party, deleting data. For the Payment Reminders and Invoices agent, this will cover, for example, issuing an invoice, a formal notice or a credit note.
Each decision will be recorded in a log: what the agent read and suggested, what you decided, who, when, with which version of the agent.
Approve, correct or reject: three actions and their effect on the agent
Approve. The suggestion is right. You accept it as it is and the action goes out. For example, a customer asks for your opening hours and the draft reply is accurate.
Correct. The substance is good, a detail is not. You edit, then you approve. For example, the overdue payment reminder has the right amount, but the tone is too curt for a loyal customer.
Reject. The suggestion starts from a wrong basis: wrong customer, wrong request, or a case the agent should have handed over to you. Nothing goes out and you handle the case yourself. For example, the agent prepares a quote for a request outside your catalogue, when this is one of its stop cases. In that case, report it too.
To choose, ask yourself one question: “If a colleague had prepared this draft for me, would I start again from it?” Yes: correct. No: reject.
What these actions change for the agent. It does not learn on its own from one day to the next: each agent will have a fixed version. Your decisions will be counted in its indicators. When the same correction comes back, we will adjust the agent. The new version will first go through a short acceptance test, that is, a series of trials, before working for you. The log will show which version handled each case.
Your actions also weigh on its level. An agent only moves up a level on the basis of the measurements and with your agreement. After a serious alert, it is planned to drop back a level.
Reporting a problem: what goes to support, what stays with you
Delicate cases stay with you. Each agent will have a written list of stop cases: it will stop and hand over to you rather than guess. For the Emails and Support agent, these are for example a serious complaint, a legal threat or a person asking to access or erase their personal data. For the Payment Reminders and Invoices agent, it is notably a customer who disputes the invoice. These cases will go to the person you have designated: they are your cases, not breakdowns.
Operating problems go to support. A repeated error, software the agent can no longer read, suggestions that are often off target. As soon as you correct the same thing two or three times, report it.
Two routes will lead to the same support tool, in preparation:
- your team, from your future client area;
- the agent itself, which will open a ticket (a support request followed through to its resolution) when it meets a bug, a repeated error or a change in the AI model it uses.
The ticket will be attached to your project: we will be alerted and you will follow it in your client area. The same repeated problem will be grouped on a single ticket. When the agent runs in your own cloud account (servers rented from a host) or on your servers, it is planned that only its indicators and its alerts reach us. The content of the reports will be defined with you.
Early pitfalls: approving everything without reading, redoing everything by hand
Approving everything without reading. When suggestions are almost always right, attention drops. The EU regulation on artificial intelligence (Regulation (EU) 2024/1689, known as the “AI Act”) describes this risk for so-called “high-risk” systems. It speaks of “une éventuelle tendance à se fier automatiquement ou excessivement aux sorties produites par un système d'IA à haut risque (biais d'automatisation)” (a possible tendency to rely automatically or excessively on the outputs produced by a high-risk AI system (automation bias)) (Article 14, paragraph 4, point b). Your agents do not necessarily fall into this category, but an approval click commits your company. At the very least, read the recipient, the amount and the sentence that commits you. What lightens the work is moving to level 2 on safe cases, not a faster click.
Redoing everything by hand. Out of habit or distrust, some people rewrite every draft. The agent then saves no time and its measurements tell you nothing. If its suggestions do not deserve your trust, reject them and report it: this is useful information for us.
Correcting in silence. The same touch-up every morning, never reported: the agent keeps the same flaw.
Letting the queue grow. A suggestion that waits is a reminder that goes out late. Decide who approves what, and who covers during absences.
Tracking your progress: approval rate and time saved
Three measures will serve the whole team:
- the approval rate: the share of suggestions approved without changes;
- errors: the suggestions rejected because they were wrong;
- time saved: the time spent by the team, compared with the way of working before.
They will begin during the shadow trial: the trial period before going live, where the agent stays at level 0. Each agent will also have its own indicators: time to first reply for emails, average payment time for reminders, bank transactions without supporting documents for accounting documents.
You will find them in a dashboard, in preparation, and at regular check-ins whose frequency will be defined with you. The safe cases, which the agent will be able to handle alone at level 2, will be written with you in its rulebook.
The typical programme of the onboarding session
Onboarding is included in every agent set-up. We do not yet announce a typical length: we will set it once we have measured it. The programme follows the everyday actions:
- Your agent's journey. Its steps, those where it asks for your opinion, its stop cases.
- Reading a suggestion. What the agent read, what it suggests, what will go out if you approve.
- Approve, correct, reject. Exercises on examples close to your own cases.
- The cases it hands over to you. Who receives them, who covers in case of absence.
- Reporting. How to report a problem to us, and what the agent will report by itself.
- Reading the measures. Approval rate, time saved, the agent's indicators.
The goal: a team that approves its agent's work on its own.
The EU AI Act also requires companies that use an AI system under their own authority (“deployers”, defined in Article 3) to take measures “pour favoriser le développement de la maîtrise de l'IA par leur personnel” (to promote the development of AI literacy among their staff) (Article 4). In its version amended in 2026, it states that they do not have to guarantee a precise level for each person. Onboarding can be part of these measures. It does not replace legal advice.
Going further
- Training: onboarding and the training courses in preparation.
- Method: the autonomy levels and the shadow trial, step by step.
- Supervision: monitoring your agents, alerts and support, in preparation.
- Request an audit: find out which agents would suit your team. At €490 excl. VAT, it is deducted from the quote if you sign it within 30 days.