In short
- At Kendu Intelligence, an AI agent always starts “in shadow”: it proposes, your team does the work. It then gains autonomy level by level, action by action.
- It moves up a level only on the basis of measurements and with your written agreement. After a serious alert, defined in its rulebook, it is planned to step back down one level on its own.
- Paying, sending a document that binds you, deleting data: these actions never go ahead without a person's approval.
An AI agent is a program that reads your emails, your invoices or your CRM (the software where you track your customers), then prepares or carries out tasks for you. The right question is not “what can it do?”. It is “what do we let it do alone?”. Our answer: at the start, nothing without your agreement. Afterwards, only what it has proved it can do.
Why “fully automatic” is a bad starting point
Three reasons.
- The agent does not yet know your business. Your discount rules, the tone of your reminders, the customer you never chase before the 15th of the month: all of this is learned on your real cases, not in a demonstration.
- People quickly trust the machine too much. The EU AI Act (Regulation (EU) 2024/1689) names this trap “automation bias”: the tendency to automatically rely, or over-rely, on the output produced by the system (Article 14, paragraph 4). For the systems it classes as “high-risk”, it requires that the people assigned to oversight can monitor how they operate, disregard or override their output, and interrupt them “by means of a ‘stop’ button or a similar procedure”. These rules cover only “high-risk” systems. We are not saying your agents are among them: each one's role under this Regulation will be examined and written in its rulebook, agent by agent. But we start from the same principle for all our agents: at the start, nothing goes out until a person has approved it.
- A published account from the field points the same way. Jason Lemkin, founder of the SaaStr site, says he has deployed more than twenty AI agents in a year. On that site, he writes that his team spends 15 to 20 hours a week managing five sales prospecting agents, and that their results follow directly the human attention they are given. This is his own account, not an independent study. But it matches our position: an agent left alone too early risks drifting.
Hence our rule: the agent prepares, your team approves. In the cases set out in its rulebook, it stops and hands over. Its freedom is then earned, level by level.
The four levels, from shadow to supervised autonomy
The level is set per agent and per action. The same agent can send an acknowledgement of receipt on its own and wait for your green light on a quote.
| Level | The agent… | Your team… |
|---|---|---|
| 0. Shadow | proposes, but does not act | does the work itself; the two are compared |
| 1. Systematic approval | prepares each action | approves, corrects or refuses each action before it goes out |
| 2. Approval by exception | acts alone on safe cases, defined by written rules and thresholds | approves the cases outside the rules and checks a sample |
| 3. Autonomous and supervised | acts alone | reads the report and receives the alerts |
Level 0 serves as a trial period, which we call the “shadow trial” (acceptance testing is the test phase before going live). For a period defined with you, the agent proposes and your team carries on working. We measure how many proposals would have been right, what mistakes the agent would have made and the time it would have saved. You then decide whether to go live.
Take a typical scenario: an SME of 30 employees entrusts the sorting of its contact inbox to an agent. During the shadow trial, the agent classifies each message and writes a draft reply, alongside the team's own. Nothing goes out in its name. At go-live, it moves to level 1: each reply waits for approval.
Going live then proceeds step by step, from level 1 to level 2, according to the measurements from the shadow trial. At level 2, what counts as a “safe case” is written in black and white in the agent's rulebook. For an email agent, it might be an acknowledgement of receipt, opening hours or order tracking. Everything else goes up to a person.
Moving up a level: on the basis of measurements and with your written agreement
An agent never changes level because it “seems to work well”. Two conditions are needed, always together.
- Measurements. The main ones are the agreement rate (the share of proposals that your team approves without correcting them) and the errors found. They will appear in the supervision dashboard, which is in preparation, and in the follow-up reports.
- Your written agreement. You decide, not us and even less the agent. We propose the move, with figures to back it up; you sign, or not.
Each agent's levels will be reviewed regularly. A level can go up (always with your written agreement), stay where it is or go down.
Stepping back down automatically after a serious alert
Autonomy is lost faster than it is gained. It is planned that after a serious alert, defined in its rulebook, the agent steps back down one level on its own, without waiting for a meeting. An agent at level 2 will go back to level 1: each action will again wait for approval. To move back up, the rule is the same as at the start: measurements and your written agreement.
Agents are also planned to report their own problems. A bug, an error that repeats, a change in the AI model they use (the engine that reads and writes for them): the agent will open a ticket, that is, a tracking record for the problem, in our support tool, which is in preparation. This ticket will be linked to your project. We will be alerted and you will see the ticket in your client area. If the same problem comes back, it will be grouped under a single ticket.
Alongside these alerts, each agent has its stopping cases, written in its rulebook. It stops there and hands over rather than guess. For example:
- a serious complaint or a legal threat in an email;
- a customer in dispute, before a reminder or a quote; a customer with an unpaid invoice, before a quote;
- an unreadable accounting document, a suspicious duplicate, an unknown supplier.
The actions that always stay at level 1
Some actions never go beyond level 1, whatever the figures. A mistake there is costly and hard to put right.
- Payments. The Accounting Documents agent is planned to suggest reconciliations between your invoices and your payments. It will never pay.
- Sending a binding document to a third party. For the Quotes and Prospects agent: sending a quote, a commercial discount, a commitment on a lead time. For the Payment Reminders and Invoices agent: issuing an invoice, a formal notice, a cancellation or a credit note, a promise of a payment deadline.
- Deleting data.
Each agent also has its own rules. For example, any entry in your accounts will wait for approval, as will any document whose amount or VAT cannot be read with certainty. Each agent's rulebook lists these blocked actions. You approve it before the agent is built.
The log of each action: who approved what, when, with which version
Each action will leave a trace in a log:
- what the agent read;
- what it suggested;
- the decision taken: approved, corrected or refused;
- who took it, and when;
- the version of the agent.
The version matters. Each deployed agent will have a fixed version number, as will the tools and connections it uses. An update will go through a short round of acceptance testing again. When you look back at an action from March, you will know which version prepared it and, if it was approved, who gave the approval.
The EU AI Act goes the same way for “high-risk” systems. They must technically allow for the automatic recording of events (logs) over their lifetime (Article 12). The business that uses them (the Regulation calls it the “deployer”) keeps the logs under its control for at least six months, unless other EU or national rules say otherwise, in particular on personal data (Article 26). The same article requires human oversight to be assigned to people who have the necessary competence, training and authority, as well as the necessary support.
We follow the same idea: we plan a hands-on introduction in every set-up. It will be a short session where your team will learn to read the proposals, approve, correct and report.
On data, the agent will send the AI model only what is written in its rulebook. Your agents will run where you have chosen: your cloud in a European region, your servers, or a dedicated server with us, in the European Union. The AI model itself may be called from a provider: see data and security. Our supervision tool, in preparation, will receive only the indicators and the alerts; our support tool, the technical detail of the tickets.