Which process should you start with? Choosing the first job to hand to an AI agent

In short

• A good first process is repetitive, follows rules you can write down, tolerates mistakes that can be put right, relies on accessible data and offers a gain you can measure.
• Rare, sensitive or “gut-feel” decisions can wait: they are not the first thing to hand to an AI agent.
• The Kendu Intelligence audit fits in a one-page report: tools and versions, volumes, personal data, expected gains. It is paid, and deducted from the quote if you sign it within 30 days.

You want to try an AI agent. The real question is: on what? An AI agent is a program that reads your emails, invoices or files and prepares an action: a draft, a filing, a payment reminder. To begin with, it submits every action to your team. Well chosen, the first agent builds confidence for what comes next. Here is how we choose it with you.

Five criteria for a good first process

1. A repetitive volume. The same task comes back every day or every week: chasing overdue invoices, sorting a shared inbox, filing accounting documents. The more cases there are, the sooner you know whether the agent does well.
2. Rules you can write down. If you can explain the task to a new employee on one page, that is a good sign: an agent can probably follow it. These rules become its rulebook, with its “stop cases”: the situations where it stops and hands over to you rather than guess.
3. Mistakes that can be put right. Start where a mistake costs a correction, not a client. A badly worded draft reminder can be corrected before it goes out. A formal notice sent by mistake cannot.
4. Accessible data. The information is in software: email, invoicing, online banking. This software often has an application programming interface, or “API”: an entry point provided by the vendor so that another program can read or write data. If the information exists only on paper that was never scanned, or in one person's head, the agent cannot read it.
5. A measurable gain. You know how long the task takes today, or you already track an indicator: average payment time, time to first reply to a client. Without a starting point, there is no way to tell whether the agent helps.

The four standard agents in our catalogue target jobs of this kind: payment reminders and invoices, emails and support, quotes and prospects, accounting documents.

Poor candidates: rare, sensitive decisions or work with no written rule

Some jobs can wait, even if they take time.

• Rare decisions. Choosing a strategic supplier, reviewing your prices once a year: too few cases to write rules and measure.
• Sensitive situations. A dispute, a serious complaint, health data, a person asking to exercise their rights over their data. Our agent pages classify these situations as stop cases: they provide that an agent that spots one flags it to a human and does not handle it.
• “Gut-feel” work. When the answer is “it depends, I can feel it”, the rule is not written yet. We can write it together, but it is not the best first project.

Our method provides that some actions are always approved by a person, whatever the agent: a payment, sending a binding document to a third party, deleting data. The agent prepares them, your team decides.

Take an inventory: tools, versions, volumes, personal data

Before choosing, we look at what exists.

• Which tools, and in which version? The version matters, because vendors change their interfaces. For example, the documentation for Odoo 19, a business management software, says that its older XML-RPC and JSON-RPC interfaces are scheduled for removal in Odoo 22, expected in autumn 2028. An agent connected to those older entry points would stop working when you move to Odoo 22. With us, each connector (the link between the agent and a piece of software) will freeze the interface version it uses, and every update will go through a new short trial.
• Which volumes? How many invoices, emails or quotes per month, and how many minutes for each.
• Which personal data? Names, addresses, message contents, employees' expense reports. The General Data Protection Regulation (GDPR), the European text that governs the use of personal data, provides for a record of processing activities: the written list of what the company does with this data. Companies with fewer than 250 employees are exempt, except in certain cases, including processing that is “non occasionnels” (not occasional) (Article 30, paragraph 5). The CNIL, the French data protection authority, gives payroll management and the management of clients, prospects and suppliers as examples. A payment reminder or support agent touches precisely on client management. In case of doubt, the CNIL recommends entering the processing in the record: this is the most prudent course, to be checked with your adviser. And if a provider processes this data for you (the GDPR calls it a “processor”), a contract must set out the subject matter, duration, nature and purpose of the processing, the type of data and the categories of people concerned (Article 28, paragraph 3).

The audit: what it contains, and why it is deducted from the quote

At Kendu Intelligence, the audit is the first step of every set-up. It is short, paid work, by quote. We look at your tools, your ways of working, your volumes and your personal data. From it we produce a one-page report, the catalogue agents we select for you, and a quote, with the estimated number of days.

If you sign, the price of the audit is deducted from the quote. You do not pay twice for the same work: the audit becomes the first part of the set-up.

Why charge for it? Several agencies offer a free first audit or diagnostic, of 20 to 45 minutes according to their pages viewed in October 2026. That is useful to get acquainted. But choosing the right first process means opening your tools and counting your volumes. This work has value in itself: it ends in a clear page on what is worth handing to an agent. The audit prices found on the market are detailed in our article on prices.

In France, the State also offers a longer scheme: the Diagnostic Data IA of the national plan Osez l'IA, delivered for the State by Bpifrance Conseil. According to the official site entreprises.gouv.fr (page updated on 28 September 2026), it is a service of 8 person-days, that is 8 working days in total. It costs €10,000 excl. VAT, of which 40% is covered by the France 2030 investment plan. It is aimed at SMEs and mid-caps (intermediate-sized companies) with 10 to 2,000 employees, registered in France in the trade and companies register. It delivers a review of the current situation, concrete use cases for AI and their ranking by priority. If you have already had one done, its priorities are a good starting point for choosing the first process. Conditions may change: check them on the official page before you start.

Estimate the gain before you start, then measure it

Before starting, a simple calculation: the monthly volume multiplied by the time spent per case. With illustrative figures, 200 payment reminders a month at 6 minutes each come to 20 hours of work. That is the time at stake. It is not yet a gain.

The real gain is measured during the “shadow trial” (level 0). Acceptance testing is the trial period before going live. It lasts a few weeks: the agent suggests, your team does the work as usual, and we compare. Suggestions accepted as they are, errors, time saved: these measurements decide what happens next. The agent moves up an autonomy level only on these figures and, as our method provides, with your written agreement. At level 1, your team approves every action before it goes out. At level 2, the agent acts alone on cases covered by written rules and thresholds; your team approves the others and checks a sample.

After that, supervision follows the same indicators. Faced with a stop case, the agent stops and hands over. After a serious alert, it can drop back a level.

Example of a one-page audit report

Typical scenario, fictional: an SME with 30 employees that sells equipment to professionals and invoices on 30-day terms.

Going further

• From the audit to supervision, step by step: our method.
• The four standard agents: our agents.
• Where your data goes: data and security.
• A process in mind? Request an audit.