Understand

Automating your business with AI

You do not automate a company. You take away from it, one at a time, the tasks nobody wants to do and nobody does well. The difference is not rhetorical: it decides whether the project succeeds. A plan that starts with “automate customer service” never lands; a plan that starts with “answer requests for duplicate invoices” lands in a week, and opens the way to the rest.

1. What changed since scripts and RPA

Business automation has been around a long time, and it left a bad memory. A script does exactly what it was told and nothing else: the day a supplier renames a column, it stops or, worse, carries on with wrong data. RPA industrialised that principle by imitating a human's clicks inside software — with the same fragility, and the obligation to redescribe everything each time a screen changes.

What changed is not execution speed, it is the ability to handle a case nobody planned for. An agent that receives a request phrased differently, a spreadsheet whose columns have moved, a client who answers beside the point, does not stop: it relates the situation to what it knows of its job and takes the next step. That is precisely where the 20% of cases sat that made classic automation unusable — because handling them separately cost more than doing everything by hand.

The practical consequence is that you no longer automate a sequence of screens, but an objective. Setting it up no longer means describing steps; it means stating a rule and saying what must come back to you before anything is done. That is what a agentic system is, and it is written in plain language, not in code.

2. What genuinely gets delegated

In almost every small company, the same quarter of the working week disappears into four families of tasks. None of them is a profession, all of them are necessary, and that is where to start.

Work that repeats identically — quotes to draw up, invoices to chase, notes to write up, files to sort. The rule fits in one sentence and the result is checked at a glance.
Work that needs consistency rather than talent — keeping a CRM up to date, calling a prospect back at the right moment, publishing regularly. A human does it better once; an agent does it every time.
Work that consists of reading and sorting — going through incoming requests, qualifying them, routing them to the right person, condensing a twenty-message thread into three useful lines.
Work that never gets done for lack of time — market watch, reactivating dormant clients, updating the documentation, search visibility. The work whose absence does not show immediately, and costs the most.

What cannot be delegated is not what is “too complex”: it is whatever commits the company. Signing, hiring, dismissing, granting a discount, paying, terminating a contract. Not because an agent would be incapable, but because a mistake there is expensive and sometimes irreversible. Those acts go through human approval — and a serious system makes that boundary explicit rather than leaving it to everyone's caution.

3. The four criteria of a good first task

The choice of the first project matters more than the choice of tool. A good first task meets four conditions, and it needs all four.

It comes back often. At least several times a week. Below that, the setting-up time is never repaid and nobody gets into the habit of using it.
Its rule fits in one sentence. If you cannot tell a new employee how to do it, you will not be able to tell an agent either — and the problem is not the agent.
Its result is checked in seconds. You must be able to say “that's fine” or “that's not fine” without an investigation. Otherwise you will never know whether the system is working.
A mistake can be undone. A clumsy email is fixed with a follow-up message; a payment that has left never comes back. Start on the reparable side.

Chasing unpaid invoices ticks all four boxes, which is why it is the most common first project — and the most legible, since it is measured in cash. Sorting incoming requests comes a close second.

4. The three mistakes that sink a project

Starting too big. “Automate admin” is not a project, it is an intention. It produces six months of analysis, a specification nobody rereads, and no result. One task delivered in a week changes a company more than a year-long programme, because it teaches the team what actually works.

Automating a broken process. An absurd approval chain stays absurd once automated — it simply becomes fast and harder to challenge. If nobody in the company knows why a step exists, the question is not how to automate it but whether to delete it. Automation exposes bad processes; it does not repair them.

Not knowing what the system did. This is the mistake that kills projects after three months, and the quietest one. Delegated work you cannot review becomes work you no longer trust, and the company quietly goes back to doing it by hand. Being able to see what was done, by whom, at what cost is not a management comfort: it is what lets you delegate more the following year.

5. What it changes inside the organisation

The real change is not a headcount reduction — it is a shift in the load. The work that disappears is the work nobody claims: re-entering, chasing, filing, reminding. The work that grows is framing: saying what you expect, setting the limits, reviewing what came out.

It calls for a new skill, and it is better said plainly: knowing how to state a clear rule. Many companies run on unwritten habits passed on by word of mouth, and discover when delegating that they had never put their own criteria into words. That is uncomfortable in the first week, and it is often the real gain of the project — quite apart from the technology.

One practical consequence: the person who should lead this is not the IT specialist, it is whoever knows the work. An agent is tuned by describing the job, not by programming it. Handing the project to the technical team is the surest way to get a correct tool nobody uses.

6. The one figure to watch at the start

Not the number of automated tasks, not the hours theoretically saved: the human rework time. How many minutes does it take each week to fix what the agent produced before it is usable?

That figure has the virtue of being honest. If it falls, the delegation is working and you can add more. If it stays flat, the rule is badly worded — the useful reflex is to rewrite the instruction, not to change the model. If it rises, the task was the wrong one: take it back by hand and pick another, without regret.

The second figure, worth watching as soon as several agents are running, is the cost per task. Delegated work whose price you do not know is work you cannot judge worth delegating.

7. Frequently asked questions

Do I need to change tools to automate my business with AI?

No — in fact the opposite. Replacing your tools at the same time as you delegate the work makes it impossible to know what worked. An agent works with what is already there: the mailbox, the spreadsheet, the CRM in place. Tool migrations are a separate project, to run afterwards if the need is confirmed.

How long before a first visible result?

For a properly chosen task — repetitive, simple rule, verifiable — a few days are enough to get a usable result, and one to two weeks to stop having to review it every time. If a first project has produced nothing after a month, it is almost never the technology: the task was too broad, or its rule could not be stated.

Does this replace jobs?

In the kind of company this sort of tool is built for, the delegated work is rarely anyone's job: it is the work nobody has time for. The follow-up that never goes out, the market watch that stopped, the half-filled CRM. The honest question is not “who does this replace”, but “who, today, is not doing it”.

How do I keep control over what is done in my name?

Through two mechanisms, and not through trust. First a declared boundary: the list of acts that require human approval before being carried out — committing money, signing, replying to a sensitive client. Then a reviewable trace of every action, with its cost. A system that does not provide both cannot be controlled, however good its answers are.

Which task should I actually start with?

Look at what you have been putting off for three weeks. That is almost always the right answer: the work you postpone is the work that needs no particular talent, only consistency — so exactly the work that delegates well. Chasing unpaid invoices and sorting incoming requests are the two most common starting points.