For French SMEs of 50 to 500 people
AI where it pays off. Your expertise where it counts most.
Your quotes, your tender responses, your reports: someone re-reads them before they go out. What that re-reading still catches, nobody counts.
In three weeks, we count it on your last twenty files, with the AI you already have. In six weeks, we write with your teams what software can take on alone in this work, what stays with the people who know, and the plan to build it. You will know which files can go out without a re-read, and which ones still need someone.
Free. Twenty minutes to set up the measure: your tool, your files, the date.
No AI at your firm yet? Tell us which document comes back most often, a written answer within 48 hours.
Julien MartinPhD in AI (Sorbonne / Inria), in the field since 2000In every file, some questions are matters of fact. Others take a judgment call.
A matter of fact
"Did the client send the five documents we asked for?"
Software settles it on its own, and nobody goes back over it.
A judgment call
"Is this price defensible on this job?"
Someone answers for it. That question stays with you, however good AI gets.
If someone has to re-read behind the software, you have gained nothing. Ours gives you its error rate, measured, and a sample drawn at random to check it. You control a number, not every file.
Three years from now, your competitors will have the same models you do. They will not have what your firm knows.
What it changes in your firm
Your files stay on your premises or on sovereign infrastructure, and nothing confidential leaves without your consent. The software is designed for the requirements of the EU AI Act.
The same line works in every firm. Where everything is reviewed, it says which files still need an opinion, and the time of the people who know goes back to those. Where each person builds and sends their own file, it says the same thing, except that it now points at files nobody was opening.
Nobody knows where that line runs in their own firm. Writing it down, on your own files and with the people who know, is the work.
ANSWER THIS WITHOUT US
Of your last hundred files, how many would have gone out differently had someone else built them?
Nobody knows, in any firm, because it is measured nowhere. It is the first number we pull from your archives, and on its own it says whether this process is worth the work.
What we offer
Measure first. Build afterwards, if the numbers say so.
If AI is already in place at your firm
The measure — three weeks, €3,000 excl. VAT, not due if the result does not let you decide
Copilot, ChatGPT or your vendor’s module: we measure it on your last twenty files, where it is right, where it is wrong, and on which questions.
In detail
You decide whether to keep it, frame it or stop it. If the answer is to keep it as it is, that is what we will write.
We need the tool as your teams use it, and your last twenty files, the last twenty and not the best, with what the right answer was. We have it redo each one, and we count. You receive the table of the twenty, question by question, and a sample to re-read so you can check our count yourself. The measure runs on your files, on your premises or on sovereign infrastructure.
Where there is no AI yet: one piece of work, not a survey of the whole company. One signature, and you choose where to stop.
- Where to start — free, an answer within 48 hoursIn writing: the piece of work to start with, what it costs you today estimated from your own figures, and the ones that do not qualify. If AI has nothing to do there, we will say so.
In detail
The right first piece of work comes back often, depends on a few people, and part of its questions are matters of fact. There are two ways to find it, your choice.
In writing, without a callYou describe in five minutes what comes back in your firm. We answer within 48 working hours.
Describe what comes back in my firm →In person, in twenty minutesThe same choice, and we also check who can say yes and where your past files are.
Book twenty minutes → - The three reference files — six weeks, €4,500 excl. VAT, deducted from the software if you go onOn a piece of work you name, your quotes or your tender responses, we go through three of your last twenty files with the people who know, half a day a week. You leave with your three files annotated, three numbers you do not have today, and the plan.
In detail
We start with a few interviews with everyone the file passes through, from the person who puts it together to the person who invoices it, to confirm it is the right piece of work. Then you give us the list of your last twenty jobs sent out, the last twenty and not the best, and say for each whether it was good or less good. We pick three together: an ordinary one, one you would no longer send, and an edge case. We go through them in detail with the people who know, half a day a week for four to six weeks.
What you receive, and keep if you stop here
- Your three files, annotated. Question by question: what made the file right, what you would no longer send, and why. It is the first time your firm writes down what it calls a job done right.
- Three numbers you do not have today. How many days a file takes to go out, how much of that time it spends waiting on someone, and how much it changes depending on who built it. They come from your archives, not from your teams’ time.
- The plan for the redesign. For each question in a file, it says who takes it. Matters of fact go to the software, with the cheapest method that holds up: AI when it is needed, a plain calculation when that is enough. Matters of judgment stay with your teams. It also says in what order to build.
These three files lead to the plan. The plan prepares the specification, whether you have it built by us, by another provider or in-house.
This stage does not yet say how many files the software will handle on its own. That is the first thing the next stage measures, on the other seventeen files on your list.
You can stop there, with no penalty and owing us nothing further.
- The process rebuilt — priced at the end of the three filesYou receive the software, on your premises or on sovereign infrastructure. No file goes out until every matter of fact has been checked, the tenth as carefully as the first. The software says which ones still need a judgment call, and your teams have the time to make it.
In detail
Before building, we apply the plan to the other seventeen files on your list. You see how many the software would handle on its own, and what it would miss. You set the threshold with the people who know, and we build around that threshold. When the software does not know, it does not guess: it stops and hands over to your teams. It keeps a record of everything it decided on its own. It runs where the file already travels, with the people who look after your tools. You hold the keys: it evolves without us, someone on your side changes a rule when needed, and you have what it takes to extend it to a second process.
Want to see what it looks like on a real file? Read AI on a tender response. You will see what it writes, where it goes wrong, and why.
- The follow-up — includedYou receive a dashboard. It shows what the software handled on its own, what it sent back to your teams, and what the work gives today against what it gave before.
In detail
The three numbers from the start are measured again, on the baseline agreed before we began. The dashboard also gives the software’s error rate, checked on a sample of files drawn at random. So you follow the software without reopening files one by one. You see where it hesitates, which is where it needs your teams, and where the line has to move.
- Once in place — the software keeps learningEvery file your teams rule on sharpens the line, and you validate every change.
In detail
Cases nobody foresaw always come in, and a line written in advance does not stay right on its own. The dashboard says where it has to move. Nothing changes in the software unless you decided it, and every change is measured like the first. That is how the software evolves with your company and with the people who run it, with no subscription to our presence. We come back when you call us.
Who you will be talking to
You work directly with Julien Martin. A PhD in AI, he has been building tools since 2000 next to the people who use them, in animation, investigative journalism, law and logistics.
The whole trade, before the tool
Tools that hold, since 2000
AI since 2006, not since 2023
In detail
The whole trade, before the tool. A file passes through several hands, and each one knows something the others do not. Before building, we understand who does what, who depends on whom, and what each person has to know for the whole to hold. We take the problem as you live it, and we turn it into a solution that holds, with software exactly where it pays.
Tools that hold, since 2000. He designed film production pipelines. He is behind the search application and the secure network of the ICIJ, the consortium behind the Panama Papers. Each time, the tool had to work that same day, for people with no time to fix it. Laureate of the Brown Institute (Stanford / Columbia). Distinguished by the CNIL-Inria privacy prize jury.
AI since 2006, not since 2023. A PhD in AI, he has led AI engineers and spent ten years in constrained optimisation. He knows how to choose an AI tool and judge it on your files, not on its demo. We use the cheapest method that holds up: AI when it is needed, a plain calculation when that is enough.
Which job comes back the most often at your firm, and costs the most when it goes wrong?
It is free. You will know whether it is doable in your firm.
Not sure which piece of work to pick? Describe in writing what comes back in your firm, we answer within 48 hours.
Frequently Asked Questions
Who is this for — and who is it not for?
For whoever answers for a process in a French SME and can decide to change it: the owner or managing director, the operations or plant director, a branch director, a technical director. You don't need to be the AI expert — that's the starting assumption.
The right process has four traits: someone exercises judgment in it, often; that judgment isn't written down anywhere; a mistake costs something you can put a number on; and you can decide to change it without a tender.
It isn't for you if ten files out of ten are handled by applying the procedure to the letter — a plain automation costs less than we do, and we'll tell you so. Nor if the method belongs to a third party, a software vendor or an accreditation body, whose sign-off any change requires. Nor if the goal is to cut headcount: those projects turn on you at the first awkward file.
What happens in the twenty minutes?
You describe a process, or your business if you do not know which one to pick yet. We ask five questions: how often it comes back and how long it takes, who exercises judgment in it, what a mistake costs when one gets through, where your past jobs live — bids, quotes, post-mortems — and in what shape, and who can say yes.
The call ends with a straight answer: the process we would start with, what it costs you today per year, estimated from your own figures, and what we would do in your place. No deck, no diagnosis to buy, no surprise quote: a quote only comes afterwards, and only if both sides want to carry on.
The person who would do the work is the person on the call.
What does it cost?
The Agenda is free. Counsel is €99/mo, fourteen days free with no card. The twenty minutes cost nothing and commit you to nothing.
Redesigning a process takes one signature. The three reference files cost €4,500 excl. VAT. If you go on, that amount is deducted from the price of the software. The process rebuilt is quoted at the end of those three files. We cannot price it earlier, because the price depends on the process and we do not know it before looking at it with you. The follow-up is included.
The measure of AI already in place costs €3,000 excl. VAT, not due if the result does not let you decide.
You can stop at the end of the three files, with no penalty. You keep your annotated files, your three numbers and the plan.
Before we start, we fix with you what this process costs you today (real lead times, rework, jobs turned down) and what will count as success. Those are the numbers you will follow the software against afterwards.
Where do our data and client files go?
Two answers, because these are two different things.
The software we build for you runs on your own infrastructure or on sovereign hosting, it can evolve without us, and there is no subscription to our presence. Your files never have to leave your walls.
We choose each AI model according to what your files contain. Whatever is confidential goes through models installed on your premises or on sovereign infrastructure: you keep control of them, and their cost does not move at a vendor's whim. A commercial model can serve on a step where it does better, provided it sees nothing confidential. It then works on documents that are not, or on files from which we have first removed the names and whatever you want protected. You decide what leaves, and the record shows which model served at each step.
The EU AI Act requires, depending on the use, that a record be kept of what software decides and that a person answers for it. The software keeps that record, and the matters of judgment stay with your teams. What the regulation requires depends on your use: we look at it with you from the three reference files onward.
Counsel, our online product, is more exposed and we would rather say so: the text of the decisions you submit to it is analysed in the United States by Google LLC (Gemini API) and Tavily Inc., processors under Article 28 of GDPR, who do not use it to train their models. Everything else — account, reading, billing — is hosted in France with OVH and does not leave the European Economic Area. The full list is in our privacy policy.