THE ENGAGEMENT · PER PROJECT

Which process in your company depends on the judgment of two or three people?

Part of it is written down: a procedure, a quality manual, a spec nobody has reopened in three years. The version people actually work from is written nowhere — it lives in the heads of the people doing the job. We rebuild the process around that know-how, instead of laying a coat of AI over how it runs today. The routine then stops going through your experts: it arrives handled.

The point is not to interrupt your experts less. It is that their time goes to the cases that deserve it: the rest comes out just as fast, and better.

This work can only be done while those people are still here. That is the only urgency on this subject, and it does not come from us.

Twenty minutes on one process, and a straight answer →See the whole ladder →

What the engagement is

The problem is neither the model nor your teams: it is everything in between. A model can do the work once; it cannot do it a thousand times in a row without slipping. What is missing is the engineering that makes it reliable, and the domain knowledge that tells it what "correct" means in your company. It is one and the same job, and that is the one we do.

One processOne real workflow where AI should already be paying off, and isn't. Bring it, or let's find it together. Not a company-wide programme.
Your knowledge, made operationalWhat your experts carry in their heads, and what sits in documents nobody ever formalized: the context no foundation model will ever have. We turn it into verifiable assets, inside the workflow itself. They stay yours.
Proven against an honest baselineWhat counts as success is agreed before we start, and measured as we go. Even a no ends with the reason for the no.
The operating model, then the handoverThe redesigned work, and what your teams do next: their judgment moved to where it pays most, never cut. Then the build-out — the solution is yours: it runs on your infrastructure or sovereign hosting, and can evolve without us.

The largest study on the subject measured exactly this: a support organization that trained its system on the work of its own top performers saw novices resolve 34% more cases per hour — and its experts gain nothing. The value did not come from the model, but from the know-how the model finally had access to. (Brynjolfsson, Li & Raymond, Quarterly Journal of Economics, 2025 — 5,172 agents, 3 million conversations.)

Where it starts

Twenty minutes, free, with nothing to prepare. If it is not for you, you hear it during the call — not after you have paid.

The engagement starts with a short slice, running on your own files.

You find out straight away whether your know-how can be captured: if the people who do the work say "yes, that is exactly it", we go on. If they say "no, that is generic", that is a real answer, and it stops there.

We first measure what that process costs you today: real lead times, rework, files turned away. You sign off on that baseline before we start, and it holds both ways — if the number does not move, it is on record.

You are not afraid of spending. You are afraid of paying for something that will not pay back.

What makes an engagement work

One process at a time. These engagements only work under five conditions, and they barely depend on us.

1
One person decides.
One yes, one signature — yours, or that of someone whose door you can walk into. No committee, no steering group, no tender. When three departments have to agree, the decision takes longer than the engagement.
2
The people who know give us their time.
Their know-how is in their heads, and in documents that no longer quite describe what is done. We need those people half a day a week — including the busy weeks — and not their documentation. This is, by a wide margin, the condition that fails most often.
3
The work takes judgment, and it comes back.
Someone decides, arbitrates, says no — and not twice a year. If ten cases out of ten can be handled by applying the procedure to the letter, a plain automation costs less than we do, and we will tell you so. And if that judgment only comes up twice a year, there is no process to rebuild: it is the decision itself that needs the work, and that is what Counsel is for.
4
You know what a mistake costs.
A delay, a rework, a lost customer, a penalty. Without that number there is no honest baseline to compare against — so there is nothing to prove.
5
You accept being measured, both ways.
We agree on what will count as success before we start. If the measurement says no, it stops there, and you leave with the reason.

When I decline

you are looking for a demo to show a shareholder or a board;
the goal is to cut jobs — those projects turn around on the first awkward case: you call back the people you let go;
you want "an AI strategy" rather than one precise process;
nobody on your side will carry the subject after we leave;
you would rather give us access to your data than to your teams.

Who you will be talking to

Led by a PhD in AI (Sorbonne / Inria) — twenty years turning expert know-how into workflows that run: a privacy-preserving search platform for the investigative network behind the Panama Papers, an augmented legal word processor, an Inria–CNIL distinguished-paper award. We have been building AI systems since 2006, well before the generative wave.

If there is a playbook for it, hire the consultancy. If there is not, that is us.

If you tick all five boxes

Twenty minutes, with the person who would do the work. The call can very well end with "this is not for you" — and with what we would do in your place.

Twenty minutes on one process, and a straight answer →

Missing one or two conditions? Start with The Agenda — what a leader needs to know this month — or with a public verdict. Nothing to pay, nothing to commit.