One of your processes, rebuilt: AI where it pays off, your expertise where it counts most.

A model that handles one file well does not handle a hundred in a row. On business tasks, success falls from about 58% in a single turn to 35% across several. What gets through is an answer given to a client with complete assurance, and wrong.

AI drafts, an expert reviews: across 106 experiments, the pair does worse on average than the better of the two alone, because the human follows the machine when it is wrong, and training does not fix that. The authors point to another route: give the machine only the parts it is clearly better at. Three experiments out of 106 had tried it.

Where a file passes through someone before it leaves, it spends most of its life waiting. Across 2,501 medical exams, turnaround fell from 60 to 48 minutes while working time did not move: only the wait went down. On business documents, Microsoft measured a 6 to 12% gain, and nothing on documents taking over a week.

To move a lead time, you need to know which of the ten files still needs an opinion. Writing faster changes almost nothing.

Whoever answers before you takes the deal, even when they are worse: before signing, a client checks price and lead time; quality, they find out afterwards.

"Why was this job priced the way it was?" — Ask Monique. "This one — how do we handle it?" — Ask Dave.

There is a Monique and a Dave in almost every company. Neither was ever given the job of knowing: they know because fifteen years of jobs crossed their desks.

We tried writing that knowledge down: forty pages nobody opens. Handing it to a model as is does not work either: it applies the rule seven times out of ten, and on the rare cases it answers wrongly with complete assurance.

So the people who know go back over all ten, because nothing says which three are wrong. Their procedure says what to do; it does not say which files deserve an opinion, and nobody ever noticed. That is where their week goes — on seven files nobody had asked them to look at. And in the firm next door, where each person builds and signs their own files, nobody goes back over anything. All ten go out, the three with them.

When the right way to handle a file is known to only two or three people, the firm cannot work without them. It cannot teach it to newcomers either. Each newcomer relearns it alone, waiting for their opinion, and what they learn leaves with them if they go. That is the only real urgency here, and we are not the ones making it up.

Seven times out of ten — SOP-Bench (2025): 2,000+ tasks drawn from expert-written procedures across 12 domains; the best model-agent combination tops out at 72.4% task success.

Your best people cannot be everywhere. The way they work can.

Once it is in the process, that way of working serves the newest hire as much as the oldest hand, and from one branch to another the same case gets the same answer. In a support desk that had trained its software on the work of its best people, beginners handled 34% more files an hour, and the most experienced gained nothing.

A team that aligns on a market model ends up answering like its competitors. A team that aligns on its own way of working resembles nobody but itself. Three years from now your competitors will have the same models you do. Not your doctrine.

(Brynjolfsson, Li & Raymond, Quarterly Journal of Economics, 2025 — 5,172 agents, 3 million conversations.)

Part of your know-how is written down. Your firm reads those texts its own way, above all when a case does not fit the box, and that way is written nowhere. It is your doctrine, in the legal sense.

1. With the people who know, we go back over jobs already sent out: this one was right, this one we would not send again, and why.
2. The software runs them again and returns two things for each: an answer, and a solidity score on that answer.
3. We read the table: cut here, and it handles six out of ten, and here is what it misses. You choose where to cut, with the people who know.
4. Below the cut, the software works alone, and a random sample checks its error rate. Above it, the file goes back to whoever can judge it. When the software does not know, it does not guess: it escalates.

Your teams stay on all their files. What we take off them is a review that was finding nothing.

AI is already in your firm, but in the corners: across 11,021 companies, 26% use it at all and about 1% use it as the tool of a business function (France Num, 2025). We do not take work whose purpose is to cut jobs: it turns on you at the first awkward case.

Sources — CRMArena-Pro, Salesforce AI Research (4,280 tasks). Vaccaro, Almaatouq & Malone, Nature Human Behaviour, 2024 (74 studies, 106 experiments, 370 effect sizes). Batra et al., American Journal of Roentgenology, 2023 (2,501 exams: turnaround 59.9 → 47.6 min, wait 33.4 → 21.4, reading 26.5 → 26.3). O’Neill et al., Radiology: Artificial Intelligence, 2021 (the alert alone changes nothing; reordering the worklist: 15.5 → 12.0 min). Dillon et al., Microsoft Research, 2025 (randomised trial, 6,000 workers, 56 firms).

Among French companies of 50 to 249 people, 31% use any AI at all (INSEE, 11,000 firms). Buying the tool was the simple part. Rebuilding one process around what your teams know is the work we do.

These measurements come from a radiology department and a vendor's own telemetry, not from a French SME. In your firm, we measure before we build.

What our claims rest on →

No — that's the normal situation, and it's the work. If everything were written down and followed to the letter, there would be nothing to redesign.

What we need isn't your documentation: it's your past jobs — bids, quotes, post-mortems — and half a day a week with the two or three people who know, including the busy weeks. That is by far the condition that fails most often.

One practical point settles itself in twenty minutes: where those files live. A shared drive, however badly organised, is enough. Scattered across individual mailboxes, no.

A first number from the twenty minutes: what this work costs you today, estimated from your own figures.

At the end of the three reference files, four to six weeks later, you have three numbers measured on your archives and the plan for the project. They are the baseline.

The software's return is measured once it is in place, against those numbers, in the follow-up. The software is built so that it can later extend to a second piece of work.