AI in industry: the field guide to move from promises to results

I spent a career running plants before doing this job. What strikes me, mission after mission, is that the problem is almost never the technology. The problem is that every single day, something is lost: an operator retires and takes years of hands-on know-how no one ever wrote down, data is entered ten times and used nowhere, and you end up chasing deviations in firefighting mode. AI is only worth it if it tackles that. Not to look good in a demo: to stop the bleeding of know-how and to make visible what you're already burying.

Why now, and why so many projects fail

The tools have changed category over the last two years. But most AI projects disappoint for very down-to-earth reasons: a technology was bought before there was a problem, a tool was rolled out without bringing the teams along, or the AI was asked vague questions. The Power BI no one opens anymore, the Copilot licenses paid for but never deployed, the homemade Excel no one can pick up: these aren't failures of the technology, they're failures of framing and adoption.

The honest promise isn't to automate everything. It's to flip the ratio: move from time lost on data entry, formatting and information hunting, to expert time on the floor. The human stays in the loop, always. It's not "AI and us": it's "AI with us".

The real use cases, by family

Capturing and transmitting know-how (the king use case)

By far the one that comes up most. A written work instruction doesn't transmit the gesture: there's a gulf between what the methods engineer meant to write and what the newcomer understands well enough to reproduce. So we film the gesture while the operator narrates what they do, and AI re-cuts the video into steps and drafts the work instruction, with human review. Everything goes into an indexed library, a kind of "internal Netflix": a newcomer, alone in front of a machine, types what they're looking for and lands straight on the right sequence. You can even adapt the procedure to the person. The know-how no longer leaves with them.

Quality and documentary compliance

In regulated sectors (pharma, fine chemicals, food), documentary formatting eats up an enormous amount of time: dossiers, protocols, work instructions, amendments never folded back in. AI pre-drafts from the existing material and spots contradictions between documents, under the control of an expert who validates. It does not release a batch and does not decide on compliance: it tips the balance toward substance rather than form.

Maintenance and inspection

Inspection robots, acoustic or thermal cameras, trend monitoring: you detect a leak, a hot spot or a drift before the breakdown, and you log readings that can't be falsified. The right question is never "is it impressive", but "does it replace a sensor that would have been enough" and "how do we explain it to the teams".

Making data visible and useful

When people ask me where to start, I almost always say the same thing: make visible the data you already capture but never use. AI does two concrete things: connect software that doesn't talk to each other so a schedule or an indicator comes out automatically, and build dashboards you generate and modify in plain language, without coding and without staying frozen for two years. Making already-captured data visible solves, on its own, a large share of operational problems.

EHS and safety

Digitalized safety induction, a draft prevention plan, last-minute risk analysis, traceable lessons learned: AI prepares and structures, the signatory validates. Here too, the decision that carries responsibility stays human.

Steering, meetings, finance

Automatic minutes, sorting and prioritizing emails, financial analyses made reproducible from a clear instruction: these are the quick wins that give back time from week one. A controller who saves fifteen minutes the day after a session; that's what creates appetite in a team.

Robotics and the physical floor

Quadruped robots or arms, where the environment is harsh, dangerous or repetitive. It's the most visible use case, and the most demanding in change management: it's prepared with the teams, not behind their backs.

To go further, family by family: capturing know-how, regulatory dossiers, inspection robots, the data that sleeps, EHS and prevention plans, and the back-office quick wins.

The method: the Fabrique and the 3-V rule

I never start with the tool. I start with a day in production: watching the flow, listening to the teams, seeing how newcomers are trained and where the blind spots are. And I always start by saying what's going well; in a plant, we're used to talking only about what's wrong.

Then we surface the irritants from the people who live them: that's the Fabrique des idées (the "idea factory"). We synthesize, score each idea on impact and feasibility, and separate quick wins from structural projects. For each retained case I propose several angles (sovereign, standard, innovative, bespoke) and the leader decides. Before tooling up, we build literacy: most people have never written a good prompt. For bespoke tools, we prototype fast, then secure the data seriously. And we always keep the human in the loop. That's my three-V rule: verify, validate, valorize. The goal, in the end, is your autonomy, not your dependence.

Where to start, concretely

Don't launch a big AI program. Take a single irritant that hurts (a critical skill about to retire, an indicator you can't find, a dossier that costs you days) and turn it into a pilot, with a goal and a date. The point isn't to buy AI: it's to recover expert time and to stop losing the most precious thing in your plant, what's inside your people's heads.