AI in the Factory: Is It Worth It?

Adoption is doubling — the factory is still waiting

Start with the question every industrial leader asks before committing a single euro: is this worth it for a company like mine? The honest answer is yes — but it immediately calls for a caveat, because the average hides what matters. At the end of 2025, 18.2% of French companies used at least one AI technology, up from 10% a year earlier (INSEE / Eurostat): a near-doubling in twelve months that puts France squarely in the European movement — the EU stands at 19.95% in 2025, across 157,000 surveyed companies (Eurostat). The shift is real, fast, and no longer confined to large technology groups.

Manufacturing, however, is running behind: 7% of French manufacturers used at least one AI technology in 2024, against 10% across all sectors — even if progress is real (5% in 2023) (INSEE). This lag is not a French anomaly: everywhere in Europe, factories adopt more slowly than services. But it means one precise thing for an industrial leader: the competitive advantage is still there for the taking. In a sector where 93% of competitors have not yet structured their approach, the first company to industrialize one profitable use case opens a gap the others will need years to close.

The decisive divide is not sectoral — it is dimensional. In France, 9% of companies with 10-49 employees use AI, 15% of those with 50-249, and 33% of those with 250 or more; the European Union reproduces exactly the same slope at a higher level: 17%, 30.4% and 55% (INSEE; Eurostat). A one-to-four ratio between the small company and the large group. The pessimistic reading — "size decides" — is contradicted by the field: several of the most profitable cases documented in recent months are SMBs, as we will see below. Size buys speed of entry (a dedicated team, an experimentation budget, a consolidated information system); it does not buy the discipline that turns a trial into a result. The lag exists, it is measurable — and it owes less to resources than to how the problem is framed.

One last signal, and it is counter-intuitive: where AI has entered the factory, it did not start with production. Among European manufacturers using AI, the dominant purposes are marketing and sales (30.4%), then administration and management (27.1%) — production does not come first (Eurostat). AI entered the factory through the offices, not through the lines. For the leader, this anomaly is strategic information: the room for progress sits exactly where industrial value is created — quoting, quality, maintenance, process — and that is precisely where the documented returns are highest.

The momentum on the leadership side confirms that the subject has changed status. Among French SMBs and mid-market companies, generative AI use by the leader personally went from 31% to 55% in one year (Bpifrance Le Lab); 43% have defined an AI strategy, and 58% consider AI important or very important to their company's survival over the next three to five years (Bpifrance Le Lab). The question is no longer "should we go" — it is "where, and with what discipline".

The paradox: more usage than ever, still almost no impact

Using AI "somewhere" has become ordinary: 88% of organizations report using it in at least one function (McKinsey). Getting a result that shows up in the accounts remains the exception: only 39% attribute any impact on operating income to AI, most often below 5% (McKinsey); barely 5% of organizations have genuinely deployed AI at scale (Capgemini / Microsoft); and about 95% of generative AI pilots produce no measurable return on the P&L (MIT Project NANDA).

These four figures tell a single story, and it is a useful one because it clears the technology. If 88% use and only 5% transform, the bottleneck is not access to models — it is everything around them: data never prepared, processes never changed, and above all a measurement of results that was never set. A pilot launched without a baseline number can neither succeed nor fail: it can only last. That is the exact mechanism feeding MIT's 95% — not failing technologies, but demonstrations never converted into decisions.

French industry illustrates this paradox with precision. Bpifrance classifies industrial leaders first among the "Experimenters" — 28% of the sample: open to AI, encouraging exploration, but held back in deployment by financial constraints and a lack of in-house expertise (Bpifrance Le Lab). The typical profile is neither the sceptic nor the convert: it is the manufacturer who has tried, has seen that "it works", and does not know how to move from demonstrator to production station. Everything that follows in this report is addressed to that leader.

One last order of magnitude deserves to be set down: within a single sector, the maturity gap between the most advanced third and the most lagging third frequently exceeds a factor of three (range — YPrtnrs Institute estimate). Your sector matters less than the state of your data. That is good news: unlike your sector, that state can be changed.

If 88% use and 5% transform, the problem is not the technology: it is everything that was not done around it.

Six use-case families, one common logic

Value is not spread evenly across the factory: it concentrates in a small number of use-case families that share three traits. The task is repeated often; its current cost is measurable; the improvement translates directly into time saved, material saved or errors avoided. When the three conditions are met, AI pays; when they are missing, it produces demonstrations that commit no one. Six families meet these conditions in production today, with published results — and the point of reviewing them is not to collect them, but to understand why each one works, so you can recognize the one that matches your situation.

Quoting and estimating is often the best entry point, for a structural reason: the data required — sales and pricing history — already exists, and the gain is measured in days from the first week. At Yaskawa, one of the world leaders in industrial robotics, the time to produce a commercial quote went from eight days to one day across a catalogue of more than 120 configurable models (Tacton). Flowserve, a maker of pumps and valves, multiplied its quote throughput by 2.5, brought unit time down to 5-20 minutes, and lifted on-time quote delivery from 85% to 95% (Oracle CPQ). Xometry takes the logic to its conclusion: its engine analyzes the geometry of the customer's CAD file and configures price and lead time in moments (Xometry). The lesson reaches beyond quoting: wherever an expert spends hours re-computing what the history already knows, the same mechanics apply.

Quality through industrial vision is the most mature family on the market — the one where the gap between large groups and SMBs has narrowed most. Bosch brought its false-reject rates down to 5% with cycle time reduced by 15%, and runs AI in about fifty plants across more than 2,000 connected lines (Bosch). BMW generalized AIQX, its camera-and-deep-learning quality control, across all of its plants (BMW Group); Michelin counts tire inspection among more than 200 AI use cases in production (Michelin / Dataiku). But the most instructive case for an SMB leader is elsewhere: Setic Pourtier, 85 people in Roanne, a maker of machines for the cable industry, integrated machine vision directly into its machines for continuous quality monitoring (France Num). What used to require a data team now requires a camera, a standard model and an integrator — automated visual inspection has become a piece of equipment, no longer a research project.

Predictive maintenance buys production time, and its economics rest on one condition: the existence of a failure history. When it exists, the return is fast — a carmaker cut unplanned downtime by 12% in twelve weeks with Siemens Senseye (Siemens); BMW avoids more than 500 minutes of assembly interruption per year in Regensburg by monitoring its conveyors (BMW Group); SNCF continuously monitors more than 1,100 trainsets (SNCF); easyJet avoided around 80 flight cancellations in two months with Airbus Skywise (Airbus / easyJet). When the history does not exist, the family disappoints: the sensor does not replace the memory, it only begins to build it. That is the first honest trade-off to make — and many disappointments come from having dodged it.

Process and energy form the most discreet and most defensible family: its gains are recurring, contractual, and fall straight to the bottom line. ArcelorMittal reports a return on investment of under two years on AI furnace control in Luxembourg (ArcelorMittal); HeidelbergMaterials cut the energy consumption of a cement plant by 2.2% through AI kiln control (HeidelbergMaterials); Unilever saved 2.8 million dollars of energy and gained 1 to 3% of productivity at its Valinhos site with a digital twin (Unilever). At the far end of the spectrum, Schneider Electric's Putuo plant shows +82% productivity per person and a product development cycle reduced by 63% (Schneider Electric) — and Air Liquide halved its night-time stock-counting errors by having generative AI "clean" the images before analysis (Air Liquide). As soon as an energy or material line weighs on your costs, this family deserves priority study.

Supply chain and procurement is the family where large groups report the biggest absolute returns — Michelin values the return on its data/AI projects at more than 50 million euros per year, led by early detection of stock-outs (Michelin / Dataiku); BMW cut tender-preparation time by 40% with a multi-agent system used by 1,800 buyers (BMW Group); Siemens identified some 150 distributors of a critical polymer in days to secure a supply line (Siemens / AWS). The effect scales with volume — which makes it a second-step family for an SMB, and a first-step one for a multi-site mid-market company.

Technical knowledge, finally, is the newest family and perhaps the most transformative over time: it makes the factory's accumulated know-how queryable. More than 100 industrial companies use Siemens' Industrial Copilot for automation engineering in natural language (Siemens); Schneider Electric halves configuration and documentation time — line changeovers that took weeks now take hours (Schneider Electric); at Airbus, teams query manufacturing work instructions in natural language instead of leafing through dozens of pages, and at Renault, the operator describes the fault to receive repair suggestions (Airbus; Renault). The prerequisite is a clean documentation base — which is why this family rewards first those who invested in their operating procedures.

Six families, one common trait: the gain does not come from the model. It comes from a repeated task whose cost was measured before — and measured after.

Proof at human scale

The classic objection — "these are large-group results" — no longer holds. The documented French cases at SMB and mid-market scale tell the same story as Bosch or Michelin, with the same rules: measure the baseline cost, instrument, decide. And they add a lesson the groups do not teach: at human scale, discipline replaces resources.

Taken together, these cases sketch the identikit of the project that succeeds at intermediate scale: a narrow scope, data that already existed, a measured baseline cost, and a decision date set before launching the trial. None of these four ingredients is expensive. Each is missing in most of the pilots that stall.

True costing: the 10/20/70 rule and the forgotten 70%

Why do so many projects overrun their budget? Because the initial costing almost always describes the same object — the license and the integration — while the real cost sits elsewhere. The effort distribution of a successful AI project is now well established: 10% algorithms, 20% technology and data, 70% processes and change management (BCG). In other words, the algorithm is the small share of the work — and it is precisely the only one most budgets price.

The accounting consequences are documented. In an AI project, 55 to 75% of the total cost sits outside the software license — setup, change management, operations (YPrtnrs Institute estimate); and the gap between the cost actually observed and the budget announced at the outset frequently reaches a factor of two to four (Gartner). The overrun is not fate: it is the accounting trace of the 70% no one wanted to see. A leader who demands a total-cost budget where change management and operations are visible on page one eliminates, before starting, the first cause of disappointment.

Against these costs, the documented returns remain clearly favorable — provided you see it through. Generative AI investment returns on average 3.7 dollars per dollar invested, and 10.3 dollars for the most advanced organizations (IDC). As for timing, the cases cited in this report bracket the range: six weeks when the data was ready (Safran/Pelico), twelve weeks on an existing failure history (Senseye), under two years on energy control (ArcelorMittal) — and up to eighteen months when everything must be built. The payback period is not a property of the technology: it is a property of the state of your data at the start.

A project's journey — and its four breaking points

Post-mortems of industrial AI projects look alike. Laid on top of each other, they reveal a typical journey — and four breaking points where the trajectory tips, almost always in the same places. Knowing them before you start changes the nature of the project: each break is prevented by a decision that costs nothing, provided it is taken before launch.

1. Break #1 — choosing the use case: the showcase pilot. The case is chosen for its demonstration effect, not for its measurable cost. Without a baseline number, the pilot can neither succeed nor fail: it joins the ≈ 95% with no measurable return (MIT NANDA). The counter: choose within one of the six documented families, where your current cost is easiest to quantify.
2. Break #2 — the data audit: optimism. Process industries (food, chemicals, plastics) often hold a dense history captured by their automation systems; discrete manufacturing (machining, boilermaking, assembly) lives with data fragmented across ERP, files and team memory. The same use case does not have the same entry cost depending on that starting point. The counter: audit the data of the chosen case only — not the whole factory — with a written verdict within fifteen days.
3. Break #3 — the costing: the invisible 70%. The budget describes the license and the integration, then discovers along the way the training, the process adaptation and the operations — hence the ×2 to ×4 factor (Gartner). The counter: demand a total-cost budget where the 70% is visible before signing.
4. Break #4 — the decision: the pilot that lasts. No decision date was ever set; the pilot is renewed quarter after quarter, "promising" without ever being settled. And when trust replaces control, the bill can be public: Deloitte Australia had to partially refund a 290,000-Australian-dollar government report containing AI-fabricated citations (The Register) — in the factory as elsewhere, every AI deliverable passes human review until its error rate is characterized. The counter: put the decision date — industrialize or stop — in the executive committee's calendar before the pilot's first day.

Each breaking point is prevented by a decision that costs nothing — provided it is taken before starting, not after.

What the leader must decide — and when

Neither a three-year transformation plan nor a cautious moratorium: the trajectory that works, judging by the cases carried through, holds in five decisions over ninety days — each with its indicator, each answering one of the breaking points above.

1. Choose a family, not a technology. Among the six documented families, take the one where your current cost is most measurable. Indicator: a validated baseline cost figure (hours, scrap, kWh, delays).
2. Audit the data of the chosen case — not the whole factory. Does it exist, is it accessible, since when? Indicator: a written verdict within fifteen days.
3. Cost at total cost. License + integration + change management + operations, applying the 10/20/70 rule and the 55-75% outside-the-license range. Indicator: a budget where the 70% is visible.
4. Set the decision date before starting. Industrialize or stop, at twelve or twenty-four weeks depending on the family. Indicator: the date sits in the executive committee's calendar.
5. Instrument the measurement from day one. The end figure is compared with the baseline figure, on the same scope. Indicator: a before/after dashboard active from week 1.

The third scenario — doing nothing — also has a price, even if it appears on no budget: the adoption gap between company sizes has widened every year since 2023 (INSEE; Eurostat), and the gains documented in this report are recurring — every year of delay leaves them to competitors. This is exactly the discipline that separates the ≈ 5% who scale from the ≈ 95% who pilot without end: not more resources — more method.