
Why AI Use Is Doubling in Italian Companies but Not Reaching the Shop Floor
AI adoption in Italy has doubled, but it stops short of the shop floor. The real limit isn't the technology, it's the shape of operational knowledge.
In August much of Italian manufacturing stops, and the stoppage is not empty time. It is the only window in the year when the lines can be worked on without taking hours away from production, so it is where extraordinary maintenance concentrates, along with component replacements and changes to tooling, often carried out by external contractors working in a closed plant. The result is that in September a plant restarts that is no longer quite the one that stopped in July.
The documentation, by contrast, is still July's. The misalignment this creates is not a formality, because the restart is the phase in which those documents are actually used. Data collected by the UK Health and Safety Executive on the offshore sector quantifies the phenomenon with unusual precision: between 1992 and 2015, of 4,656 reported accidental hydrocarbon releases, 773 cases, or 16.6%, had reinstatement or start-up as the operational mode at the time of the event. Since these two phases account for a minimal share of a plant's operating life, the authors conclude that reinstatement and start-up are particularly high-risk activities (HSE, Hazards 30, 2020). The disproportion, not the absolute figure, is what matters.
A restart concentrates three conditions that rarely occur together. The HSE itself, in its COMAH guidance on operating procedures, notes that many of a plant's potential hazards are realised precisely during start-up or shut-down, and for that reason requires specific procedures that account for every eventuality (HSE, COMAH, operating procedures). In discrete manufacturing the consequence is not the release of a hazardous substance, but the mechanism that produces it is the same, and it surfaces as scrap, rework and a longer ramp-up than planned.
The gap that opens in August comes from three distinct mechanisms, worth keeping separate because each is corrected differently:

The third mechanism deserves an example, because it is the one that dismantles confidence in formal control. Among the cases documented by the HSE, one concerns a plant where modifications had been made years earlier without being recorded on the process drawings: an isolation valve had been added, and other valves had been removed with no record made. The visual checks carried out before restart did not detect the discrepancies and identified as correctly positioned valves that no longer existed on the plant (HSE, Hazards 30, 2020). The control was in place, the procedure had been followed, and for exactly that reason nobody had cause to doubt it.
The same report identifies the underlying causes of the weaknesses found during inspections: inadequate supervision and oversight of reinstatement activities, weaknesses in hazard awareness and competence, failure to manage change, and insufficient consideration of human factors, including the provision of suitable operating instructions. It adds a detail that holds well beyond the offshore perimeter: in some cases the decision whether to apply formal recording systems was a matter of custom and practice at a particular installation, rather than something mandated by company standard. That is the difference between a practice that depends on who is on shift and a standard that holds regardless of the people involved.
This is the same Physical Data Gap described when we wrote about the gap between written procedures and shop-floor practice, observed here at a specific moment rather than as a slow drift. The summer shutdown does nothing more than compress into three weeks a misalignment that elsewhere matures over years, and the debt accumulated in that window is settled all at once in the first week of production.
Added to this is an availability constraint that makes after-the-fact reconstruction harder every year. According to the July 2026 Excelsior bulletin, recruitment difficulty affects 42.6% of the positions companies are offering, down from 45.4% the year before. It remains concentrated, however, exactly where the restart needs it: skilled production workers are hard to find in 57.3% of cases and technical professions in 50.7% (Unioncamere and Excelsior, 2026). The breakdown by occupation is sharper still: craft mechanics, fitters, repairers and maintenance workers stand at 67.4%, and maintenance technicians for fixed and mobile machinery at 66.5%. These are precisely the profiles that, in September, are supposed to reconstruct the method nobody wrote down.

The quantification cited comes from the process industry, where what is at stake is the containment of hazardous substances, and the authors themselves extend their conclusions to the perimeter of plants handling such substances. Transferring those numbers to a machine shop would be wrong. What does transfer is the mechanism, not the measure: documentation describing a superseded configuration, and a check conducted against that documentation, produce a falsely reassuring result in any context, because the defect sits in the reference and not in the execution of the check. In discrete manufacturing this translates into out-of-tolerance parts and hours of set-up, items no accounting system isolates and which for that reason are never traced back to the August shutdown as their cause.
The useful conclusion is not to write longer manuals, because the problem is not the quantity of documentation but the moment at which it is produced. As long as updating remains an activity that follows the work, it will fall to someone who was not present when the work was done, and it will go on being the first thing to be dropped. Some manufacturing companies are approaching this the other way round, capturing the setup directly from whoever performs it, while they perform it, as video that becomes a structured, versioned procedure. The August modification is then recorded at the moment it is made, instead of being reconstructed from memory in September.
It is worth distinguishing where this genuinely pays. On a marginal work instruction the misalignment is tolerable, and formalising it costs more than it returns. It becomes decisive on the tooling that determines product conformity, on the phases critical to safety, and on everything physically touched during the shutdown, which is also the shortest and most manageable list. A multi-site operation has a further reason to do it, because the same intervention carried out in two different plants produces two divergent configurations if neither is recorded.
The summer shutdown comes round every year, and every year it hands back a plant slightly different from the one given over to maintenance. The difference between a company that accumulates that misalignment and one that closes it lies not in the tools available, but in the moment it decides to record what it has changed. The information asset that stays in the head of whoever carried out the work is not lost because nobody considers it valuable, but because the right moment to collect it lasts as long as the shutdown, and by September it has passed.

AI adoption in Italy has doubled, but it stops short of the shop floor. The real limit isn't the technology, it's the shape of operational knowledge.

One in four cyberattacks on manufacturing worldwide hits an Italian company. With NIS2, governing factory data becomes a legal obligation.

ISO 9001:2026 lands this autumn, but the real audit risk isn't the new clauses: it's documented procedures that no longer match how the work is actually done on the floor.