AI literacy in manufacturing: where to graft without distorting

In short. In manufacturing, AI training meets a training framework that already exists and is required by law, the one for workplace safety. It is the most efficient graft point and at the same time the risk, because discussing AI inside the safety module leads to treating it as a hazard rather than as a tool to govern. Three roles need individual evidence, and one of them, the product designer, doesn't yet know it at many companies.

In a factory, training has a long tradition and a defined structure, with supervisors, operators and managers, roles the law names and assigns specific duties to. Whoever comes to build the AI plan finds language, calendar and tracking already in place, and the temptation to lean entirely on that framework is strong. It's worth partly resisting.

Two worlds that don't talk

Manufacturing houses two populations with different skills and cultures. The OT world, which governs machines and lines, with a culture of reliability and safety built over decades. The IT world, which governs management systems and data, with a culture of process and project.

AI systems fall between them, and that creates a specific training problem: whoever knows the machine doesn't know the model, whoever knows the model doesn't know the production process. A plan that trains the two populations separately leaves precisely the contact point exposed, which is where problems start.

The workable answer isn't training everyone on everything. It's building one short shared module on the contact points, then separate tracks on the specifics.

The three roles needing individual evidence

Those designing products that will embed AI. This is the role many companies don't yet recognise as critical. An AI system embedded in a product subject to harmonisation legislation falls under Annex I, and conformity assessment is designed alongside the product. Whoever shapes those choices needs to know their consequences, and the 2028 deadline doesn't make the training deferrable, because products in design today reach the market then.

Those introducing systems that measure people. Productivity monitoring, task allocation, performance evaluation: these are Annex III systems, and in Italy employment law constraints stack on top. Whoever proposes or configures them needs to know what that entails.

Those exercising human oversight over systems with a safety function. Here the competence adds to what accident prevention legislation already requires, and the documentation has to hold both together.

Where it grafts, without distorting

Existing framework	What it already covers	What to add for AI
Safety training for operators	Correct machine use, hazards	Behaviour of adaptive systems and when to stop
Supervisor training	Oversight and accountability	What to report on a system that decides
Designer training	Product conformity and marking	Requirements arriving from the AI Act on embedded systems
Training on a new line	Operational use	Known system limits and cases where the data doesn't hold
Quality training	Controls and nonconformities	Difference between a deterministic and a probabilistic control

The last row produces the most immediate value. Quality people think in thresholds and tolerances, and a model-based visual control system works differently: it errs in a different way, and how it errs has to be taught.

When to train, inside a production cycle

In a factory, training has a natural window other sectors lack, coinciding with the introduction of a new line or machine, when training is already budgeted, people have been released from production and the supplier's engineer is physically present on the floor. It is the best possible condition.

Grafting AI content there costs almost nothing extra and produces competence tied to the specific system, which is exactly what clause 7 asks for. Outside that window, structured training competes with production and loses.

The machine supplier isn't the model supplier

A complication typical of manufacturing concerns the chain. The machine builder integrates third-party components, which in turn embed models developed by others. When something goes wrong, working out who answers requires tracing a chain nobody has mapped.

On the training side the company therefore has to require known-limit information from the builder rather than settling for the user manual. Without that information the training stays generic, and in the event of an incident it demonstrates nothing.

Where to start

The prerequisite is the AI system inventory, which in manufacturing must separate plant systems from product-bound ones, because the two groups carry different regimes and deadlines.

From there the matrix gets built, described in a roles-competence matrix for AI. The full requirements are in clause 7 explained without jargon.

The sector's regulatory picture is on the AI governance for manufacturing page.

Our approach to role-based tracks is on the AI Training page. To review your company's plan, you can book a meeting.