AI adoption in manufacturing: plant or product, the choice that comes first
The same model changes regime depending on where it ends up. The candidate map, the historical data check almost nobody runs, and why a pilot on one line doesn't replicate itself.
In short. In manufacturing the choice of a first use case turns on a separation that precedes every other assessment: whether the system stays in the plant or ends up inside a product that gets sold. In the first case the company is a deployer and the path is short; in the second it becomes a provider under the regulation and the path runs through product conformity assessment. Starting from the plant lets you learn before taking on product obligations, which is why it pays, beyond caution.
In a factory the AI conversation almost always starts from predictive maintenance and quality control, and this time the instinct is right. The problem comes later, when the same technology gets proposed for the product without anyone noticing the change of regime.
A separation that comes before everything
A computer vision model checking parts on the line is a system the company uses for itself. The same model, embedded in a machine the company sells to a customer, becomes a system the company places on the market, with everything that follows in documentation and liability. Everything changes.
The first carries user obligations. The second carries provider obligations, with technical documentation, conformity assessment and liability towards the buyer. The technology is identical, the regime isn't.
This separation has to happen before building the portfolio, because it determines timing, cost and the competence required. A company discovering it mid-project ends up with a product in development and no preparation on the conformity path.
The map of candidate use cases
| Use case | Value | Regulatory friction | Suitable as first |
|---|---|---|---|
| Predictive maintenance on plant | High, in downtime avoided | Low | Yes |
| Visual quality control, not safety-related | High, in scrap reduction | Low | Yes |
| Scheduling and energy optimisation | Medium-high | Low | Yes |
| Documentation assistant for maintenance staff | Medium, in time | Low | Yes |
| Control with a safety function | High | High, changes class | No, not as a first |
| AI embedded in products sold | Strategically the highest | High, Annex I | No, needs a dedicated path |
| Worker productivity monitoring | Medium, contested | High, Annex III and employment law | Assess with worker representatives |
The last row deserves a note that isn't regulatory. These systems often get introduced as efficiency tools, and they produce internal conflict once people understand what they measure. The cost of that conflict rarely enters the business case.
The constraint sits in the data, not the model
In manufacturing the factor determining the result is almost always the quality of historical data. Predictive maintenance works when correctly labelled failure data exists, and in many plants that data sits in a maintenance system filled in inconsistently.
The check to run before any pilot is trivial and many skip it: take a year of history, count how many failure events carry a reliable cause and date, and see whether there are enough to train anything.
When the answer is no, the first sensible project concerns data collection rather than a model. Management doesn't want to hear it, and hearing it avoids spending on a pilot that cannot succeed.
Which framework does your company actually need?
AI Rating measures maturity across the four areas of the model and shows where to start, with priorities and estimated effort.
Start your AI RatingA pilot on one line doesn't replicate itself
The right perimeter in a factory is a line, a cell or a plant. The baseline is measured on that line, not on the plant average, and acceptance criteria get set beforehand.
There is a specificity, though: what works on one line doesn't automatically work on another, because machines, materials and operators differ. Extension has to be treated as a new project with its own verification rather than as a copy. Many companies discover this after buying licences for the whole plant.
Who has to be able to stop the system
Accountability for an AI system in a factory can't sit with IT, which doesn't know the process, nor with the production manager, who is measured on output and has no time. The balance point is usually the maintenance manager or the quality manager, depending on the use case, with a contact in operations management.
The person must have the authority to stop the system when needed. An owner who can't switch it off remains one on paper only.
Where to start
The sequence begins with the AI system inventory, separating plant and product from the very first row.
The general selection criteria are in AI adoption: how to choose your first use case, and the cost model in AI project budgeting.
The sector's regulatory picture is on the AI governance for manufacturing page.
To assess which use cases make sense in your company, you can book a meeting.