AI in industrial production: where it really helps, where it over-promises

AI in production is not short of spectacular demonstrations. What is missing is a way to tell what actually holds up on the shop floor from what remains a trade-show promise.

The difficulty is real: the same phrase, "AI in production", covers both scheduling optimisers that have been running for years in some plants and an "autonomous factory" that, for its part, still belongs largely to the sales pitch. An industrial decision-maker needs to sort through it.

This article does that sorting, use case by use case. For each one, a single question: on which problem does AI in industrial production deliver a measurable gain today, and where does it over-promise?

Planning and scheduling: the most tangible help

This is the use case where AI, or more precisely constrained optimisation, delivers a visible gain the fastest. Scheduling a production run means solving a combinatorial problem: sequencing work orders while accounting for changeovers, machine availability, customer deadlines and shift teams. The number of combinations quickly exceeds what a planner can explore by hand.

An optimisation engine proposes sequences that cut changeover times, smooth the workload and respect priorities. In food processing, this translates into fewer intermediate clean-downs; in metallurgy, into campaigns better grouped by grade or by thickness.

The honest point: the tool proposes, it does not decide. A good scheduler remains a decision-support system. The planner keeps control to handle the unexpected (an urgent order, a breakdown, a supplier hiccup) that the model did not see coming. Tied to the real availability of equipment, scheduling in fact meets the challenge of reducing unplanned downtime: a plan that is optimal on paper is worth nothing if the machine goes down on Tuesday.

Optimising process settings

On a continuous or semi-continuous process, many parameters influence yield and quality: temperature, pressure, flow rates, dosing, residence time. Here AI is used to model the link between these settings and the outcome, then to suggest set points that reduce variability, scrap or material consumption.

In chemicals, this might target a more stable reaction yield; in food processing, a target content held more consistently batch after batch. The models sometimes act as a soft sensor, estimating a quantity that is hard to measure online from the variables that are available.

The condition for success is strict: the process must be instrumented and stable enough for the model to learn something other than noise. And the suggestion must stay bounded by process safeguards and validated by the operator. An "optimal" set point that falls outside the safety envelope is not an optimisation, it is an incident in the making. To explore settings without risking the real line, a digital twin offers a safer testing ground than production.

Forecasting demand: useful, but not magic

Demand forecasting feeds the sales and operations plan, and therefore stock levels and line loading. Statistical and machine-learning models do better than naive averages when there are seasonal patterns, promotional effects or recurring trends. The gain shows up on two fronts: fewer stockouts and less overstock.

Here too, the limit is clear. A model learns from the past; it does not predict a market disruption, a sudden regulatory change or a geopolitical shock. Its value is to reduce everyday uncertainty, not to eliminate it.

Tracking OEE and detecting drifts

Overall equipment effectiveness (OEE, known as TRS in French) is the reference indicator for a line's performance. Its definition is standardised (NF E60-182), which matters: without a shared definition, two workshops are not talking about the same figure.

AI's contribution is not to calculate OEE, a simple automated feed already does that, but to make use of what hides beneath the aggregated indicator. A stable average OEE can mask a proliferation of micro-stoppages at one specific station, or a slow drift in throughput over a given time slot. Automatically bringing together the stoppage events, the reasons logged and the machine parameters surfaces these root causes that manual reading lets slip.

The value is in the filter, not in exhaustiveness: flagging the three emerging drifts that deserve action beats a dashboard of a hundred curves that no one looks at. It is the same principle as anticipating failures: the useful signal often already exists in the data, it is simply not brought together.

Energy: an often under-used opportunity

Energy is a major cost item in chemicals, in metallurgy and on any drying or refrigeration installation. It is also an area where AI provides concrete services, within an energy management system along the lines of ISO 50001.

Three uses hold up. Consumption forecasting, to shift loads towards the cheapest or lowest-carbon time slots. Anomaly detection, to spot an overconsumption that signals a leak, fouling or a drifted setting. And tracking energy per unit produced, which reveals the stations that are truly energy-hungry rather than the overall bill.

The gain is all the more solid because it is measured directly in kilowatt-hours avoided, a figure that is hard to dispute.

What is oversold: the "autonomous factory"

That leaves the most conspicuous promise: a factory that runs itself, without human intervention. It has to be treated honestly, because it clouds the assessment of everything else.

Closed-loop autonomy exists, but on narrow, well-understood scopes: a control loop, a sorting station, a dosing loop. Extending that autonomy to a whole site assumes controlled variability and a complete model that few installations bring together. On top of that comes a fundamental requirement: on sensitive uses, European regulation (the AI Act) mandates effective human oversight, precisely because full automation is neither desirable nor safe by default.

• Confusing demonstration with deployment.An impressive pilot on a chosen scope says nothing about how it holds up at scale, on noisy data and unforeseen cases.
• Aiming for the autonomous factory as the goal.The useful goal is reliable decision support, not removing the human. The second framing sells better and delivers less.
• Believing a generic model knows your process.Without data specific to your installation, the optimisation reasons about an average factory that does not exist.
• Promising a percentage gain before any data audit.The figure announced almost never survives first contact with the real history.
• Piling up dashboards.The more indicators there are, the less they are looked at. A useful tool flags deviations, it does not put data on display.

The real prerequisite: data quality

All the use cases above share the same dependency, and this is the honest limit to state upfront. AI does not invent information: it works with what you give it. If the readings are incomplete, poorly timestamped, logged with inconsistent stoppage reasons or scattered across systems that do not talk to each other, the optimisation gets it wrong, and it gets it wrong with confidence.

This is the specific trap of these tools: a model fed dubious data produces a clean, quantified, convincing recommendation, and a false one. The error is all the more dangerous because it is presented without the doubt that would accompany it in a human.

The practical consequence is simple. Before investing in an algorithm, you have to invest in data reliability: shared definitions, disciplined logging of stoppages, consistent timestamping, history consolidated by line and by piece of equipment. It is less impressive than an AI demonstration, but it is what decides whether the rest will hold. This requirement is the same as for predictive AI and equipment life: without clean degradation data, the prediction drifts.

To place these uses within an overall trajectory, from the first pilot to industrialisation, see the industrial AI guide. And for documentary and conversational uses, distinct from optimisation, generative AI in industry answers other needs.