OECD manufacturing evidence: process optimisation is ahead of proof

External-source research synthesis, not an OfferSignal client result. Based on the retained October 2026 research and source-specific dates.

A refinery, metals plant or fuel distributor can have years of sensor, maintenance, laboratory, supplier and transport records without having a safe operational decision system. The 2026 OECD evidence is useful because it separates adoption from impact: it shows where European manufacturers say they use artificial intelligence, but does not turn that adoption into proof of higher throughput, lower energy use or higher revenue.

What the OECD evidence actually shows

Manufacturing represented 16% of European Union value added in 2023 and employed more than 30 million people. In 2024, 2.8% of all EU manufacturing enterprises used artificial intelligence for production-process optimisation. Petroleum enterprises were a notable outlier at 10.7% for that use, while basic metals recorded 9% overall AI adoption, compared with 26% in pharmaceuticals and 25% in electronics. These are enterprise survey statistics, not plant-level results or causal outcomes. The survey covers enterprises with at least 10 employees, not individual mines, refineries, furnaces, tonnes or shipments. 1 2

Among manufacturing enterprises already using AI, 26% reported process optimisation and 9.4% logistics use. The evidence base combines literature review with semi-structured interviews conducted from December 2024 to April 2025. Only 45 of roughly 250 prospective participants agreed to interviews, making the qualitative sample voluntary and non-representative. Provider claims and early-adopter enthusiasm may therefore be over-represented. 1 3

The practical mechanism is decision support

For a continuous mineral or fuel operation, the strongest application is a bounded workflow rather than autonomous control. Historian, supervisory-control, laboratory, maintenance, enterprise-resource-planning and logistics data can feed anomaly detection, fault diagnosis, remaining-useful-life estimates or constrained scheduling. An operator or process engineer then reviews the recommendation before changing a set-point, booking an inspection, ordering a part or scheduling a shutdown.

The same pattern applies to transport. Demand, inventory, telematics, traffic, fuel prices, route availability and delivery windows can inform dispatch, multi-stop routing or modal coordination. The mobility evidence identifies route and schedule optimisation and fuel or energy optimisation as relevant mechanisms, but also points to real-time data, legacy integration, fragmented actors and interoperability as barriers. Its freight shares—67.8% maritime, 24.9% road, 5.5% rail, 1.6% inland waterways and 0.2% air by EU tonne-kilometres in 2022—are context, not AI outcomes. 4

For sustainability, route objectives can include fuel and greenhouse-gas measures, and supplier ranking can include carbon data, certificates and compliance history. The OECD’s Elio.Earth example describes extracting emissions and transport information from documents and operational datasets. That is a provider/interview example of capability, not an audited reduction in fuel, energy or carbon.

What has been reported—and what has not

The chapter records heterogeneous provider or interview claims. Sensorminds reported a 50% improvement in detection rates versus manual inspection in fabric printing. Another machinery and AI provider reported clients with up to 40% operational-efficiency improvement from monitoring and defect detection. A German, Austrian and Swiss survey of 247 industrial enterprises estimated up to 10.7% potential profitability uplift from consistent generative AI use, while 0.74% of that potential had been realised when cited. None is an independently verified mineral or fuel result; the samples, baselines, periods and denominators are not disclosed. 1

The evidence also does not demonstrate an end-to-end custom-work or procurement system: no verified chain joins archived drawings, equipment capability, engineering constraints, current capacity, quote assumptions, approval and acceptance. A similar document or supplier record is evidence for review, not permission to issue a binding offer.

An illustrative first pilot

An illustrative fuel terminal could start with maintenance alerts or delivery scheduling. Preserve asset identity, sensor units, timestamps, material or product grade, approved operating limits, source and revision. Let the system propose an alert or route; let the maintenance planner, process engineer or dispatcher remain authoritative. Measure false positives, missed failures, review time, schedule adherence, downtime, fuel use and service performance against a defined baseline. Record the recommendation, sources, approval, override reason and realised result.

This discipline addresses the barriers the evidence records: lack of expertise above 7.5%, data availability or quality at 5.0%, incompatible systems at 4.8%, legal consequences at 4.9%, privacy at 4.4% and high cost at 3.9% among reported reasons for non-use. Digital twins and broad supply-chain optimisation remain conditional future signals; the OECD notes that pilots often fail to scale because of maintainability, integration and process misalignment. Secure data exchange, skills, compute and clearer compliance remain prerequisites, not completed outcomes. 1 5

Sources and next reading

Explore OfferSignal · Focused Business & AI Workshop · All research

References