Industrial AI as a Location and Reshoring Factor

The policy challenge

Industrial AI is an industrial-capacity issue

Industrial AI encompasses systems that influence the design, execution or coordination of physical production. Relevant applications include predictive maintenance, machine-vision inspection, process control, production scheduling, demand forecasting, energy optimisation, autonomous or semi-autonomous robotics, engineering copilots, digital twins, supplier-risk analysis and AI-supported logistics.

This definition excludes the assumption that all enterprise AI adoption is equally relevant to industrial competitiveness. A manufacturer using a general-purpose text assistant in its marketing department should not be treated as equivalent to one using secure machine-learning models to increase first-pass yield, reduce production stoppages or reconfigure a supply chain. Public policy therefore requires a more granular measure of adoption than the current binary question of whether a firm uses at least one AI technology.

The distinction matters because the economic barriers are different. General office applications can often be acquired as standardised software subscriptions. Industrial AI generally requires integration with machinery, sensors, production control, enterprise-resource-planning systems and operational procedures. It must perform reliably under demanding conditions, incorporate engineering knowledge, manage safety constraints and remain functional over the long life cycles of industrial assets.

Reshoring should be distinguished from related strategies

Reshoring refers to the return of activities to the firm’s or investor’s home country. Nearshoring relocates production closer to its principal market, including from outside to within the EU. Friend-shoring places production or sourcing in trusted partner economies. Regionalisation develops several geographically distributed production and supplier bases rather than one globally concentrated system.

These strategies can contribute to resilience, but they are not interchangeable. A German manufacturer moving a component operation from East Asia to Poland is nearshoring within the Single Market, not reshoring to Germany. A European producer adding a second supplier in Canada or the United States may reduce geopolitical risk without increasing EU manufacturing. Conversely, a firm can improve European productive capacity without formally reshoring, for example by expanding an existing EU plant, creating a European repair and remanufacturing operation or localising a critical tooling stage.

Industrial policy should therefore assess capability and exposure, not merely count relocation announcements.

The central market failure is one of complementary investment

Industrial AI projects frequently depend on several investments that must occur together:

• machinery must generate accessible data;

• data must be cleaned, structured and governed;

• production and IT systems must be connected securely;

• workers must be trained;

• models must be adapted to the production context;

• legal and cybersecurity requirements must be understood;

• management must be able to calculate and capture the value created.

A subsidy for an AI model alone does not solve these constraints. Nor does an additional research programme necessarily lead to factory deployment. The OECD’s manufacturing assessment finds that pilots often fail to progress because of scalability, maintainability and business-process integration problems. Public intervention is therefore justified primarily where it reduces coordination costs, provides shared infrastructure, corrects financing constraints or produces knowledge and standards that can diffuse across firms and supply chains.

How industrial AI changes the manufacturing-location calculus

A production location is selected on the basis of more than nominal wages. Firms consider total landed cost, productivity, quality, energy, logistics, taxation, regulatory conditions, supplier depth, access to engineering and customers, intellectual-property risks, geopolitical exposure and the cost of disruption. Industrial AI can influence several of these variables simultaneously.

Labour productivity and the high-wage-location penalty

Automation and AI can increase the output generated by each employee, while allowing scarce technicians and engineers to oversee more equipment or more complex processes. AI-supported diagnostics and digital work instructions can also extend the capabilities of less experienced workers.

This can reduce the weight of direct labour costs in location decisions. The effect will be strongest where labour represents a significant but automatable share of total cost, and where the local workforce possesses complementary engineering, maintenance and process knowledge. It will be weaker where labour-intensive tasks remain difficult to automate or where the principal cost disadvantages concern energy, raw materials or scale.

Emerging firm-level evidence is encouraging but should be interpreted carefully. An ECB assessment cites evidence associating AI adoption by EU firms with a 4% productivity increase through capital deepening and no adverse aggregate employment effect in the short term. This is neither a manufacturing-only estimate nor a guaranteed return for individual projects. The ECB also stresses that realised gains depend on digital infrastructure, skills, business dynamism and broader productive capacity.

Quality, yield and compliance

In high-cost economies, competitiveness often depends less on producing the lowest-cost standard item than on manufacturing highly reliable, regulated or technically differentiated products. Computer vision, anomaly detection and process analytics can identify defects earlier, improve traceability and reduce scrap and rework.

This is particularly relevant to pharmaceuticals, medical devices, aerospace, electronics, specialised machinery, automotive components, energy equipment and maritime systems. In these sectors, the economic value of avoiding a failure or maintaining certification may exceed the value of reducing direct labour input.

AI can therefore reinforce locations with strong engineering, testing and regulatory ecosystems. It can make proximity between research, product design, process engineering and manufacturing more valuable, particularly where production feedback improves the next generation of products.

Capital utilisation and operational continuity

European production is capital-intensive. Unplanned downtime reduces the return on expensive machinery and can make smaller domestic production runs uneconomic. Predictive maintenance and AI-assisted condition monitoring can increase equipment availability, improve maintenance scheduling and extend asset life.

This matters for brownfield sites because their competitiveness often depends on improving the productivity of existing capital rather than constructing entirely new facilities. However, legacy machinery may not contain modern sensors or use interoperable communication protocols. The practical policy challenge is therefore not simply to fund predictive-maintenance algorithms, but to enable secure data extraction, instrumentation and integration.

Flexibility, customisation and speed

AI-enabled scheduling, adaptive robotics and digital twins can reduce the cost of switching between products or operating shorter production runs. This supports European business models based on specialised machinery, customised components, industrial services and close customer integration.

Greater flexibility can also increase resilience. A plant that can change inputs, modify recipes, use alternative components or shift production between sites is less vulnerable than one optimised only for a single product at maximum scale. Resilience should therefore be measured partly by reconfigurability, not simply by the geographic origin of output.

Energy and material efficiency

AI can optimise machine settings, production batches, heating and cooling, industrial loads, material flows and internal logistics. These applications can reduce energy peaks, material loss and emissions. The OECD identifies energy-efficient manufacturing, sustainable procurement and circular supply-chain optimisation as important emerging applications.

However, AI cannot compensate for all structural location disadvantages. In basic metals, chemicals, hydrogen derivatives, e-fuels and other energy-intensive production, energy availability and cost remain decisive. AI may improve conversion efficiency and asset utilisation, but it will not make an otherwise uncompetitive energy system competitive. Industrial AI policy must therefore be integrated with energy, grid, infrastructure and decarbonisation policy.

Supply-chain visibility and strategic optionality

AI can improve demand forecasting, inventory allocation, supplier monitoring, logistics planning and the identification of bottlenecks. These capabilities do not necessarily relocate production. They can make an international supply chain more manageable and therefore reduce the incentive to reshore.

At the same time, better visibility can identify previously hidden single points of failure, quantify time-to-recovery and reveal which production stages have high strategic externalities. This enables more selective resilience measures: dual sourcing, reserve capacity, alternative specifications, regional production, repair capability or the localisation of one critical component rather than an entire value chain.

Knowledge retention and demographic resilience

Manufacturing knowledge is often embedded in experienced technicians, engineers and operators rather than formal documents. As workers retire, firms can lose tacit information about machinery, materials and production anomalies. AI-assisted knowledge systems can structure maintenance records, manuals, photographs and operating experience, making expertise more accessible to new employees.

This can support production in regions facing demographic constraints, but it requires careful design. Capturing worker knowledge without participation or fair recognition may undermine trust. AI systems should augment professional judgement rather than create opaque instructions or excessive workplace monitoring.

The double-edged location effect

The overall implication is that industrial AI is a conditional location factor. It can:

• reduce the high-wage penalty of European production;

• increase the value of co-locating engineering and manufacturing;

• improve the viability of smaller, flexible production runs;

• strengthen the operational resilience of European plants.

But it can also:

• reduce the cost of coordinating distant production;

• enable remote supervision of offshore plants;

• standardise processes across international networks;

• improve the productivity of competitors outside Europe.

Industrial policy cannot assume that subsidising AI adoption will automatically result in European investment. It must combine deployment support with a broader location proposition: skills, energy, infrastructure, suppliers, finance, market access and predictable governance.

What the evidence does and does not show about reshoring

Automation can support reshoring, but the relationship is heterogeneous

Research on industrial robots offers the closest long-run empirical analogue, although contemporary industrial AI covers a broader set of technologies. One cross-country study found that an additional robot per 1,000 workers was associated with greater reshoring activity, while also finding that gains accrued primarily to higher-skilled occupations. A later meta-analysis concluded that the average relationship between automation and reshoring was positive, but that results varied materially with methodology.

A separate cross-country working paper reached the opposite conclusion: automation adoption was associated with less reshoring between 2008 and 2019, particularly in relation to ICT and additive manufacturing. Its interpretation is that digital coordination and productivity technologies can reinforce internationally fragmented production rather than reverse it.

These findings are not necessarily contradictory. Automation may cause some labour-intensive tasks to return while enabling other activities to remain distributed. It may increase domestic value added without returning final assembly, or result in a new highly automated plant that employs fewer people than the activity previously located abroad. Sector, firm size, technology, trade exposure and the availability of complementary skills all influence the outcome.

Policy should therefore avoid three assumptions:

First, adoption statistics are not evidence of reshoring. Second, a reshored investment does not necessarily create large numbers of jobs. Third, employment should not be the only measure of value: technical capability, supplier spillovers, tax base, innovation, security of supply and surge capacity may also be relevant.

Widespread reshoring has not materialised

The OECD’s 2026 estimates find little evidence of broad reshoring during 2023–2024. Across 41 economies, the domestic value-added share in gross exports changed only marginally, while the average domestic contribution to manufacturing stabilised rather than rising sharply. The OECD also cautions that an increase in domestic value added may reflect greater domestic-service content rather than the return of physical manufacturing.

European enterprises have instead pursued combinations of digitalisation, supplier changes, inventories and regionalisation. For 2021–2023, Eurostat reports that 32.3% of surveyed enterprises increased the digitalisation of global-value-chain processes, 30.1% concentrated on reliable suppliers, 29.2% undertook nearshoring within the EU and 27.6% increased inventories. These were generally moderate adjustments rather than wholesale reorganisations.

European Commission survey evidence similarly suggests that firms contemplating relocation usually prefer another country to their home market. Among firms choosing or considering relocation, almost two-thirds favoured friend-shoring; in industry, four-fifths of firms considering relocation preferred another country over reshoring. More adjusting firms expected their operational costs to rise than to fall.

Blanket localisation is neither efficient nor reliably resilient

The OECD estimates that broad relocalisation scenarios could reduce global trade by more than 18% and global real GDP by more than 5%. More importantly, the modelling does not find a consistent resilience benefit: economic volatility increased in more than half of the economies examined.

Geographical proximity can reduce some transport and geopolitical risks, but concentration within Europe can create other vulnerabilities, including common energy shocks, cyber incidents, natural disasters, shortages of specialised inputs or common regulatory bottlenecks. A resilient European industrial system should therefore combine:

competitive EU productive capacity;

• geographically diverse and trusted supply;

• interoperable production systems;

• inventories for selected critical inputs;

• repair, recycling and substitution capability;

• rapid access to transport and logistics;

• reliable data and contingency planning.

The relevant policy question is not “How much production can be brought home?” It is “Which capabilities must Europe be able to maintain, reconfigure or scale under adverse conditions, and how can this be achieved at proportionate cost?”

Assessment of the current European policy architecture

The EU has created many of the necessary components for an industrial AI strategy. The policy gap is increasingly one of integration and delivery.

A substantial AI infrastructure is being established

The Apply AI Strategy identifies manufacturing, robotics, transport, energy, engineering and other strategic sectors as priorities. It positions European Digital Innovation Hubs as regional Experience Centres for AI and links them to AI Factories, Testing and Experimentation Facilities and regulatory sandboxes.

As of April 2026, 19 AI Factories and 13 associated antennas were operational. These facilities combine high-performance computing, data and expertise, and the Commission plans further AI-optimised supercomputing capacity. The AI-MATTERS Testing and Experimentation Facility provides manufacturing test infrastructure across eight European locations. The EDIH network provides a regional entry point for SMEs and measures changes in firms’ digital maturity.

This architecture is valuable, but it is not yet equivalent to a factory-deployment system. An SME with incompatible machinery, poor operational data and uncertain cash flow does not primarily need frontier-model training capacity. It needs a sequenced pathway from plant diagnosis and data readiness to testing, finance, workforce preparation and multi-site deployment.

Europe has strengthened its industrial-data framework

The Data Act has applied since September 2025 and improves business users’ access to data generated by connected machinery while facilitating switching between data-processing providers. The Data Union Strategy seeks to link data spaces and AI ecosystems, expand sectoral data access, support synthetic data and protect sensitive non-personal information.

The challenge is implementation at industrial scale. Legal access to machine data does not ensure that the data are complete, semantically consistent or technically usable. Nor does it solve commercial reluctance to share production information. Policy must support common data models, trustworthy intermediaries, contractual templates and architectures that preserve trade secrets and enable controlled access.

Finance is available, but the deployment gap persists

The EIB Group’s TechEU programme plans €70 billion in financing through 2027, with the objective of mobilising €250 billion in total investment. It includes mature corporations and digital-infrastructure projects as well as start-ups and scale-ups.

Yet industrial AI deployment often falls between traditional categories. It may be too operational for research funding, too customised for standard SME digitalisation programmes and too small or uncertain for conventional infrastructure finance. The benefits may also diffuse to suppliers or customers, while the investing firm bears most initial costs.

The Clean Industrial Deal State Aid Framework provides greater flexibility for clean energy, industrial decarbonisation and clean-technology capacity until the end of 2030. It can support industrial AI where AI is demonstrably part of an energy-efficiency, decarbonisation or clean-technology investment. It should not, however, be treated as a general legal basis for all industrial AI aid. Broader deployment will require appropriate use of existing research, innovation, SME, regional and investment-aid rules, subject to project-specific legal assessment.

Industrial and digital policies are insufficiently joined

The Commission’s proposed Industrial Accelerator Act would introduce low-carbon and European-origin considerations into selected procurement and public-support decisions, while simplifying permitting for strategic sectors. As of July 2026, it remains a proposal rather than adopted legislation.

Industrial AI should become an enabling layer across this broader industrial agenda. Low-carbon production, European manufacturing and strategic capacity will not be competitive merely because demand is supported. Factories must also meet demanding productivity, quality, energy and delivery requirements.

The Commission’s 2026 Digital Decade assessment reaches a similar system-level conclusion: Europe has established important foundations, but scale, speed and coordination remain inadequate, with continuing gaps in advanced technology adoption, cybersecurity, digital skills and scale-up capacity. The phase-out of the Recovery and Resilience Facility also creates a risk of investment discontinuity.

Strategic objective and policy design principles

The proposed objective should be:

To make Europe a preferred location for strategically and economically viable advanced manufacturing by lowering the fixed costs and risks of industrial AI deployment, strengthening adaptable productive capacity and addressing verified supply-chain vulnerabilities.

This objective requires six design principles.

First, support productive outcomes rather than technology acquisition. Eligibility should depend on a credible operational problem and measurable value, not the purchase of software labelled as AI.

Second, focus on additionality. Public funds should address investments that would otherwise be delayed, reduced or not undertaken because of uncertainty, externalities or coordination failures. Routine software replacement should not qualify.

Third, distinguish resilience from localisation. Projects should be assessed against concentration, substitutability, time-to-recovery and strategic spillovers. Domestic content alone is an inadequate measure.

Fourth, require interoperability and portability. Public support should not create proprietary data silos or long-term dependence on a single platform, model or cloud provider.

Fifth, combine European scale with regional delivery. Standards, cross-border infrastructure, finance and state-aid coordination belong primarily at EU level. Factory diagnostics, training, cluster relationships and implementation require national and regional institutions.

Sixth, incorporate workers and cybersecurity from the beginning. Human oversight, skills, occupational safety and secure industrial control cannot be retrofitted after deployment.

Policy recommendations

Recommendation 1: Establish a European Industrial AI Deployment Compact

The European Commission and Member States should create a joint deployment framework under the Apply AI Strategy and the Digital Decade governance process. The Compact should not become another stand-alone funding programme. It should align existing resources, institutions and national measures around a common set of industrial outcomes.

By the end of 2027, the Compact should establish a harmonised baseline for core-production AI, distinguishing factory-floor, engineering, energy, maintenance and supply-chain applications from generic administrative tools. On this basis, the EU should set a 2030 objective of at least doubling the proportion of manufacturing SMEs and small mid-caps with one independently verified core-production AI deployment.

The Compact should concentrate on replicable application families:

• machine-vision quality assurance;

• predictive and prescriptive maintenance;

• energy and material optimisation;

• adaptive production planning;

• digital engineering and industrial knowledge systems;

• supplier-risk, inventory and logistics optimisation;

• human–robot collaboration and worker assistance.

Projects should report comparable operational indicators while retaining sector-specific flexibility. The central measure should not be the number of funded pilots, training events or software licences. It should be the number of sustained deployments that improve verified production performance and are subsequently replicated across plants or suppliers.

Recommendation 2: Create a Factory AI Readiness Facility for brownfield industry

National governments, supported by EU cohesion, Digital Europe and investment instruments, should establish a staged facility for firms whose principal constraints precede model deployment.

The first stage should provide a standardised and subsidised Factory AI Readiness Assessment covering:

• operational and enterprise data;

• sensor coverage and machine connectivity;

• OT–IT integration;

• cybersecurity exposure;

• data governance and legal rights;

• workforce capability;

• prospective use cases and business value;

• interoperability and vendor dependence.

The second stage should co-finance enabling investment: sensors, industrial gateways, secure edge computing, data historians, integration between manufacturing-execution and enterprise systems, machine interfaces, connectivity and cybersecurity controls.

The third stage should support application deployment only when readiness conditions and a credible value case are demonstrated.

This staged model would reduce waste by preventing firms from purchasing AI systems that cannot access reliable data or integrate with production. It would also create a consistent pipeline for banks and public investors. Support intensity should be higher for SMEs, first movers in traditional sectors and cross-company supplier consortia, where demonstration and spillover effects are greatest.

All publicly financed architectures should provide data-export functionality, documented interfaces and credible switching arrangements. Proprietary solutions should remain eligible where economically justified, but public support should not finance avoidable lock-in.

Recommendation 3: Turn existing AI institutions into end-to-end deployment corridors

European Digital Innovation Hubs should serve as the regional front door. Testing and Experimentation Facilities should provide real-world validation. AI Factories should supply relevant compute, models and expertise. Regulatory sandboxes and legal helpdesks should support compliance. The EIB, national promotional banks and commercial intermediaries should finance scale-up.

At present, these institutions are connected conceptually but not always operationally. The Commission should require participating networks to publish a common customer journey:

1. plant and use-case diagnosis;

2. data-readiness and cybersecurity assessment;

3. solution design and provider matching;

4. testing under realistic conditions;

5. legal and standards review;

6. financing decision;

7. deployment and worker training;

8. independent performance verification;

9. replication across plants and supply chains.

Each firm should have a lead institution responsible for coordination rather than being referred repeatedly between programmes. Service-level standards should specify response times, cost transparency and responsibility for follow-through.

EDIH performance evaluation should give greater weight to verified deployment, productivity and replication outcomes. Workshops and maturity assessments remain useful, but they should be treated as intermediate outputs rather than policy success.

Cross-border specialisation is essential. Not every region requires identical expertise. A manufacturer should be able to access a specialised European centre for industrial vision, robotics, maritime systems, energy equipment or process manufacturing even when it is located in another Member State.

Recommendation 4: Build interoperable industrial data commons

Industrial data policy should move from creating abstract data spaces to enabling operationally useful data exchange.

The EU should finance sector-specific reference architectures, semantic models and standard interfaces for production data. Priority should be given to information with broad productivity and resilience value: equipment condition, component quality, energy consumption, materials, maintenance events, supplier performance and product traceability.

Participation must remain commercially credible. Firms will not contribute sensitive data to systems that expose intellectual property, weaken bargaining power or create uncontrolled access. Appropriate models include federated learning, data trustees, secure processing environments, synthetic data, confidential computing and standardised access contracts.

The Data Act’s access rights should be translated into practical industrial guidance. SMEs require assistance in determining which machine data are available, on what terms they may be shared with an independent maintenance or AI provider, and how trade secrets can be protected.

Public procurement and grant agreements should require beneficiaries to maintain machine-readable records, documented data lineage and interoperable interfaces. Where public funds support a shared model or dataset, access terms should encourage diffusion while protecting legitimate commercial interests.

European industrial-domain models merit support where they solve demonstrable market failures for example, multilingual engineering assistants, models for machine signals or models capable of using technical drawings and maintenance histories. The objective should not be European origin as an end in itself. It should be access to secure, controllable and competitive options that can operate in industrial environments, including edge and hybrid configurations.

Recommendation 5: Establish a risk-sharing finance window for industrial deployment

A dedicated industrial AI deployment window should be created within relevant EIB, InvestEU and national promotional-bank instruments. It should complement, rather than duplicate, TechEU’s innovation finance.

The window should contain three financing tiers.

Readiness finance should fund assessments, data preparation, cybersecurity and systems integration. These are often small investments with substantial transaction costs and should use standardised vouchers, grants or highly simplified loans.

First-deployment finance should support technically validated but commercially uncertain projects. Instruments could include contingent loans, guarantees, milestone-based grants and risk-sharing contracts. Public support should be conditional on co-investment, independently verified implementation and a plan for operational ownership after the pilot period.

Scale-out finance should provide loans or guarantees for replicating proven applications across plants, suppliers or regions. Once a use case has demonstrated value, support should shift from grants to repayable finance.

Payment schedules should be tied to project milestones and operational outcomes. Where agreed performance or dissemination obligations are not met without adequate justification, clawbacks or reduced support should apply. At the same time, programmes must accommodate the uncertainty inherent in experimentation; they should not penalise well-managed projects solely because a technically plausible approach fails.

Member States may also consider accelerated depreciation or investment tax allowances for qualifying industrial digital assets. Eligibility should require interoperability, cybersecurity and workforce preparation. Tax measures should not be used as an indiscriminate subsidy for general software expenditure.

To prevent fiscal capacity from determining industrial geography, a substantial share of risk-bearing support should be financed or guaranteed at EU level. Common evaluation criteria and aid-intensity guardrails would reduce subsidy competition between Member States.

Recommendation 6: Link AI policy to selective resilience and location decisions

Public authorities should develop a common methodology for identifying strategically consequential production nodes. Assessment criteria should include:

• import and supplier concentration;

• availability of substitutes;

• time required to restore supply;

• dependence of downstream industries;

• defence, health, energy or infrastructure relevance;

• knowledge and supplier spillovers;

• environmental and resource implications;

• realistic European production economics;

• availability of capacity in trusted partner countries.

Support should focus on the smallest economically efficient intervention capable of reducing systemic exposure. This may involve localising a critical component, tooling capability, maintenance operation, data infrastructure, recycling process or reserve production line rather than an entire final product.

A time-limited resilience premium could be added to industrial AI support where an investment demonstrably creates European capacity in a high-risk node. It should be conditional on competitive productivity, contingency planning and continued operational capability. It should not compensate indefinitely for a structurally non-viable plant.

Resilience scoring should reward firms that:

• establish qualified alternative suppliers;

• design products for component substitution;

• maintain transferable production recipes and tooling;

• create surge or restart capacity;

• improve repair and remanufacturing capability;

• use interoperable systems across several plants;

• reduce critical material and energy dependencies.

This approach would be particularly relevant to energy-system equipment, semiconductors and electronics, medical and pharmaceutical production, specialised machinery, defence and aerospace, maritime equipment, port and logistics infrastructure, and selected clean-technology components. For energy-intensive basic materials, location support must be integrated with energy and decarbonisation policy; AI alone cannot determine competitiveness.

Public procurement can create lead markets for secure, resource-efficient and resilient production. Tender criteria should emphasise performance, lifecycle cost, cybersecurity, traceability, delivery continuity and environmental impact. European-preference provisions, where used, should be proportionate, transparent, reciprocal and compatible with international obligations. They should not shelter poor performance or exclude trusted partners without a risk-based justification.

Recommendation 7: Establish an industrial AI skills and workforce compact

Industrial AI deployment requires hybrid capability. Data scientists without process knowledge may develop technically sophisticated but operationally irrelevant systems. Production engineers without data expertise may be unable to specify or govern AI projects. Operators and technicians need sufficient understanding to challenge outputs, detect failure modes and use systems safely.

Member States should support modular programmes combining:

• process engineering and data analytics;

• industrial automation and machine learning;

• operational technology and cybersecurity;

• AI project economics and procurement;

• data governance and regulatory compliance;

• human factors and safety;

• sector-specific applications.

Vocational education and apprenticeships should be adapted alongside university programmes. A particular priority should be the creation of industrial AI translators: professionals capable of converting production problems into data and system requirements and of coordinating engineering, IT, management and workers.

Publicly supported projects should include paid training and a workforce-impact plan. Workers and, where applicable, works councils should participate in use-case design, performance monitoring and decisions about workplace data. This is necessary both for legitimacy and for technical quality: operators often hold the knowledge required to interpret anomalies and avoid unsafe automation.

Support should be conditional on safeguards against disproportionate worker surveillance and opaque automated personnel decisions. Productivity support should not become a mechanism for intensifying work without corresponding investment in safety, capability and job quality.

Reshoring policies should also avoid promising large numbers of routine manufacturing jobs where production will be highly automated. The more credible employment proposition is one of skilled production, maintenance, software, engineering, supplier services and industrial innovation. Regional policy must prepare workers for this occupational structure.

Recommendation 8: Provide coordinated regulatory and cybersecurity implementation

Legal certainty is an adoption condition. Eurostat’s finding that more than half of non-adopting firms that had considered AI cited uncertainty over legal consequences indicates that implementation support has economic value.

The EU should publish a manufacturing-specific AI implementation guide covering common use cases, responsibilities across the value chain, documentation expectations, human oversight, product-safety interaction and examples of systems likely to fall within or outside high-risk categories.

Following the June 2026 AI Omnibus, the application dates for stand-alone high-risk systems and high-risk systems embedded in products have been moved to December 2027 and August 2028 respectively. Machinery has a specific interaction with sectoral safety legislation, and the Commission is expected to provide further guidance. The additional time should be used to build practical compliance capacity rather than to postpone preparation.

Regulatory sandboxes should prioritise cross-border and industrial cases. A firm testing the same system in several Member States should not receive materially different interpretations without a substantive legal reason. National authorities, the AI Office, market-surveillance authorities and sectoral regulators should coordinate templates and technical positions.

Cybersecurity must be treated as part of production engineering. Connecting legacy machinery, external models, cloud services and suppliers expands the attack surface of operational technology. Publicly supported projects should therefore require:

• asset and dependency mapping;

• network segmentation;

• secure identity and access management;

• controlled software and model updates;

• logging and anomaly detection;

• tested manual fallback and recovery procedures;

• supplier-risk assessment;

• lifecycle patch and vulnerability management.

The Cyber Resilience Act establishes lifecycle cybersecurity duties for products with digital elements, while the EU cybersecurity framework is moving towards more coordinated and risk-based supply-chain security. Industrial AI schemes should align with these requirements and avoid creating parallel compliance systems.

Division of responsibilities

European Union

The EU should provide the common definition of core-production AI, cross-border infrastructure, data and interoperability standards, shared testing resources, financing guarantees, state-aid coordination, regulatory guidance, strategic dependency analysis and a European performance scoreboard.

It should also ensure that poorer or smaller Member States are not excluded by limited fiscal or institutional capacity. Cohesion resources and EU-level guarantees should maintain a minimum level of deployment support across industrial regions.

European policy should preserve cooperation with trusted partners. Technological sovereignty should mean the capacity to choose, govern and substitute critical technologies—not the mandatory localisation of every supplier. Transatlantic cooperation remains important for semiconductors, cloud and edge infrastructure, cybersecurity, standards, research and strategic supply chains.

Member States

National governments should organise plant-level programmes, co-finance readiness investment, align tax and depreciation provisions, reform vocational education, mobilise national promotional banks, integrate energy and permitting policy, and use public procurement strategically.

National authorities should identify regional industrial strengths rather than reproduce the same sectoral priorities everywhere. A country with a strong maritime, machinery, chemicals, pharmaceutical or automotive base requires a deployment strategy tailored to its assets and constraints.

Regions and industrial clusters

Regional agencies, chambers, technology centres and clusters should undertake outreach, aggregate SME demand and organise supplier consortia. They are best placed to identify firms that are operationally capable but lack internal AI expertise.

Cluster-level projects are especially relevant where the value of AI depends on supply-chain data. A large manufacturer should be encouraged to deploy interoperable tools with its SME suppliers rather than improve only its own visibility while transferring data and compliance burdens downstream.

Firms and social partners

Beneficiary firms should co-invest, establish internal ownership, maintain data and cybersecurity governance, involve workers and disclose agreed performance information in anonymised or aggregated form. Public support should purchase measurable public value, including knowledge diffusion and supplier development.

Implementation roadmap

Phase I: Build the framework, July 2026–June 2027

The Commission and Member States should agree the Industrial AI Deployment Compact and incorporate its priorities into the forthcoming revisions of national Digital Decade roadmaps. Existing programmes should be mapped against the complete deployment journey to identify gaps, duplication and inaccessible interfaces.

A common Factory AI Readiness Assessment and core-production AI taxonomy should be developed. Several industrial regions should pilot end-to-end deployment corridors using existing EDIHs, AI-MATTERS facilities, AI Factories and financing partners.

The Commission should issue initial manufacturing guidance on the Data Act and AI regulation and establish a joint industrial AI and cybersecurity implementation group.

Phase II: Scale deployment, July 2027–December 2029

The readiness facility and risk-sharing finance window should become available across Member States. Public procurement pilots should be launched in infrastructure-intensive sectors such as energy, transport, ports, healthcare and defence.

Sectoral data architectures should be deployed in selected value chains. Cross-border supplier consortia should receive support where shared adoption can reduce systemic risk or improve the competitiveness of an entire European ecosystem.

The first independent evaluations should compare supported firms with credible control groups. Instruments with weak additionality or low replication should be redesigned or terminated.

Phase III: Consolidate productive capability, 2030 onwards

Proven measures should be incorporated into the next Multiannual Financial Framework and future European competitiveness instruments. Support should move progressively from grants to guarantees, loans and procurement as technologies and applications mature.

Strategic-capability assessments should be updated regularly. Resilience premiums should expire when a market becomes competitive, a dependency is resolved or a supported capability no longer has strategic relevance.

Performance measurement and accountability

A credible programme requires measurement beyond expenditure and participation.

Adoption and depth

Authorities should measure the share of manufacturers using AI in core production, the number of production lines and sites covered, the duration of use after the pilot stage and the proportion of SMEs and mid-caps adopting.

Operational performance

Projects should select relevant indicators such as equipment availability, unplanned downtime, first-pass yield, scrap, energy use, material use, lead time, production-changeover time, inventory, forecast accuracy and maintenance cost.

No universal minimum improvement should be imposed across all applications. Baselines, measurement methods and verification should nevertheless be standardised.

Resilience

Relevant indicators include supplier concentration, time-to-recovery, availability of qualified substitutes, share of critical inputs with dual sourcing, restart time, surge capacity, interoperability across plants and dependence on single technology providers.

Diffusion

Public value depends on whether knowledge moves beyond the first beneficiary. Metrics should include replication across sites, adoption by suppliers, reuse of common architectures, cross-border deployment and the number of firms using shared testing or data assets.

Workforce outcomes

Reporting should cover training participation, occupational changes, safety, employee involvement, job quality, recruitment of technical roles and the distribution of productivity gains. Headcount alone is an inadequate measure.

Competition and technological control

Evaluation should examine switching costs, data portability, concentration among solution providers, dependence on non-substitutable cloud or model services and the availability of European or trusted alternatives. A simple count of European vendors would not capture actual strategic control.

Public additionality

Authorities should estimate the proportion of investment advanced or expanded by public support, private finance mobilised, cost per sustained deployment and spillovers to other firms. Independent evaluation and publication of aggregated results should be mandatory.

Principal risks and safeguards

Subsidy competition and fragmentation

National industrial AI schemes could favour Member States with greater fiscal capacity, distort investment and create incompatible national systems. EU-level guarantees, common criteria and shared infrastructure are required to preserve the Single Market.

Public support for commercially weak investments

Political interest in reshoring can result in support for production that remains structurally uncompetitive. Time limits, co-investment, productivity benchmarks, competitive selection and clawbacks should prevent indefinite protection.

Vendor lock-in and data extraction

A publicly subsidised deployment may strengthen a dominant platform while leaving the manufacturer unable to access its own data or switch providers. Interoperability, data portability and transparent contractual terms should be eligibility conditions.

Cybersecurity and operational safety

More connected factories may become more vulnerable. Industrial AI can also produce erroneous recommendations or unsafe control actions. Risk-based assurance, human override, secure architecture and tested fallback procedures are essential.

Unequal labour-market outcomes

Automation-enabled reshoring may increase demand for engineers and technicians without restoring routine employment. Evidence from robotics suggests that benefits can be skill-biased. Training, mobility support, worker participation and regional adjustment policies should accompany investment support.

Excessive localisation and trade conflict

Broad domestic-content requirements can raise costs, reduce competition and invite retaliation. Preference should be used only where proportional to a defined resilience or reciprocity problem. Trusted international partnerships remain an essential component of European security.

Energy and environmental rebound

Efficiency improvements may be offset by increased production or computing demand. Supported projects should report both unit efficiency and absolute energy implications where material, and computing architecture should be proportionate to the industrial task.

Conclusion

Industrial AI can become an important European location factor, but only when embedded in a competitive industrial system. It can reduce the labour-cost disadvantage of European production, improve the utilisation of capital and energy, strengthen quality and flexibility, preserve industrial knowledge and create more adaptable supply chains. These effects are particularly relevant to complex, specialised, time-sensitive and quality-critical manufacturing.

The technology does not, however, constitute an automatic reshoring mechanism. AI may strengthen domestic plants, enable nearshoring or make distant production easier to coordinate. Recent evidence shows adjustment and regionalisation rather than a general return of manufacturing. Broad localisation would impose substantial economic costs and would not necessarily improve resilience.

National and European industrial policy should therefore avoid presenting reshoring as an end in itself. The strategic objective should be competitive and adaptable European productive capability. Where industrial AI makes an economically viable European investment possible, policy should accelerate it. Where a verified strategic dependency requires reserve, substitute or local capacity, policy may provide proportionate additional support. Where neither condition is present, public authorities should not subsidise relocation for symbolic reasons.

Europe already possesses many of the necessary institutions: industrial companies, engineering expertise, AI research, high-performance computing, testing facilities, regional innovation hubs, data legislation and public financing capacity. The unresolved task is to join these assets into a deployment system that reaches the factory floor.

The proposed European Industrial AI Deployment Compact would provide that connection. It would treat data, connectivity, skills, cybersecurity, finance and operational integration as one industrial-policy challenge. It would reward measurable productivity and resilience rather than nominal adoption. It would strengthen European capacity without equating sovereignty with autarky.

The guiding principle should be clear: Europe should make itself the economically superior location for strategically relevant advanced manufacturing, not attempt to make the European location administratively compulsory.