AI for manufacturing and industry
Intellectual property, designs and supply-chain data.
Research
Context
For manufacturers, designs, process knowledge and supplier data are the business. AI pays off in engineering and operations when that knowledge stays inside the company.
A manufacturer's edge is its designs, recipes and process knowledge. Most AI services need that data to leave the company. Lindstead helps decide which AI stays in-house and runs it on infrastructure the company controls.
Case study (anonymised): Keeping customer drawings in-house at a machinery manufacturer A European machinery manufacturer, around 2,000 staff.
Need a first decision? The AI Sovereignty Scan covers readiness, regulation and hosting in ten working days.
Factory AI use ranges from 12% to 41% by sub-sector
In 2025, 17.3% of EU manufacturers with ten or more employees used AI, up from 6.8% in 2023. The average across all sectors is 20.0%.
In 2025, 41% of EU pharmaceutical manufacturers used AI, about three times the rate in metals, food and textiles.
Share of manufacturing enterprises using at least one AI technology, EU-27, 2025, %
¹ Motor vehicles, trailers, semi-trailers and other transport equipment (NACE C29 and C30). ² Rubber and plastic products and other non-metallic mineral products (NACE C22 and C23). Enterprises with 10 or more persons employed, EU-27, reference year 2025. AI use means at least one of eight technologies: text mining, speech recognition, language or code generation, image, video or audio generation, image recognition, machine learning for data analysis, AI-based process automation, and autonomous movement of machines. Selected manufacturing sub-sectors.
Source: Eurostat, Artificial intelligence by NACE Rev. 2 activity (isoc_eb_ain2), database updated 15 June 2026
View data table
| Category | Value | Group |
|---|---|---|
| Pharmaceuticals | 41.3% | 1 |
| Computers and electronics | 37.4% | 1 |
| Chemicals | 28.5% | 1 |
| Electrical equipment | 25.9% | 1 |
| Machinery and equipment | 25.5% | 1 |
| Motor vehicles and transport¹ | 23.9% | 1 |
| Rubber, plastics, minerals² | 16.8% | 2 |
| Basic and fabricated metals | 14.3% | 2 |
| Food, drinks and tobacco | 12.5% | 2 |
| Textiles, clothing, leather | 12.0% | 2 |
Group 1: Above the all-sector EU average (20.0%). Group 2: Below the all-sector EU average. Unit: %.
Challenges
Process and high-tech industries moved first. Of manufacturers using AI, 30.4% apply it to marketing or sales and only 20.2% to production. Most has yet to reach the shop floor, where quality and maintenance gains sit.
- Know-how leaving the company Drawings, recipes and process settings stay trade secrets only if the holder takes reasonable steps to protect them. Pasting them into public AI tools weakens that claim.
- Customer and OEM obligations Suppliers hold customer drawings and prototype data under confidentiality agreements and TISAX assessments. Sending them to an outside model provider can breach contract.
- Machine data and Data Act Since 12 September 2025 the Data Act lets a manufacturer use customer machine data only under a contract with the user. That limits AI training on field data.
- Fragmented engineering data Specifications, test reports and quality records sit across PLM, ERP, MES and file shares. An AI system that cannot reach them all answers confidently from half the facts.
- Overlapping product regulation AI in machine safety functions falls under the Machinery Regulation from 20 January 2027. Cyber Resilience Act reporting has applied since 11 September 2026, and NIS2 covers many manufacturers. Separate assessments leave gaps.
Half of stalled factory AI plans now cite data protection
Eurostat asks manufacturers that considered AI but did not use it why. Missing expertise still tops the list. Legal and data protection concerns grow fastest.
Among EU manufacturers that considered AI but held back, data protection concerns rose from 42% to 49% in a year. Legal uncertainty reached 51%.
Reasons for not using AI, share of manufacturing enterprises that had considered using AI, EU-27, 2024 and 2025, %
¹ Lack of clarity about the legal consequences. ² Concerns regarding violation of data protection and privacy. ³ Difficulties with the availability or quality of the necessary data. Base: manufacturing enterprises (NACE C) with 10 or more persons employed, EU-27, that did not use AI but had considered using it. Multiple reasons possible. Incompatible systems: incompatibility with existing equipment, software or systems.
Source: Eurostat, Artificial intelligence by NACE Rev. 2 activity (isoc_eb_ain2), database updated 15 June 2026
View data table
| Category | Series 1 | Series 2 |
|---|---|---|
| Lack of expertise | 73.5% | 71.1% |
| Legal uncertainty¹ | 47.9% | 51.1% |
| Data protection² | 42.3% | 49% |
| Data quality³ | 48.2% | 45.3% |
| Incompatible systems | 46.4% | 44.7% |
| Cost too high | 38.3% | 39.8% |
| Ethical concerns | 20% | 21.4% |
Series 1: 2024. Series 2: 2025. Unit: %
Key regulation
Every capability barrier eased slightly from 2024 to 2025. Legal uncertainty rose 3.2 points, data protection concerns 6.7. Where models run and who sees design data now matter as much as the use case.
| Framework | Relevance |
|---|---|
| EU Data Act | Rules on access to and sharing of industrial and product data, applicable since 12 September 2025. |
| EU AI Act | AI safety functions in machinery follow the Machinery Regulation; AI Act requirements arrive through delegated acts due by 2 August 2028. |
| NIS2 | Manufacturers of machinery, electronics, vehicles and medical devices can fall within scope. |
| Trade secrets | Protection depends on reasonable measures to keep know-how confidential. |
How Lindstead helps
- AI strategy Workload map by IP sensitivity Lindstead sorts each use case by the IP and customer data it touches. The board gets a verdict per workload: in-house, European cloud or external API.
- Model selection Models tested on technical documents Lindstead tests open-weight models on the organisation's own specifications and test reports, and checks each licence. The organisation gets a shortlist with results per task.
- Regulation and risk Data Act, NIS2, CRA and machinery map Lindstead maps each AI system's data flows against the Data Act, NIS2, the Cyber Resilience Act, the Machinery Regulation and supplier contracts. The map shows dates, owners and gaps.
- Infrastructure and cost Cost model for plant and group Lindstead sizes the GPU capacity needed and compares owned servers, European cloud GPUs and API spend over three years, with every assumption shown.
- Deployment Engineering assistant on own servers Lindstead deploys a first use case, such as specification search, on servers the organisation controls. Access follows project and customer, with audit logs, before handover to IT.
- Model stewardship Model and regulation watch Each month Lindstead checks new models, licences and rules such as the Machinery Regulation delegated acts. The board gets a quarterly note on upgrades and deadlines.
Case study (anonymised)
Keeping customer drawings in-house at a machinery manufacturer
A European machinery manufacturer, around 2,000 staff
Engineers wanted AI for specifications and test reports, and some already used public chatbots. Automotive customers required assessed handling of their drawings.
The board wanted one ruling on where AI could run and on what data.
What Lindstead did
- Inventory and classification Lindstead logged all AI use, including public chatbots, and sorted each case by the data it touched. Interviews with engineering, quality, IT and legal identified the rules that applied.
- Models and cost The team tested open-weight models on the company's own documents in three languages. It compared owned servers, a European cloud and APIs over three years.
- Pilot engineering assistant Lindstead deployed search and drafting across specifications and test reports on company servers, with access by project and customer. Engineers used it on live work against agreed quality criteria.
- Board decision paper The team combined model results, costs and the regulatory map into one recommendation. It covered Data Act and Machinery Regulation questions on machine learning in products.
What the client received
- Use-case register by IP sensitivity
- Model tests on own technical documents
- Three-year cost model with break-even volume
- Map of Data Act, NIS2, CRA and machinery rules
- Pilot engineering assistant on own servers
The decision
The board set which AI workloads could use external services and one policy for customer and OEM data.
It also ruled on extending the assistant to other sites.
EU robot density is 75% above the world average
Robot density counts operational industrial robots per 10,000 manufacturing employees. On this measure, European industry is among the world's most automated.
Germany ranks third worldwide with 449 robots per 10,000 manufacturing employees. The EU-27 averages 231, 75% above the global average of 132.
Operational industrial robots per 10,000 employees in manufacturing, 2024, robots per 10,000 employees
¹ Based on updated labour market data from China's National Bureau of Statistics; China ranks 22nd worldwide. ² Average across all economies covered by World Robotics. Selected economies, reference year 2024. Singapore (818) ranks second worldwide and is omitted as a small economy with few manufacturing employees.
Source: International Federation of Robotics, World Robotics 2025 robot density release, 8 April 2026; installations: World Robotics 2026 release, 24 September 2026
View data table
| Category | Value | Group |
|---|---|---|
| Republic of Korea | 1220 | 2 |
| Germany | 449 | 1 |
| Japan | 446 | 2 |
| Sweden | 377 | 1 |
| Denmark | 329 | 1 |
| Slovenia | 315 | 1 |
| United States | 307 | 2 |
| EU-27 | 231 | 1 |
| China¹ | 166 | 2 |
| Global average² | 132 | 2 |
Group 1: Europe. Group 2: Asia, Americas and global average. Unit: robots per 10,000 employees.
Opportunities
The robots are largely in place. The gap is the data and AI layer on top. Investment is slowing: EU installations fell 11% to 60,500 units in 2025, while Asia and the Americas each grew 14%.
- Engineering knowledge search Engineers search specifications, standards, change requests and test reports for earlier solutions. That core design know-how belongs on the company's own infrastructure.
- Quality and deviation analysis AI finds recurring causes in non-conformance reports, complaints and audit findings. These records name customers, parts and suppliers, so in-house processing keeps contracts simple.
- Supplier documentation review AI extracts data from supplier certificates, material declarations and contracts, and flags gaps. Confidentiality clauses often limit third-party processing of that data.
- Service and maintenance support Technicians query manuals, service history and machine data during repairs. Data Act contracts with customers are easier to honour on the manufacturer's own systems.
Frequently asked questions
-
The main frameworks are:
- EU Data Act: Rules on access to and sharing of industrial and product data, applicable since 12 September 2025.
- EU AI Act: AI safety functions in machinery follow the Machinery Regulation; AI Act requirements arrive through delegated acts due by 2 August 2028.
- NIS2: Manufacturers of machinery, electronics, vehicles and medical devices can fall within scope.
- Trade secrets: Protection depends on reasonable measures to keep know-how confidential.
-
Organisations in this sector most often face three issues:
- Know-how leaving the company: Drawings, recipes and process settings stay trade secrets only if the holder takes reasonable steps to protect them. Pasting them into public AI tools weakens that claim.
- Customer and OEM obligations: Suppliers hold customer drawings and prototype data under confidentiality agreements and TISAX assessments. Sending them to an outside model provider can breach contract.
- Machine data and Data Act: Since 12 September 2025 the Data Act lets a manufacturer use customer machine data only under a contract with the user. That limits AI training on field data.
- Fragmented engineering data: Specifications, test reports and quality records sit across PLM, ERP, MES and file shares. An AI system that cannot reach them all answers confidently from half the facts.
- Overlapping product regulation: AI in machine safety functions falls under the Machinery Regulation from 20 January 2027. Cyber Resilience Act reporting has applied since 11 September 2026, and NIS2 covers many manufacturers. Separate assessments leave gaps.
-
The use cases with the clearest value are:
- Engineering knowledge search: Engineers search specifications, standards, change requests and test reports for earlier solutions. That core design know-how belongs on the company's own infrastructure.
- Quality and deviation analysis: AI finds recurring causes in non-conformance reports, complaints and audit findings. These records name customers, parts and suppliers, so in-house processing keeps contracts simple.
- Supplier documentation review: AI extracts data from supplier certificates, material declarations and contracts, and flags gaps. Confidentiality clauses often limit third-party processing of that data.
- Service and maintenance support: Technicians query manuals, service history and machine data during repairs. Data Act contracts with customers are easier to honour on the manufacturer's own systems.
-
Yes. Open-weight models can run on-premise or in a European private cloud, which keeps data and model access under the organisation's control. Whether that is cheaper than an external API depends on sustained volume, and whether it is required depends on data sensitivity and regulation. Lindstead's break-even analysis shows the volume at which own infrastructure pays off.
-
The Digital Omnibus on AI (Regulation (EU) 2026/1744) placed the Machinery Regulation in Section B of Annex I to the AI Act. The Commission must adopt delegated acts on AI safety components by 2 August 2028. From 20 January 2027 the Machinery Regulation already requires notified body assessment for safety components with self-evolving machine learning.
-
Only under a contract. Article 4(13) of the Data Act lets a data holder use readily available non-personal product data only by contract with the user, and never against the user's commercial position. Personal data in machine logs also falls under the GDPR. Lindstead advises checking customer terms first.
-
It helps, but not on its own. The Trade Secrets Directive protects information only if its holder takes reasonable steps to keep it secret. Own infrastructure with access control and logging is one such step. It also avoids the US CLOUD Act, which can force US providers to hand over data wherever it is stored.
Related research
- Open-weight models for internal engineering data
- Self-hosting or API: where is the break-even point?
Sources
- EUR-Lex, Regulation (EU) 2023/2854 (Data Act), 13 December 2023
- Data Act, Article 4: rights and obligations of users and data holders
- Data Act, Article 50: entry into force and application
- EUR-Lex, Regulation (EU) 2026/1744 (Digital Omnibus on AI), 8 July 2026
- artificialintelligenceact.eu, Digital Omnibus on AI: amended application dates and Machinery Regulation, 2026
- European Commission, Machinery: Regulation (EU) 2023/1230, accessed September 2026
- Unchained Robotics, EU Machinery Regulation 2027: what applies to robot cells
- European Commission, Cyber Resilience Act: reporting obligations, September 2026
- European Commission, Cyber Resilience Act, accessed September 2026
- EUR-Lex, Directive (EU) 2022/2555 (NIS2), Annex II manufacturing and chemicals, 14 December 2022
- reuschlaw and VDMA, Working paper on Annex II No. 3 NIS2 Directive and REACH, April 2024
- EUR-Lex, Directive (EU) 2016/943 on the protection of trade secrets, 8 June 2016
- ENX Association, TISAX: information security assessment for the automotive industry
- Cornell LII, 18 U.S. Code 2713 (CLOUD Act): required preservation and disclosure
Discuss the priorities in manufacturing and industry.
A 30-minute meeting covers the use case, sector requirements and next step.