Sovereign Agentic AI in European Manufacturing: Architecting Autonomous Engineering Workflows under the EU AI Act

Executive Summary

The European industrial sector is navigating a profound technological inflection point in 2026. The integration of artificial intelligence into engineering, Computer-Aided Design (CAD), and Product Lifecycle Management (PLM) has transitioned from the deployment of passive, generative assistants to the implementation of autonomous, goal-directed Agentic AI systems.1 For enterprises managing complex manufacturing value chains, this evolution offers unprecedented opportunities to automate intricate workflows, optimize supply chain data, and accelerate product development.3 However, the deployment of such advanced computational intelligence within the European Union is heavily gated by two absolute imperatives: rigorous compliance with the EU AI Act and the unyielding demand for digital sovereignty.5

This technical whitepaper provides an exhaustive architectural blueprint for designing, deploying, and governing sovereign multi-agent systems within demanding B2B engineering environments. Developed through the lens of advanced system integration, custom software development, and specialized Siemens NX Open API architecture, this report addresses the critical technical requirements for operationalizing AI in high-risk industrial contexts. By examining the integration of Large Language Models (LLMs) with deterministic engineering platforms via the Model Context Protocol (MCP), and mapping these integrations against the strict mandates of the EU AI Act and the Gaia-X Trust Framework, this analysis establishes a comprehensive roadmap for Chief Technology Officers (CTOs), enterprise architects, and engineering leaders.8 The objective is to provide a robust framework for achieving computational sovereignty, mitigating non-deterministic AI outputs, and driving compliance-driven innovation in the era of autonomous industrial digital transformation.

1. The Paradigm Shift: From Generative Intelligence to Agentic Engineering Workflows

1.1 Deconstructing the Agentic Engineering Loop

Generative AI, characterized by its reliance on static context windows and human-initiated prompting, has reached a plateau in enterprise utility.2 While highly capable of synthesizing text and code based on probabilistic patterns, generative models are fundamentally reactive. The next frontier, Agentic AI, shifts the operational model from a question-and-answer paradigm to an autonomous, goal-directed execution framework.12

Agentic systems operate on a continuous perception-reasoning-action (PRA) loop, which allows them to function dynamically within complex industrial environments.13 Within a manufacturing and PLM context, this cognitive loop functions through four distinct phases:

• Perception: The agent ingests and interprets multimodal data streams. This includes unstructured requirements documents, real-time sensor telemetry from shop floor machinery, and highly structured CAD metadata and geometric parameters.14
• Reasoning: Utilizing advanced LLMs as core reasoning engines, the system decomposes complex, high-level engineering goals into sequential, actionable sub-tasks. The agent evaluates historical design patterns, assesses constraints, and retrieves necessary contextual data via vector databases.2
• Action: The agent executes deterministic operations across external systems. This may involve invoking the Siemens NX Open API to alter a specific geometric tolerance, querying a PLM database for material compliance, or generating automated routing scripts for Manufacturing Execution Systems (MES).16
• Reflection and Learning: System outcomes, error logs, and human-in-the-loop (HITL) corrections are fed back into the cognitive memory architecture. This reflection phase refines future iterations, establishing an evolving institutional memory that adapts to the specific engineering nuances of the enterprise.13

1.2 Overcoming the Single-Platform Limitation in CAD and PLM

Historically, the application of artificial intelligence in engineering has suffered from a critical structural bottleneck: intelligence has remained confined within the proprietary walled gardens of individual software vendors.19 Native tools, such as the Siemens Teamcenter BOM Copilot and the NX Design Copilot, offer substantial usability improvements.19 They enable conversational BOM exploration, allowing engineers to utilize natural language queries to traverse complex product structures and discover relevant software features.20 However, these single-platform solutions generally lack the cross-system semantic interoperability required for end-to-end digital thread orchestration.19

True agentic engineering requires the establishment of “Product Memory”—an intelligent, persistent, and connected knowledge model that spans entirely disconnected architectures, including Enterprise Resource Planning (ERP), MES, disparate CAD environments, and external supplier databases.19 To achieve this level of integration, organizations must move beyond rigid legacy automation scripts. Instead, they require sophisticated multi-agent orchestrators capable of interpreting high-level engineering intent, dynamically interacting with multiple APIs simultaneously, and reconciling data fragmentation across the global supply chain.22

1.3 The Complexity Threshold and Multi-Agent Orchestration

As industrial tasks surpass the “complexity threshold” of any single Large Language Model, robust technical architectures demand specialized Multi-Agent Systems (MAS).24 In a production-grade MAS deployment, a primary orchestrator agent receives a high-level user request—for example, a mandate to optimize a high-voltage assembly for thermal dissipation while simultaneously updating all relevant compliance documentation and PLM records. The orchestrator decomposes this massive goal and delegates the specific sub-tasks to a specialized team of worker agents.12

In this architecture, one agent may interface directly with a Computational Fluid Dynamics (CFD) solver to simulate thermal loads.26 Simultaneously, a geometric agent utilizes the NX Open API to adjust the heat sink fin density based on the CFD feedback, while a regulatory agent queries an external database to ensure the newly selected materials comply with European chemical safety standards.16 This decentralized, multi-agent approach systematically mitigates context window overload, reduces latency, and isolates execution boundaries, thereby significantly enhancing the overall reliability and fault tolerance of the system.24

1.4 The Agentic Trust Fabric

The viability of these autonomous workflows relies entirely on the establishment of an Agentic Trust Fabric.13 This framework integrates the Agent Runtime Environment (ARE)—which enforces strict execution boundaries, provides isolation, controls resets, and ensures lifecycle management—with a comprehensive Trust Envelope.13 The Trust Envelope governs security, logging, protocols, and observability, ensuring that the autonomous system does not drift, repeat mistakes, or fail silently.13 This unified adaptive framework guarantees that trust is reinforced at every API boundary and within every cognitive loop, allowing multi-agent systems to progress beyond proof-of-concept demonstrations and operate responsibly at enterprise scale.13

2. Architecting Interoperability: The Model Context Protocol (MCP)

To operationalize Agentic AI within complex CAD and PLM ecosystems, the underlying technical architecture must bridge the gap between stochastic reasoning engines and deterministic engineering platforms. The traditional approach of building custom, hard-coded API wrappers for every LLM-to-system interaction results in highly fragmented, brittle infrastructure that is difficult to maintain and scale.28

2.1 Standardizing AI-to-Data Integration

The Model Context Protocol (MCP) has rapidly emerged as the definitive open-source standard for resolving API fragmentation in agentic architectures.10 Functioning analogously to a universal interface standard for artificial intelligence, MCP dictates exactly how context is delivered to models.10 It enables developers to build secure, two-way connections between their proprietary data sources and AI-powered tools, allowing autonomous agents to interface securely with local data spaces, PLM databases, and CAD APIs without requiring bespoke integration code for every new model iteration.10

The MCP architecture is predicated on three foundational pillars that define the boundaries of agent interaction:

2.2 Implementing MCP for Siemens NX and PLM Environments

By utilizing MCP, an engineering organization or specialized software integration partner can deploy an MCP server that securely wraps the Siemens NX Open API or Teamcenter endpoints.16 When a design engineer requests an automated assembly modification, the local MCP client translates the LLM’s cognitive intent into a secure, structurally validated API call.28

This architecture strictly enforces data containment. The model is constrained to accessing only the data explicitly permitted by the protocol’s boundaries, reducing the surface for unauthorized data access.32 Furthermore, this structure intrinsically mitigates the risk of AI hallucinations. Because the agent relies entirely on structured, mathematically verified responses returned from the MCP server—rather than synthesizing information from its own latent, probabilistic memory—the outputs are grounded in actual, current engineering data.32

2.3 Overcoming Token Consumption and Latency Bottlenecks

As the ecosystem of connected tools grows within a massive PLM environment, loading all tool definitions upfront and passing intermediate geometric results through the LLM’s context window can severely increase latency and API costs.28 Excessive token consumption from complex tool schemas degrades the agent’s reasoning efficiency.28

Advanced MCP implementations solve this by enabling code execution capabilities.28 Rather than forcing the LLM to process massive JSON payloads of geometric data, the agent can generate and execute Python or C++ scripts via the NX Open API directly within a secure, sandboxed runtime environment.16 The script interacts with the CAD kernel, performs the necessary calculations or modifications, and returns only the final, validated result to the agent. This approach minimizes token overhead, accelerates execution speed, and allows the agentic system to handle vastly more complex, multi-tool workflows.28

3. Bridging Deterministic CAD Systems and Probabilistic AI

The fundamental friction between artificial intelligence and industrial engineering lies in the concept of determinism. CAD systems, structural solvers, and MES platforms require absolute mathematical precision and repeatable outcomes.3 Conversely, Large Language Models are inherently probabilistic; they generate slightly different text, code, or parameters even when provided with identical input prompts.37 Unchecked non-determinism in industrial design is unacceptable and can lead to catastrophic physical failures, manufacturing defects, and severe safety violations.

3.1 RAG vs. Fine-Tuning for Engineering Context

Equipping an LLM with deep, proprietary engineering knowledge requires a sophisticated, highly deliberate data architecture. The industry consensus for industrial applications points toward a hybrid approach, strategically deploying both Retrieval-Augmented Generation (RAG) and model Fine-Tuning.11

A robust, enterprise-grade architecture leverages Fine-Tuning to teach the model how to think—instilling a deep semantic understanding of the complex hierarchical relationships within a CAD assembly tree—while utilizing RAG coupled with MCP to provide what it should think about, retrieving the exact, current tolerances required for a specific part from a live, authoritative PLM database.31

3.2 Mitigation Strategies for Non-Deterministic Outputs

To safely deploy agentic workflows in mission-critical environments, technical architectures must implement stringent, multi-layered mitigation strategies:

• Algorithmic Temperature Control and Seeding: The stochastic nature of the LLM must be tightly constrained during engineering tasks. This is achieved by setting inference temperature parameters to near-zero and utilizing fixed random seeds to ensure maximum reproducibility during the generation of automation scripts or API calls.37
• Deterministic Fallbacks and Bounding: AI agents should be utilized strictly for workflow orchestration, generative design space exploration, and script generation.22 However, the actual mathematical validation and structural sign-off must be immediately routed to deterministic tools—validated solvers, established finite element analysis (FEA) methods, and certified calculation procedures—ensuring all results meet uncompromising engineering rigor.17
• Continuous Output Validation via Judge Models: Implementing specialized “Judge LLMs” or rigorous rule-based linters that automatically verify AI-generated NX Open scripts against established geometric and syntactic constraints before any execution is permitted within the CAD environment.32
• Human-in-the-Loop (HITL) Checkpoints: Designing the Agent Runtime Environment to automatically pause execution and flag critical decision matrices or anomalous geometric proposals for human engineer approval. This shifts the AI’s role from a unilateral, autonomous decision-maker to a high-speed, proactive cognitive assistant that amplifies human engineering capability.13

3.3 Vectorizing CAD and Engineering Data

Traditional RAG systems are optimized for text. However, the vectorization of CAD data is vastly more complex.43 Engineering models contain deep hierarchical metadata, spatial constraints, parametric relationships, and extensive version histories that must be accurately captured to be useful for AI retrieval.43 Advanced implementations require parsing proprietary CAD files (such as Siemens NX PRT files) to extract structured geometric and spatial features, converting them into sequence representations that support downstream natural language transcription and spatial reasoning by the LLM.44 By leveraging advanced vector search engines to manage these multi-dimensional datasets, organizations can enable LLMs to reason accurately about physical parts and assemblies.43

4. Sovereign AI Infrastructure: Ensuring Digital Independence in Europe

As European industrial entities aggressively adopt AI, the reliance on US-based hyperscale cloud providers introduces severe compliance, legal, and intellectual property vulnerabilities.45 The implications of the US CLOUD Act, which can compel US companies to provide data stored on European servers to US authorities, conflict directly with European data sovereignty ambitions.45 In response, the deployment of “Sovereign AI” infrastructure has transitioned from a theoretical preference to an absolute strategic imperative for European B2B technology leaders.7

4.1 The Strategic Imperative for Bare-Metal and On-Premise Deployments

Sovereign AI mandates that an organization maintains absolute, cryptographically secure control over its entire artificial intelligence infrastructure; sensitive engineering data never leaves the corporate network boundary, and the organization completely dictates how models are governed, trained, and executed.46 For regulated European manufacturing industries, deploying advanced LLMs on-premises or on private, European-hosted bare-metal servers provides three undeniable strategic advantages:

1. Absolute Data Sovereignty and GDPR Compliance: Localized data processing inherently neutralizes all cross-border data transfer risks. It eliminates the threat of devastating GDPR penalties (which can reach €20 million or 4% of annual revenue) associated with transmitting sensitive IP, proprietary CAD models, or employee data to external, third-party APIs.47
2. Superior Performance and Deterministic Latency: Executing inference on local bare-metal infrastructure eliminates external network round-trips and the unpredictability of shared public cloud resources. This enables sub-20ms inference latency, which is an absolute prerequisite for real-time shop floor automation, robotics integration, and high-frequency autonomous agent loops.47
3. Predictable Cost Economics: While public cloud AI APIs charge incrementally per token, high-volume agentic workflows—which require continuous background polling, iterative tool calling, and constant metacognition—can lead to exponential and unpredictable cloud computing costs. Bare metal infrastructure offers a flat, predictable operational expenditure. For workloads exceeding 60% utilization, the financial break-even point for purchasing dedicated GPU hardware is typically reached within three to eight months.47

4.2 Architecting a Production-Grade Sovereign LLM Stack

A modern, production-grade Sovereign AI stack capable of supporting complex B2B engineering workflows requires the sophisticated orchestration of several distinct, high-performance open-source technologies:

4.3 Integration with Gaia-X and Federated Data Spaces

To scale sovereign capabilities beyond the internal corporate firewall and collaborate across the broader European supply chain, industries are rapidly adopting the Gaia-X Trust Framework.48 It is critical to understand that Gaia-X is not a new cloud platform; rather, it is a decentralized architectural standard and ruleset that ensures transparent, controllable, and interoperable data exchange across Europe, preventing vendor lock-in and the concentration of power.6

The technical foundation of Gaia-X, specifically the “Danube” release, enables organizations to securely federate data across diverse ecosystems using Cryptographic Self-Descriptors, digital identifiers, and Verifiable Credentials.8 For a European automotive manufacturer, this architecture allows an internal Agentic AI system to securely query a tier-one supplier’s material database for carbon footprint metrics—necessary for environmental compliance—without the supplier ever exposing their proprietary CAD designs or manufacturing intellectual property.49

The integration of Agentic AI with Gaia-X Data Space Governance Authorities (DSGAs) enables the automated enforcement of sector-specific compliance rules.8 This ensures that multi-agent systems operate strictly within the bounds of European digital sovereignty, seamlessly facilitating cross-border collaboration while maintaining absolute cryptographic trust.8

5. Governing High-Risk AI: The EU AI Act Framework

The implementation of artificial intelligence, particularly autonomous agentic systems, in European manufacturing is heavily dictated by the European Union’s Artificial Intelligence Act (EU AI Act). Designed to ensure safety, fundamental rights, and accountability, the Act imposes rigorous, uncompromising technical requirements on the development and deployment of AI.5

5.1 Risk Classification in the Manufacturing Context

Under the EU AI Act, AI systems are legally categorized by their potential risk: Unacceptable (which are prohibited), High, Limited, and Minimal.9 In the context of industrial automation, CAD generation, and PLM orchestration, AI agents that directly influence the design, safety parameters, or quality control of safety-critical products (e.g., machinery, automotive components, medical devices, or infrastructure) can fall under the High-Risk category, namely where the AI system is a safety component of a product covered by Annex I that requires third-party conformity assessment, or where it falls under an Annex III use case (Article 6).9

This classification triggers mandatory, rigorous compliance with Chapter III, Section 2 of the Act, fundamentally altering the software development lifecycle, risk management strategies, and operational governance for B2B engineering software providers and system deployers.56

5.2 Articles 9-15: Technical and Operational Mandates

Providers (developers) of high-risk AI systems must architect their solutions from the ground up to satisfy the stringent requirements detailed in Articles 9 through 15 of the EU AI Act.9 Compliance is not merely a legal checkbox; it requires deep, architectural integration of governance mechanisms.

5.3 Fulfilling Annex IV Technical Documentation Requirements

Article 11 dictates that the technical documentation must contain, at a minimum, the exhaustive elements specified in Annex IV of the Act.59 This requirement forces organizations to maintain a deep, structured, and auditable record of their entire software architecture and development process.61 For B2B software vendors, system integrators, and enterprises deploying customized Siemens NX integrations or private cloud AI solutions, the Annex IV technical file must comprehensively document:

1. General Description: The intended purpose of the AI system, exact versioning histories, details on how the system interacts with external hardware and software, and the specific forms of market placement (e.g., embedded software, standalone APIs, or cloud platforms).59
2. System Elements and Development Process: A highly granular breakdown of the system architecture. This must include documentation of any pre-trained third-party models utilized, the underlying design logic, the mathematical goals for parameter optimization, and the specific computational resources allocated for training and inference.59
3. Data Governance Documentation: Comprehensive datasheets outlining all training methodologies, data provenance, labeling protocols, and data cleaning techniques applied to the datasets.59
4. Monitoring, Functioning, and Control: Detailed performance metrics, explicitly defined capability boundaries, foreseeable unintended outcomes, and the specific technical provisions implemented for human oversight.59
5. Validation and Testing: Signed, dated reports of all test logs, rigorously detailing the system’s accuracy, robustness, and any assessments regarding discriminatory impacts.59
6. Post-Market Monitoring Plan: A detailed, operational framework conforming to Article 72, outlining exactly how the system’s performance, safety, and compliance will be continually evaluated in the live production environment.59

Given the immense complexity of compiling and maintaining this documentation, forward-thinking organizations are increasingly utilizing specialized LLMs and compliance tools to automatically identify gaps in technical files, ensuring that the architectural documentation remains perfectly aligned with the evolving legal provisions of the Act.63

6. The Digital Product Passport (DPP): A Multi-Agent Use Case

The convergence of Agentic AI, Sovereign Infrastructure, and European regulation is most clearly exemplified by the impending requirements of the Digital Product Passport (DPP). Mandated under the Ecodesign for Sustainable Products Regulation (ESPR), the DPP is a comprehensive digital record designed to drastically enhance the traceability, circularity, and sustainability of products placed on the EU market.27

6.1 Regulatory Timelines and Technical Specifications

Manufacturers must prepare for strict rules on data accuracy, interoperability, and serialization, with widespread implementation required rapidly following the establishment of the DPP systems.27 The regulation mandates that the DPP must be accessible via a machine-readable data carrier (such as a QR code, RFID chip, or standardized URL) affixed physically to the product.27

The technical framework for the DPP is highly exacting. It dictates that DPP data must be transferable without vendor lock-in, hosted by independent third-party service providers, and highly interoperable across the entire supply chain.27 The passport itself must aggregate multiple dimensions of highly specific engineering, legal, and lifecycle data, creating a massive data management challenge for manufacturers.27 Key data requirements include:

• Unique Product Identifiers (UID, batch numbers, serial numbers, TARIC commodity codes).64
• Exact chemical composition and granular material breakdowns.64
• Environmental performance metrics, specifically including Global Warming Potential (GWP), comprehensive carbon footprint calculations, and resource use data.27
• Legal compliance documentation (CE markings, safety certifications).27
• End-of-life handling instructions, comprehensive repairability guidelines, and recyclability indices.27

6.2 Leveraging Agentic AI for DPP Data Aggregation

The primary architectural challenge in implementing the DPP is extreme data fragmentation. The requisite information for a single product passport is rarely stored in a unified repository; rather, it is distributed across complex CAD models, financial ERP systems, unstructured supplier PDFs, and disparate Lifecycle Assessment (LCA) tools.64 Manual compilation and verification of this data for every single product variant is economically and operationally unfeasible.

Agentic AI systems provide the necessary architectural connective tissue to automate DPP generation effectively.21 By utilizing the Model Context Protocol (MCP) to standardize tool calling, a multi-agent system can dynamically orchestrate the entire data collection and synthesis process:

1. A specialized CAD Analysis Agent extracts geometric volumes, mass properties, and material assignments directly from the native Siemens NX model via the Open API, without requiring human intervention.
2. Simultaneously, a PLM Retrieval Agent queries Teamcenter to identify specific supplier nodes, part variants, and historical procurement data.
3. A Compliance Agent autonomously parses unstructured supplier Material Safety Data Sheets (MSDS) utilizing an on-premise RAG pipeline, standardizing the chemical composition data to ensure it meets ESPR requirements.
4. Finally, a Synthesis Agent aggregates these verified data streams, calculates the baseline carbon footprint, and formats the final output into the strict, standardized JSON/XML schema required for the European registry submission.52

This sovereign, agentic workflow not only ensures continuous compliance with evolving DPP mandates but also establishes a unified sustainability data layer that updates automatically and in real-time as engineering designs iterate within the PLM environment.64

7. Strategic Implementation and Architectural Integration

The realization of Sovereign Agentic AI within European manufacturing cannot be achieved through the procurement of off-the-shelf software. It requires deep, specialized expertise in custom software development, advanced cloud and on-premise infrastructure, and intricate systems integration. The architectural approach necessitated by the confluence of the EU AI Act, Gaia-X, and MCP requires a holistic digital transformation strategy.

Organizations must prioritize the development of bespoke middleware and custom agent orchestrators that can seamlessly bridge the gap between high-performance LLMs and deterministic engineering kernels like Siemens NX. The implementation of robust, automated digital processes is essential to manage the massive data pipelines required for RAG, Fine-Tuning, and continuous compliance monitoring. By integrating deep expertise in Siemens NX/Open API development with cutting-edge, localized AI infrastructure, enterprises can build custom, compliance-native software solutions that operate securely within the stringent parameters of European data sovereignty.

8. Conclusion: The Future of Autonomous European Manufacturing

The integration of artificial intelligence into European manufacturing has definitively transcended the implementation of isolated, generative algorithms. The industry is rapidly advancing toward the deployment of complex, multi-agent systems capable of autonomously navigating the rigid mathematical structures of engineering platforms like Siemens NX, while interacting seamlessly and securely across federated external data spaces.12

However, the operational viability of these autonomous workflows relies entirely on the establishment of a robust Agentic Trust Fabric, underpinned by uncompromising adherence to data sovereignty and regulatory frameworks.13 To thrive in 2026 and beyond, B2B CTOs, engineering leaders, and technology architects must prioritize the deployment of Sovereign AI infrastructure. By anchoring agentic architectures within secure, on-premise environments, standardizing tool integration via the Model Context Protocol, and rigorously aligning all development practices with the mandates of the EU AI Act and Gaia-X, organizations can safely unlock the transformative operational efficiencies of autonomous engineering.8 This strategic alignment supports regulatory compliance and IP protection and helps build a sustainable competitive advantage in the future of European industrial innovation.

Works cited

1. Agentic AI Architectures with Patterns, Frameworks & MCP, accessed on March 10, 2026, https://mehmetozkaya.medium.com/agentic-ai-architectures-with-patterns-frameworks-mcp-25afcc97ae62
2. What is Agentic AI? – IBM, accessed on March 10, 2026, https://www.ibm.com/think/topics/agentic-ai
3. AI-DRIVEN COMPUTER-AIDED DESIGN (CAD) SYSTEMS: LEVERAGING NEURAL NETWORKS FOR OPTIMIZED ENGINEERING PRODUCT DEVELOPMENT – TPM – Testing, Psychometrics, Methodology in Applied Psychology, accessed on March 10, 2026, https://tpmap.org/submission/index.php/tpm/article/download/2331/1782/5049
4. Realizing the role of agentic AI – Thought Leadership, accessed on March 10, 2026, https://blogs.sw.siemens.com/thought-leadership/realizing-the-role-of-agentic-ai/
5. European approach to artificial intelligence | Shaping Europe’s digital future, accessed on March 10, 2026, https://digital-strategy.ec.europa.eu/en/policies/european-approach-artificial-intelligence
6. Focus on digital sovereignty in Europe: Gaia-X as a central topic at the 2025 Digital Summit, accessed on March 10, 2026, https://gaia-x.eu/focus-on-digital-sovereignty-in-europe-gaia-x-as-a-central-topic-at-the-2025-digital-summit/
7. Building a Sovereign AI Stack: From Zero to POC (Guide) – DEV Community, accessed on March 10, 2026, https://dev.to/janealesi/building-a-sovereign-ai-stack-from-zero-to-poc-guide-1id6
8. Gaia-X Enters Season 2.0 of Data Spaces and Digital Ecosystems …, accessed on March 10, 2026, https://gaia-x.eu/gaia-x-enters-season-two-of-dataspaces-and-digital-ecosystems-with-summit-2025/
9. Whitepaper The European Artificial Intelligence Act – Fraunhofer …, accessed on March 10, 2026, https://www.iks.fraunhofer.de/content/dam/iks/documents/whitepaper-eu-ai-act-fraunhofer-iks.pdf
10. Introducing the Model Context Protocol – Anthropic, accessed on March 10, 2026, https://www.anthropic.com/news/model-context-protocol
11. RAG vs. fine-tuning – Red Hat, accessed on March 10, 2026, https://www.redhat.com/en/topics/ai/rag-vs-fine-tuning
12. Siemens introduces AI agents for industrial automation | Press | Company, accessed on March 10, 2026, https://press.siemens.com/global/en/pressrelease/siemens-introduces-ai-agents-industrial-automation
13. New Book: Agentic AI Engineering for Building Production-Grade AI Agents – Medium, accessed on March 10, 2026, https://medium.com/generative-ai-revolution-ai-native-transformation/new-book-agentic-ai-engineering-for-building-production-grade-ai-agents-4dc121128e45
14. What is agentic AI: A comprehensive 2026 guide – TileDB, accessed on March 10, 2026, https://www.tiledb.com/blog/what-is-agentic-ai
15. AI Business Trends 2025 | Google Cloud, accessed on March 10, 2026, https://cloud.google.com/resources/ai-trends-report
16. NX Open API Programming – Siemens Xcelerator Academy, accessed on March 10, 2026, https://training.plm.automation.siemens.com/ilt/iltdescription.cfm?pID=TR13110_______NX___10.0___2500
17. Agentic Engineering: How AI Agents Can Accelerate Advanced Modeling and Simulation Workflows – Rescale, accessed on March 10, 2026, https://rescale.com/blog/agentic-engineering-for-modeling-and-simulation-workflows/
18. Manufacturing’s 2026 Mandate: From AI Pilot to Agentic Profit – Dataiku, accessed on March 10, 2026, https://www.dataiku.com/stories/blog/manufacturing-ai-trends-2026
19. AI in PLM: What Are Vendors Really Delivering, What Did I Learn …, accessed on March 10, 2026, https://beyondplm.com/2025/11/23/ai-in-plm-what-are-vendors-really-delivering-what-did-i-learn-and-is-the-hard-problem-still-unsolved/
20. AI Capabilities in NX CAD | Design faster with AI – Blogs – Siemens, accessed on March 10, 2026, https://blogs.sw.siemens.com/nx-design/ai-cad/
21. Best AI for PLM: PLM Search Tools and Assistants (2026) – CoLab Software, accessed on March 10, 2026, https://www.colabsoftware.com/post/best-ai-for-plm-plm-search-tools-and-assistants-2026
22. AI Agents Transform Engineering Workflows – Design News, accessed on March 10, 2026, https://www.designnews.com/artificial-intelligence/closing-the-gap-between-ai-s-promise-and-engineering-impact
23. The Model Context Protocol (MCP): A Game-Changer for Agentic AI …Part 1 | by nikhil goyal, accessed on March 10, 2026, https://medium.com/@goynikhil/the-model-context-protocol-mcp-a-game-changer-for-agentic-ai-6a55c180efb4
24. Agent studio: A multi-agent system for systems engineering | The Art of the Possible, accessed on March 10, 2026, https://blogs.sw.siemens.com/art-of-the-possible/agent-studio-a-multi-agent-system-for-systems-engineering/
25. 25 Agentic AI Terms Every Business User Needs To Know – Salesforce, accessed on March 10, 2026, https://www.salesforce.com/blog/agentic-ai-definitions/
26. Siemens brings AI copilot, immersive design and integrated fluid and thermal simulation to NX, accessed on March 10, 2026, https://news.siemens.com/en-us/siemens-designcenter-nx-summer-2025/
27. Digital Product Passport: What’s a DPP – OneClick LCA, accessed on March 10, 2026, https://oneclicklca.com/en/resources/articles/digital-product-passport-whats-a-dpp
28. Code execution with MCP: building more efficient AI agents – Anthropic, accessed on March 10, 2026, https://www.anthropic.com/engineering/code-execution-with-mcp
29. Building Agentic AI apps with Model Context Protocol, accessed on March 10, 2026, https://www.youtube.com/watch?v=xPedq8X87vg
30. What is the Model Context Protocol (MCP)?, accessed on March 10, 2026, https://modelcontextprotocol.io/introduction
31. Agentic AI & MCP – developer.siemens.com, accessed on March 10, 2026, https://developer.siemens.com/building-x-openness/dev-guide/mcp.html
32. Model Context Protocol Explained: The Key to Agentic Healthcare – Artera, accessed on March 10, 2026, https://artera.io/blog/model-context-protocol-explanation/
33. How to build a simple agentic AI server with MCP | Red Hat Developer, accessed on March 10, 2026, https://developers.redhat.com/articles/2025/08/12/how-build-simple-agentic-ai-server-mcp
34. Disruptive Innovation or Industry Buzz? Understanding Model Context Protocol’s Role in Data-Driven Agentic AI | Informatica, accessed on March 10, 2026, https://www.informatica.com/blogs/disruptive-innovation-or-industry-buzz-understanding-model-context-protocols-role-in-data-driven-agentic-ai.html
35. What is the Model Context Protocol (MCP)? – Databricks, accessed on March 10, 2026, https://www.databricks.com/blog/what-is-model-context-protocol
36. Has anyone tried to use AI using NX Open ? specifically calling a LLM end point with a prompt and then processing that answer within a NX session ? Just curious to see if anyone has tried this and learn from any experiences. – Community, accessed on March 10, 2026, https://community.sw.siemens.com/s/question/0D5Vb000012iaRAKAY/has-anyone-tried-to-use-ai-using-nx-open-specifically-calling-a-llm-end-point-with-a-prompt-and-then-processing-that-answer-within-a-nx-session-just-curious-to-see-if-anyone-has-tried-this-and-learn-from-any-experiences
37. What are non-deterministic AI outputs? – Statsig, accessed on March 10, 2026, https://www.statsig.com/perspectives/what-are-non-deterministic-ai-outputs-
38. RAG vs Fine-tuning vs Prompt Engineering: Everything You Need to Know | InterSystems, accessed on March 10, 2026, https://www.intersystems.com/resources/rag-vs-fine-tuning-vs-prompt-engineering-everything-you-need-to-know/
39. RAG vs fine-tuning vs. prompt engineering – IBM, accessed on March 10, 2026, https://www.ibm.com/think/topics/rag-vs-fine-tuning-vs-prompt-engineering
40. RAG vs Fine-Tuning: What Actually Works in Production? | by Pratik Chaudhari – Medium, accessed on March 10, 2026, https://medium.com/@pratikchaudhariworks/rag-vs-fine-tuning-what-actually-works-in-production-b05bde81db0f
41. RAG vs. Fine-Tuning: Choosing the Right Approach, accessed on March 10, 2026, https://www.thenile.dev/blog/fine-tuning-or-rag
42. Generative AI for CAD Automation: Leveraging Large Language Models for 3D Modelling, accessed on March 10, 2026, https://arxiv.org/html/2508.00843v1
43. Private AI for CAD and Engineering Data – inteliScience, accessed on March 10, 2026, https://inteliscience.net/case-studies/private-ai-for-cad-and-engineering-data/
44. HistCAD: Geometrically Constrained Parametric History-based CAD Dataset – arXiv, accessed on March 10, 2026, https://arxiv.org/html/2602.19171v1
45. Gaia-X warns US hyperscalers about selling cloud sovereignty | Channel Dive, accessed on March 10, 2026, https://www.channeldive.com/news/gaia-x-warns-us-hyperscalers-about-selling-sovereignty/807362/
46. Sovereign AI | alfatier, accessed on March 10, 2026, https://alfatier.io/en/services/sovereign-ai/
47. Own Your LLM Stack: On-Premise Deployment, accessed on March 10, 2026, https://blog.easecloud.io/ai-cloud/own-your-llm-stack/
48. Gaia-X Trust Framework, accessed on March 10, 2026, https://gaia-x.eu/wp-content/uploads/2024/05/An-Introduction-to-the-Gaia-X-Trust-Framework_2024-V4.pdf
49. Gaia-X: the bid for a sovereign European cloud – Polytechnique Insights, accessed on March 10, 2026, https://www.polytechnique-insights.com/en/columns/digital/gaia-x-the-bid-for-a-sovereign-european-cloud/
50. European Initiative for Data Sovereignty Released a Trust Framework – InfoQ, accessed on March 10, 2026, https://www.infoq.com/news/2026/01/data-sovereignty-trust-framework/
51. GAIA-X: Technical Architecture – bundeswirtschaftsministerium.de, accessed on March 10, 2026, https://www.bundeswirtschaftsministerium.de/Redaktion/EN/Publikationen/gaia-x-technical-architecture.pdf?__blob=publicationFile&v=1
52. How to prepare for the Digital Product Passport: our solution for European businesses, accessed on March 10, 2026, https://www.rina.org/en/media/News/2026/03/06/digital-product-passport
53. Gaia-X context – Gaia-X Architecture Document – main version (40f2e2e8), accessed on March 10, 2026, https://gaia-x.gitlab.io/technical-committee/architecture-working-group/architecture-document/gaia-x_context/
54. Gaia-X Summit 2025 Showcases the Next Phase of Trusted Digital Ecosystems, accessed on March 10, 2026, https://gaia-x.eu/gaia-x-summit-2025-showcases-the-next-phase-of-trusted-digital-ecosystems/
55. White Paper on Artificial Intelligence – A European approach to excellence and trust, accessed on March 10, 2026, https://commission.europa.eu/system/files/2020-02/commission-white-paper-artificial-intelligence-feb2020_en.pdf
56. High-level summary of the AI Act | EU Artificial Intelligence Act, accessed on March 10, 2026, https://artificialintelligenceact.eu/high-level-summary/
57. Technical documentation obligations for AI providers: drafting manuals – Hogan Lovells, accessed on March 10, 2026, https://www.hoganlovells.com/en/publications/technical-documentation-obligations-for-ai-providers-drafting-manuals
58. White Paper: Effectively Implementing the EU AI Act – Governance, Control and Transparency for AI Systems – Technology Insights, accessed on March 10, 2026, https://tcblog.protiviti.com/2025/08/12/white-paper-effectively-implementing-the-eu-ai-act-governance-control-and-transparency-for-ai-systems/
59. Annex IV: Technical Documentation Referred to in Article 11(1) | EU …, accessed on March 10, 2026, https://artificialintelligenceact.eu/annex/4/
60. Article 11: Technical Documentation | EU Artificial Intelligence Act, accessed on March 10, 2026, https://artificialintelligenceact.eu/article/11/
61. Technical Documentation for High-Risk AI Systems – Blue Arrow, accessed on March 10, 2026, https://bluearrow.ai/hrai-technical-documentation/
62. AI Act Service Desk – Annex IV – European Union, accessed on March 10, 2026, https://ai-act-service-desk.ec.europa.eu/en/ai-act/annex-4
63. Simplifying software compliance: AI technologies in drafting technical documentation for the AI Act – PMC, accessed on March 10, 2026, https://pmc.ncbi.nlm.nih.gov/articles/PMC11965209/
64. Digital Product Passport Requirements 2026 | EU DPP & ESPR Guide for Manufacturers, accessed on March 10, 2026, https://www.eandox.com/resources/digital-product-passport-requirements-2026-eu-dpp-espr-guide-manufacturers
65. Digital Product Passports (DPPs) required by EU legislation across sectors: ESPR, toys, detergents, batteries, and more – Circularise, accessed on March 10, 2026, https://www.circularise.com/blogs/dpps-required-by-eu-legislation-across-sectors
66. The EU’s new Digital Product Passport (DPP): everything you need to know, accessed on March 10, 2026, https://blog.worldfavor.com/the-eus-new-digital-product-passport-dpp-everything-you-need-to-know
67. Unlocking the power of Model Context Protocol (MCP) on AWS | Artificial Intelligence, accessed on March 10, 2026, https://aws.amazon.com/blogs/machine-learning/unlocking-the-power-of-model-context-protocol-mcp-on-aws/