Reid Hoffman — LinkedIn founder, billion-dollar investor at Greylock, and one of the most influential minds in Silicon Valley — said a sentence at the AI Action Summit in Paris in February 2025 that still resonates in Europe today:

"You look at AI as a World Cup football match between the U.S. and China. And if all Europe is trying to do is be the referee, there's two problems. One, they never win, and two, no one really likes the referee."

That is blunt. But it hits a sore point: Europe lost the race for major internet platforms and foundational AI models. Looking in the rear-view mirror will not help now. The much more relevant question is: What is the next big thing?

The next round belongs to industry

Thomas Saueressig, SAP Executive Board member, gave one possible answer in an AFP interview in March 2026: "AI for use in industry is where I believe Europe can succeed big time around the globe — because of the industrial knowledge and the industrial data and competency we have in Europe." In other words: AI that creates value where Europe has led for decades — in manufacturing, mechanical engineering, and industrial value creation. EU manufacturing generates EUR 2.5 trillion per year, almost on par with the U.S. (even though we often see ourselves as much smaller).

A spirit of momentum — and a blind spot

For months, I have been speaking with venture capital investors, owners of large industrial groups, managing directors, and leading professors in manufacturing engineering and AI. Despite very different perspectives, I see one consistent pattern: everyone wants to shift into forward gear. Despite all global disruptions. The frustration of falling behind technologically has turned into new determination — the will to combine European industrial capability and AI to build a new, globally competitive industry.

But what does "combining AI with engineering and industrial knowledge" mean in concrete terms? The same terms keep coming up: Knowledge Graph. RAG. GNN. Digital Thread. Agentic AI. All are powerful approaches, but for many operational leaders and employees in industry, they are still abstract. Many ask me: "What does this mean in practice for my work or my company?".

That is exactly what I will cover in this article. I will explain these concepts — Knowledge Graph, Digital Thread, GNNs, RAG, etc. — using an industrial example: automated creation of routings/work plans.

Why it is worth focusing on routing/work plan creation

In variant engineering and make-to-order manufacturing, the same manual process starts after every incoming order: Design hands over a CAD model and usually a 2D manufacturing drawing — and Industrial Engineering must turn this into a complete routing/work plan. The process is labor-intensive, ties up experienced specialists for hours per part, and has a major impact on lead time and product cost. Because what is missed in planning cannot be fixed later on the shop floor, no matter how committed the team is.

One 3D model, one drawing, and a lot of experience

Because we engineers are (fortunately) trained to understand the problem first and then develop the solution, here is my view of how work is often done (especially for custom variants):

Design transfers a 3D model and a 2D drawing to Industrial Engineering. The 3D model describes part geometry: surfaces, edges, curvatures, spatial relationships. What it usually does not include is manufacturing information. The model shows what the part looks like — not how it should be produced. Technically, tolerances, surface finishes, and manufacturing requirements can be embedded directly in the 3D model — the standard exists. In practice, however, most companies still derive 2D drawings from the model and add manufacturing requirements manually. The result: the most relevant planning information sits in a document that cannot be easily processed by machines. Whether a cylindrical recess is a bore that must be reamed, or a through-hole that is only drilled, is not decided by geometry but by the tolerance on the drawing and the planner’s experience.

This is where the work of the process planner / industrial engineer begins. They open the 3D model, place the 2D manufacturing drawing next to it, and start identifying manufacturing features: holes, pockets, chamfers, relief grooves, threads. For each feature, they define the manufacturing method, machine, tools, and sequence. They rely on experience: Have they planned a similar part before? Is there a comparable routing/work plan in the ERP? They do this without system support — searching order history from memory, by part numbers, or by free-text descriptions. What they find, they interpret, adapt, and transfer manually into the ERP: operations, machine work centers, standard times, setup times.

This gives the process five characteristics that make it structurally hard to scale. First: CAD models contain no manufacturing semantics — identifying features from pure geometry is an intellectual task done by the planner. Second: tolerance and quality data are mostly on 2D drawings and therefore not machine-readable. Third: manufacturing knowledge — which feature on which machine, in what sequence, with which tool — is implicit experience and not formalized anywhere. Fourth: the search for similar reference parts depends on planner memory, not systematic similarity analysis. And fifth: there is no end-to-end continuity between CAD and ERP — every piece of information is transferred manually, again for every variant.

Why traditional solutions cannot deliver end-to-end automation

Any process planner asking why existing software tools do not take over more work quickly runs into several structural issues:

The first issue is the CAD system landscape. The most capable commercial feature recognition works only on each vendor’s native CAD model. There, the system recognizes parametric features: a hole modeled as a hole, a pocket modeled as a pocket. But once a STEP file from a third-party system arrives, that parametric information is gone. What comes in during import is raw B-Rep geometry: surfaces, edges, solids, but no feature tree, no design intent, no semantics. Recognition logic that works in native format breaks on STEP imports. In many companies, different sites also run different CAD systems — NX here, SOLIDWORKS there, CATIA at a third site. A solution that works only in one CAD ecosystem can therefore cover, at best, a fraction of real parts.

The second issue is recognition quality. Even where feature recognition is possible in principle, all commercial systems are rule-based: software searches geometry for predefined patterns — a cylinder with a certain depth-to-diameter ratio is classified as a hole, a rectangular recess as a pocket. This works for simple, isolated features. But as soon as features overlap or interact geometrically — a hole running through a pocket, intersecting machining operations, features with tolerance cascades — these rule sets fail. This is not an edge case; it is the norm in real machined parts. For the process planner, this means: they cannot rely on automatic recognition, must manually verify and rework every suggestion, and end up back in a manual process.

And the third issue is the most fundamental: PLM systems do provide geometric similarity search — they compare part shapes and find visually similar parts. But they do not compare manufacturing requirements. A part with tolerance class H7 requires reaming or grinding, while the same part with H11 only needs drilling — for a shape search, both are identical. And no system available today learns from what a company has actually manufactured in the past. Historical routings/work plans in ERP — thousands of records showing which feature was produced on which machine, with which tool, in which sequence — remain unused.

None of the solutions on the market automates the full process: from STEP file to robust feature recognition, tolerance evaluation, company-specific manufacturing knowledge, and ERP-compliant routing/work plan. This gap is exactly why process planners still plan many parts manually.

What does this have to do with Knowledge Graph, RAG, Digital Thread, GNN, and Agentic AI?

The starting point stays the same: a 3D CAD model arrives. But instead of a process planner opening the model and planning feature by feature manually, an automated pipeline runs.

The first stage solves what rule-based systems cannot: recognizing manufacturing features from raw geometry. This is where Graph Neural Networks are used — neural networks that work directly on CAD model structure. At its core, a CAD model is a network of surfaces, edges, and their relationships. A Graph Neural Network learns from thousands of annotated sample parts which patterns represent a hole, a pocket, a chamfer — even when features overlap or influence each other. And it works directly on STEP files, independent of the CAD system. But even the best neural networks do not detect every feature perfectly — especially with unusual geometries or parts that differ strongly from training data. That is why it is critical that the pipeline includes human-in-the-loop at the right points: the system states its confidence for each detection. Where confidence drops below a threshold, the process planner is involved in a targeted way — not for everything, only for complex special cases. Compared with the status quo of purely rule-based systems, the planner is involved far less often. Manual effort drops dramatically.

In parallel, manufacturing specifications are extracted — tolerances, surface finishes, fits. If a modern STEP AP242 file is available, they are stored in machine-readable form in the model. In the more common case — a 2D manufacturing drawing — a vision-language model handles extraction: an AI system that can read technical drawings like a human, but automatically. It recognizes tolerance fields, GD&T symbols, and surface specifications and maps them to features. The result: the quality requirements are defined for every manufacturing feature. Here as well, there will be specific cases where drawing content cannot be detected cleanly and the process planner must step in. But hours of drawing review become just minutes of targeted feedback.

The next stage is manufacturing knowledge. This is where the Knowledge Graph comes in — simply put, a machine-readable knowledge map. It connects three information worlds that are currently siloed: features detected from the CAD model, manufacturing data from the drawing, and historical routings/work plans from the ERP system. The latter contain the company’s materialized experience — which feature on which machine, with which tool, in which sequence. But this knowledge is buried in free-text descriptions, inconsistent time entries, and company-specific numbering schemes. Large language models — LLMs, the technology behind ChatGPT (which can also run on-premises so knowledge stays inside the company) — do most of this structuring work: they extract structured manufacturing knowledge from cryptic free-text fields and transfer it into the Knowledge Graph. But — and this is important — a Knowledge Graph is not created at the push of a button. It grows iteratively: AI extracts, manufacturing experts validate and correct, and each cycle improves precision. The key difference from classic master data projects is not that humans no longer review anything — it is that usable results are reached in days instead of years. And another important effect appears: the implicit experience of the best process planners — knowledge that previously existed only in individual heads — is systematically converted into a machine-readable knowledge structure. This is where the key difference from the internet economy becomes clear: it is not enough just to connect systems, analyze data automatically, and train models. For process planners, the Knowledge Graph is the most far-reaching component in the long term: manufacturing knowledge is available instantly and can be processed automatically by AI models. No traditional approach has achieved this so far. A large share of working time is currently spent searching for information. The potential time savings here are substantial. Another point: planning errors create high follow-up costs on the shop floor. Using Knowledge Graphs can significantly reduce these errors.

Similarity search runs on this knowledge base — and this is where RAG becomes concrete. RAG stands for Retrieval Augmented Generation: before an AI makes a decision, it searches the Knowledge Graph for relevant reference cases. For a new part, the system searches the entire historical part inventory — not by shape, but by manufacturing features, tolerances, material, and available machine park. So the question is not "Does this part look similar?" but "Has a comparable feature already been manufactured in a similar way?". This significantly improves language model output quality and creates the required acceptance by the process planner. Because only when people trust AI results will they use them in day-to-day work.

From the best reference case, a routing/work plan is derived automatically: operations, machine stations, standard times, setup times — in the format of the company’s ERP system, ready for direct import. Where the requested part differs from the reference part, the system adjusts affected positions in a targeted way and documents every adjustment. This is the Digital Thread in practice: an end-to-end data thread from CAD file to ERP routing/work plan. Time spent on manual data transfer is eliminated. Where the system is uncertain — for example, because no sufficiently similar reference part exists — it flags that point for review. The process planner fills gaps selectively instead of creating the full plan from scratch.

What I just described — a system that orchestrates multiple specialized AI components autonomously, makes decisions at each step, and involves people when uncertain — is what is called Agentic AI. Not one model solving one task, but an interplay of agents executing a complex process: detect, extract, connect, retrieve, derive, verify. It is the final component that can raise automation from below 20% to well above 80% — and massively reduce cost and lead time.

What does this mean for practitioners? It means that, for the first time, several things become possible that were previously structurally excluded. First: accumulated experience from the best process planners is formalized and made usable for everyone in the company — regardless of who is available today or who retires tomorrow. Second: a company can learn from its own historical data — not from industry averages, but from what it has actually produced on its own machines. And third: the routing/work plan is no longer created from scratch, but starts as a validated, adapted proposal — process planning shifts from highly repetitive tasks to complex new planning and AI model training.

We will not be the referee watching from the sideline

Automated routing/work plan creation is one concrete, focused example. But it represents a much larger field: CAD/PLM queries, drawing derivations, change justifications, specifications, inspection plans, NC/CAM preparation, P&ID/EPLAN-related documentation, design-for-manufacturing feedback, variant configuration, BOMs generation, quotation creation, manufacturing cost calculation — in Technical Sales, Engineering, and Industrial Engineering there is enormous untapped potential that can be unlocked with the same methods. There is a major opportunity for European companies to unlock this potential.

I know this article shows a certain degree of optimism. Many people say Europe has already lost the race in many areas. But that mindset never wins — not in sports, not in business. European engineers and entrepreneurs built the automotive industry, shaped modern mechanical engineering, and brought medical technology to the world’s top tier. In none of these cases was success guaranteed. In each case, success came from people going all in. Industrial AI can be one of the next fields like this. And Europe has everything it needs to win. Reid Hoffman called Europe the referee who never wins. We will prove him wrong.