AI projects: what mechanical engineers must focus on

Leading companies are currently working on first creating a uniform framework that includes data strategy and ontology, data model and management. Each individual AI project feeds data into this structure - until all data is available for consumption in a uniform structure. Further AI projects then require only a fraction of the effort and AI agents access the necessary, contextualized data. So much for the idea. But in practice, the situation currently looks very different, especially for SMEs.

“Most companies already fail at basic digitalization homework. Unfortunately, you cannot, so to speak, put an aircraft cockpit on a two-stroke engine,” states Doris Aschenbrenner, Professor of Digital Methods in Production at Aalen University and head of the “Human in Command” working group. “80 percent of an AI project is data. But even more expertise is also needed. Guidance is provided, for example, by Plattform Industrie 4.0 with a four-level stack for Industrial AI that companies can use as orientation,” says the expert.

“Many companies lack the fundamentals: They often have no strategy or struggle with strategy development. As a result, it is then very difficult for them to find use cases that bring them added value,” also confirms Rosalind Salecker, change consultant for AI transformation. However, selecting the right use cases is essential, because otherwise frustration quickly arises when the effort invested in AI does not bring the hoped-for financial potential.

AI practice at the Mechanical Engineering Summit

Which AI projects are delivering productivity, quality, speed or new business models today? And why do they fail?  What is hype, what is a robust business case? Where is it worth getting started immediately? What data, processes and competencies are needed?

These are the guiding questions of the session "Digitalization & AI practice: What does it really deliver?" at the Mechanical Engineering Summit on November 10 and 11, 2026 in Berlin. One of the speakers will be  Doris Aschenbrenner.

Join us! More information and tickets are available here!

Data are the central pivot and focal point

Topics such as Industry 4.0 and IoT had indeed laid the foundation, especially in the production and logistics environment, for collecting a great deal of data. “However, some companies did not take this step consistently. As a result, the high-quality data for AI projects are now missing,” says Salecker. “Projects often fail because data are collected and optimized locally, but are not orchestrated across systems,” says Dr. Daniel Seiler-Thull, CTO Bosch Manufacturing Co-Intelligence. At Bosch, Manufacturing Co-Intelligence means the coordinated collaboration of humans, machines, and AI agents on a uniform, understood data basis.

“A key success factor is not to create new system silos, but to flexibly connect existing systems via a semantic umbrella,” Seiler-Thull is certain. The most labor-intensive, but essential preliminary work is the definition of standardized data models in order to describe physical data objects clearly and uniquely once, instead of building individual interfaces and mappings for each new software project. “In mechanical engineering, highly different data classes - from CAD and PLM structural data from development to ERP bills of materials and on to time-critical machine signals and unstructured service reports - must be brought together,” explains the CTO.

How industrial AI is changing the shop floor

Industrial AI is intended to connect simulation, shop floor and maintenance more closely. AWS sees Physical AI and agent-based systems above all as drivers for flexible production. An interview with Jan Metzner, Principal Specialist Solutions Architect Manufacturing.

Anja Ringel

Classic machine learning has long been established, but reaches its limits when interpreting unstructured data and in flexible, cross-system process control, says Daniel Seiler-Thull: “ GenAI and large language models introduce a new quality here, since they can act as translators between human language and complex machine data,” Seiler-Thull makes clear. Not an isolated AI, but rather the coordinated interaction of qualified employees, existing physical machines and agentic systems solves real problems throughout the entire product life cycle.

Without context, AI runs into a void

“Generative AI and AI agents require not only a high volume of individual, current company data, but also context,” says Gernot Schäfer, partner at Efeso Management Consultants. Contextual understanding for AI requires a company-wide semantic model: the more context, the better the results. AI also needs many partly unstructured data from emails, chat and voice even in production topics. “All data involved in the production process are needed,” Schäfer makes clear. This also includes bills of materials, maintenance manuals, service reports, quality logs, quotation data, complaints and customer feedback. A lack of contextual understanding is accordingly, alongside missing data, a key aspect of why AI projects often do not achieve the hoped-for success.

“Many production data can be valuable - provided they are cleanly contextualized and made usable. They should first of all be integrated into a so-called semantic framework - only then can they be used across the board,” explains Aschenbrenner. This includes, for example, OPC-UA modeling. Topics such as connectivity and time-series harmonization are also important steps, which are still proving challenging, especially for SMEs. 

An example: In the case of an injection molding machine, not only sensor data but also the respective machine settings after each setup process must be documented. If this context is missing, anomalies in the data can hardly be interpreted correctly anymore. The open standard MCP (Model Context Protocol) now helps to consistently connect AI applications with machines, software and data sources without having to build individual interfaces each time.

Industrial AI forms the basis for a resilient industry

Artificial intelligence alone does not make industry fit for the future. What is decisive is whether companies connect AI and digital twins into a new industrial architecture. Sabine Scheunert of Dassault Systèmes explains how this can succeed.

Anja Ringel

Qualification: People must be able to critically assess AI

Above all, the organization must also be brought along, transformation expert Salecker names this as one of the most important success factors for AI. If the workforce does not know where the journey with AI is heading, whether jobs are at risk or tasks are shifting, fear arises. “Then a defensive attitude can develop that stands absolutely in the way of a successful implementation of new AI tools or AI automations,” says Salecker. 

According to the current AI study by Bitkom, 43 percent of companies still do not offer training for AI - a figure that Salecker also considers realistic for mechanical engineering. In addition, general training for example on prompting has only limited effect. Instead, training should be adapted as individually as possible to the situation, strategy and use cases in the company in order to truly bring the workforce along.

Today this is even mandatory: “Since February 2025, the EU AI Act obliges companies that use AI to build up sufficient AI competencies among their employees,” notes Rosalind Salecker. “The human factor is often underestimated: It must be clarified who is responsible for result validation, result adjustment and corrections, so that the LLM can learn from the interaction with the human being,” adds Schäfer. “The major challenge always remains to evaluate and critically review the results and interpretations of AI models, because they are not absolutely reliable,” explains the transformation expert.

All the more important, therefore, are interdisciplinary teams from IT, data specialists and experts from the areas. From Aschenbrenner's point of view, it is also problematic that otherwise mistakes are quickly made, for example the attempt to apply an LLM to numerical data from production and expect helpful results. All the experts see competence building - not surprisingly - as one of the decisive success factors.

The days of static systems are numbered

“Machine and plant manufacturers must understand that the days of deterministic, static systems such as ERP or MES, where it is clear what they can do and how to train them, are definitively over,” Gernot Schäfer is certain. “AI agents act dynamically and do not have predictable, technically controllable behavior of their own. That means I now have an active, digital entity in production,” says the expert. One example: The agent for worker assistance continues learning and adapts its behavior accordingly, thus providing different instructions than at the beginning, for example.

For the introduction of AI systems, an iterative “learn-and-adapt approach” is therefore needed, in which the governance model, compliance, responsibilities, security and the involvement of people must be clearly defined. He sees a transparent enterprise architecture as the central hub for all AI projects. “Anyone who does not properly answer all these questions for themselves cannot scale AI in production and only moves on from one individual use case to the next,” Schäfer believes.

“The exciting thing is that machine builders have to build an AI system architecture that translates between a highly deterministic and reliable system, namely production, and a non-deterministic external system: the environment including people,” Aschenbrenner describes it. Until now, people had tried to adapt the workers to the deterministic machine processes. Now, for example in the area of high-mix-low-volume production with many activities in which humans are “better,” such “translation interfaces” could ensure greater productivity. 

Precisely where it is not always clear what is coming, as in recycling, re-manufacturing or in areas of process anomalies and uncertainty, AI architectures can be particularly helpful, says the AI expert. “The real strength certainly comes from a hybrid intelligence, as it is being discussed on the Plattform Industrie 4.0 in the Working Group Work,” believes Doris Aschenbrenner.

AI must be embedded in overarching governance

An important scaling factor lies in the democratization of software development. With modern tools that link LowCode/NoCode with GenAI, specialists can very easily develop their own solutions. In quality management or production planning, small automations could thus replace the manual transferring and reformatting of data, for example, Salecker cites as an example. “The interaction with Enterprise Architecture Management and IT security is central as soon as AI systems not only analyze, but actively support processes, coordinate them or, prospectively, intervene in manufacturing in a controlling manner,” says Seiler-Thull. 

On the one hand, the technical barrier to entry has dropped significantly: Even specialist departments could, for example, build AI agents in a decentralized way. “However, this is precisely where one of the greatest governance challenges lies. What works well locally on one line does not automatically have to be scalable from line to line, from plant to plant or worldwide,” warns the expert. If scalability is only subsequently implemented into individual solutions, high integration, operating and security costs often arise, which could in some cases exceed the original benefit.

Use regulations as a booster

“In the interest of robust AI governance, it is therefore not only a matter of complying with technical security standards such as role-based access controls, tenant isolation, OAuth2, OpenID Connect or standardized interfaces such as OPC UA and REST APIs,” Seiler-Thull states. What is also crucial is to avoid the uncontrolled proliferation of local homegrown solutions and to embed AI agents in a controlled, reusable and enterprise-capable architecture.

Although dealing with regulatory requirements such as the EU AI Act is administratively and technically demanding, as it requires seamless traceability and validation of data flows. It is nevertheless worthwhile to see regulation not as a pure obstacle, but as a catalyst for a clean data architecture. “Anyone who is forced to document their material flows and greenhouse gas balances in a tamper-proof and machine-readable way, inevitably creates the standardized data basis that is also required for advanced AI applications,” says Seiler-Thull. The technical complexity can be managed by translating regulatory requirements directly into standardized data models such as the IDTA’s Asset Administration Shell and providing them as out-of-the-box functions in the software infrastructure.