Artificial intelligence is moving beyond the screen. New generations of models, vision-language-action models, and significantly more powerful robotics are bringing AI into real-world production processes. In the first half of 2026 alone, nearly 50 billion US dollars flowed into physical AI startups—about four times as much as in the same period the previous year. For industrial companies, the question is no longer whether these technologies will arrive, but what foundations they must lay now to realize the benefits later.

‍

In a conversation with Dr. Matthias Foerth (AI & Engineering Lead at Tacto), Marc Krüger-Sprengel (Co-Founder & CEO at context/fab, formerly Bosch Rexroth), Michael Pfeiffer (VP Chief Expert AI at Bosch Research), and Robert Richarz (Scout at Andreessen Horowitz) discuss the technical differences between physical AI and language models, the data architecture required for AI agents in manufacturing, and why agent rollouts often fail in practice.

‍

AI is moving beyond the screen. What technically distinguishes physical AI from language models

‍

Michael Pfeiffer begins by clarifying the terminology: Physical AI primarily refers to robotics and autonomous driving, while Industrial AI is broader, encompassing engineering, procurement, and manufacturing. Technically, both are often based on the same backbone, supplemented in robots by vision-language-action models that translate speech, images, and other multimodal signals into actions. The crucial factor is that the reasoning and world knowledge of language models are preserved. In the past, a specialized model had to be built for every problem and trained over a long period with proprietary data; today, a base exists that can be fine-tuned. At the same time, compute, robotics, and sensor technology have improved significantly. Marc Krüger-Sprengel notes that AI has been used in production for over twenty years, but until now, it has been a problem-specific and highly predictable machine learning solution. The real innovation is currently taking place less in the models themselves and more in governance, guardrails, and context serving.

‍

The architecture behind the intelligence. Context is the limiting factor

‍

Marc Krüger-Sprengel looks back at the history of Industry 4.0: The assumption that simply plugging an Ethernet cable into every machine would suffice primarily produced large amounts of data junk across more than 200 different silos. That is why context/fab is building a data layer optimized not for human-readable tables, but for agents. In this "Context Graph," physical relationships, business processes, material flows, process variants, and intralogistics paths are mapped and linked with actions, allowing for control over which agent is permitted to perform which action within which scope and under which guardrails. Ultimately, this same structure makes data more accessible to humans—a realization that, according to Matthias Foerth, Tacto arrived at from the other side. Michael Pfeiffer sees additional potential in making agents self-learning, for example, by having them build their own wiki of previously solved problems while they work, as industrial knowledge is rarely openly available on the web but hidden in internal documents. Robert Richarz sums it up: The capability of AI today is limited less by the intelligence of the models than by the available context. Much knowledge is implicit; it is exchanged at the coffee machine or emerges while walking through the production line. Structuring this knowledge is currently the biggest limiting factor.

‍

AI agents at work. Why rollouts in manufacturing fail

‍

In Marc Krüger-Sprengel’s experience, agent rollouts rarely fail due to technology, but rather due to the timing of stakeholders. Often, the suggestion comes from plant management or operational excellence to let the results speak for themselves first and bring IT in later. He strongly advises against this, because it is precisely this IT department that will later be responsible for security concepts and, in the case of on-premise deployments, will take over parts of operations. Michael Pfeiffer adds the question of acceptance: He does not consider starting with a "dark factory" scenario a smart strategy, nor does he recommend equipping employees with sensors to collect data for their own replacement. A humanoid is rarely the first solution anyway; often, mobile manipulators on wheels, a cobot, or a single arm are sufficient. For him, the first step remains a data strategy: making data from the plants accessible, utilizing explicit knowledge from design manuals and regulations, and then tapping into implicit knowledge. Both warn against isolated solutions. Intelligent robot islands that optimize themselves while the rest of the factory does not keep pace are of little use. Referring to the Theory of Constraints, Marc Krüger-Sprengel puts it this way: If you make five percent of a process twice as efficient without adjusting the rest, you end up in the same place as before.

‍

A look at the next three years. What industrial companies should do now

‍

Robert Richarz sees AI as a basic tool in documentation, reporting, and financial processes today. Next, other parts of knowledge work will follow, such as understanding technical drawings and relationships in engineering, before the benefits gradually translate into physical production. He does not expect widespread use of humanoid robots within the next three years. His recommendation to industrial companies is therefore to use the coming 24 months to lay the foundations and clarify which context is relevant to their own organization and how it can be made accessible. Michael Pfeiffer sees Bosch playing a dual role here—through its own efficiency in manufacturing and through robotics as a business field—and points to the starting position in Germany: its own plants, concrete use cases, and domain knowledge extending into the mid-sized sector, all of which can be combined with modern AI. Marc Krüger-Sprengel observes that industrial companies now have almost no fear of contact with young providers on these topics, because the speed of development often leaves them no other option.

‍

Conclusion

‍

The discussion between context/fab, Bosch Research, and Andreessen Horowitz makes it clear that the bottleneck for AI in industry is no longer the model, but the context. Those who prepare data, as well as explicit and implicit knowledge, in a way that agents can utilize, are laying the foundation to realize the benefits in physical production later on. The key is to think about processes end-to-end rather than building isolated intelligent islands.

‍

Marc Krüger-Sprengel (Co-Founder & CEO at context/fab), Michael Pfeiffer (VP Chief Expert AI at Bosch Research), and Robert Richarz (Scout at Andreessen Horowitz) join Matthias Foerth (AI & Engineering Lead at Tacto) to discuss how artificial intelligence is transforming industrial value creation: the distinction between Physical AI and Industrial AI, the data architecture behind AI agents, common pitfalls during agent rollouts in manufacturing, and the foundations industrial companies should be laying over the next 24 months. For context: nearly 50 billion US dollars were invested in Physical AI startups in the first half of 2026 alone.

‍