Manufacturing has long relied on structure: repeatable processes, precise planning, and predictable machinery. For decades, software mirrored this approach. Manufacturing execution systems (MES) and supervisory control architectures were designed to enforce order — and they succeeded because they operated in environments where inputs were stable and well understood.
Artificial intelligence is now challenging that paradigm. Yet most industrial AI deployed over the past decade has remained limited to narrow use cases or confined to pilot projects. Generative AI, despite its transformative impact in office settings, has struggled to deliver similar value on the factory floor. Its dependence on large, domain‑specific datasets and single, centralised models makes it difficult to align with the fragmented, inconsistent nature of industrial data and legacy systems.
A new direction is emerging. Instead of relying on one centralised intelligence, manufacturers are beginning to orchestrate AI as a network of specialist agents. These agents operate independently but collaborate when needed — much like colleagues with different expertise solving a shared problem. The goal is not to replace human operators but to extend their capabilities and, over time, help fill skills gaps as experienced workers retire.
From Centralised Control to Collaborative Intelligence
The concept of manufacturing co‑intelligence positions AI as an active participant in operations rather than a passive analytics layer. Each AI agent is designed with a defined role, decision‑making autonomy, and access to relevant data and tools. It can resolve issues independently or coordinate with other agents to handle more complex tasks.
In traditional manufacturing IT, the MES sits at the centre, dictating operations and managing system integration. Adding new capabilities often requires upgrading or replacing the platform. Co‑intelligence avoids this disruption. Agents sit on top of existing systems, interfacing via APIs, retrieving data, and issuing instructions just as a human operator would — but faster and without fatigue.
This makes the model particularly suitable for brownfield environments where replacing infrastructure is impractical. Legacy systems remain in place, but their capabilities are effectively enhanced by intelligent agents that bridge gaps, eliminate inefficiencies, and unlock insights hidden in existing data.
Why Semantic Structure Matters
For AI agents to operate effectively, the quality and structure of the data they consume is critical. When data is semantically aligned — consistently organised and described — agents can retrieve, interpret, and act on it far more reliably. Without this structure, agents expend unnecessary processing effort imposing order, increasing cost, complexity, and the risk of errors.
Experiments highlight the impact. Tests comparing semantically structured data with inconsistently formatted data showed a 60% improvement in output quality. The effect is even more pronounced in multi‑site operations. If each plant stores and labels data differently, agents must learn the nuances of every site before they can function effectively. Standardised formats make scaling easier and decisions more accurate.
The semantic layer is therefore foundational to co‑intelligence. It enables interoperability, reduces the likelihood of AI hallucinations, and ensures consistent performance whether agents are troubleshooting a machine in one facility or coordinating production across a global network.
Changing the Role of People and Systems
Agentic AI reshapes the relationship between people, systems, and processes. Many deployments will initially act as assistants — AI co‑workers rather than replacements. A troubleshooting agent, for example, can ingest decades of maintenance records and make that expertise instantly available to less experienced operators, reducing downtime and improving repair consistency.
In the near term, the most visible benefit will be augmenting human capability. A production engineer facing a complex issue can now rely on a digital colleague that not only identifies likely causes but also instructs connected systems to take corrective action. This reduces manual intervention and cross‑department coordination.
Over time, as experienced workers retire, some agents will assume entire functions previously handled by humans. This is not about removing people but preserving institutional knowledge in a form that continues to add value. Digital colleagues can retain and apply expertise that would otherwise be lost.
The technology also challenges traditional software categories. By enabling agents to orchestrate tasks across MES, SCADA, and ERP layers, manufacturers can avoid vendor lock‑in. Instead of relying on monolithic systems with unused or suboptimal modules, companies can adopt best‑of‑breed solutions, knowing agents can integrate them. This flexibility supports the gradual dismantling of the traditional automation pyramid in favour of a more fluid, task‑oriented model.
Building Agility Into Industrial Intelligence
Co‑intelligence represents a lighter, faster, and more open model than traditional IT programmes. Agents can be deployed individually to solve specific problems or combined into collaborative teams. They can come from different vendors and still work together, avoiding the proprietary lock‑in that has long defined industrial software.
This modularity allows manufacturers to respond quickly to new challenges without re‑engineering their entire IT landscape. If an agent proves effective in one plant, it can be replicated across other sites with minimal integration work. If an agent underperforms, it can be replaced without disrupting the broader system.
Crucially, agentic AI also changes who can build and adapt solutions. Much of the interaction occurs in no‑code or low‑code environments. Production engineers — those closest to the process — can design or modify agents themselves. They can test hypotheses directly, query data without waiting for IT or data science teams, and formalise successful experiments into production workflows.
Moving Quickly Without Overcommitting
For manufacturing leaders concerned about cost or disruption, the recommendation is to start small but act decisively. A fast pilot in a single plant, deploying two or three agents against well‑defined use cases, can demonstrate tangible business impact with minimal investment. Once validated, these use cases can be expanded or replicated across other facilities.
The urgency stems from competitive pressure. Early adopters of agentic AI will gain improvements in operational efficiency, problem resolution, and knowledge retention. Those who delay risk falling behind as benefits compound for early movers. Unlike the slow rollout of Industry 4.0, the agentic model is lighter and faster to implement — but its competitive window will be shorter.
The technology is ready. The data exists in many organisations. What is needed now is the willingness to treat AI not as a monolithic platform but as a team of adaptable, collaborative experts. Manufacturing has always been about orchestrating people, machines, and processes in harmony. Co‑intelligence extends that principle into the digital realm — and the manufacturers who embrace it first will set the pace for the rest.
By Martin Richter, Head of AI Go‑to‑Market at Bosch Connected Industry