This is the second installment in a three-part series derived from the standing-room-only session, “From Data Streaming to Autonomous Action,” held at the Industrial AI Summit 2026. While the first installment laid the essential groundwork regarding data ingestion and the contextual foundations necessary for industrial intelligence, this article shifts the focus to the “how.” We are moving beyond theoretical frameworks into the technical architecture that enables modern manufacturers to bridge the gap between Large Language Models (LLMs) and the shop floor.
To navigate this complex intersection of OT (Operational Technology) and generative AI, we sat down with Peter Sorowka, CEO of Cybus, to address the practical, often messy realities of implementation. From the Model Context Protocol (MCP) to the future of agent-to-agent governance, Sorowka provides a roadmap for engineers and CTOs looking to move from digital dashboards to truly autonomous action.
The Technical Architecture: Connecting the LLM to the Shop Floor
The primary hurdle for many manufacturers is no longer the availability of data—it is the translation of that data into a language an LLM can parse and act upon. The industry is currently experimenting with a variety of integration strategies, but the emergence of the Model Context Protocol (MCP) represents a pivotal shift.
What is the MCP Server Concept for Industrial AI?
In the context of the factory, the MCP serves as a standardized bridge. “MCP is a standard way for an agent to discover and call tools,” explains Sorowka. “Applied to a factory, the ‘tool’ is the factory model itself: namespaces, assets, live values, all exposed with the same access rules that apply to human operators.”
The design philosophy behind MCP is to provide a secure, standardized entry point into the plant. By utilizing this protocol, any LLM—whether it is a simple chat-based interface or a complex autonomous agent—can query the factory state, understand the relationships between assets, and execute commands within the bounds of predefined security policies. This effectively democratizes access to complex OT data, allowing models to understand not just that a machine is "hot," but what that heat signifies in the context of the broader production line.
The Industrial AI Landscape: Platform Choices and Strategic Deployments
Manufacturers are currently split into three distinct camps regarding their AI platform choices. As Sorowka observes, the industry has yet to settle on a single monolithic architecture, leading to a landscape characterized by trial and error.
- Hosted Assistants: Many firms are leveraging managed services like Microsoft’s Copilot Studio. These platforms reach into the factory environment via API keys, offering a low-friction entry point for enterprises already embedded in the Microsoft ecosystem.
- In-House GPT Front-Ends: A significant portion of the market is building custom interfaces that wrap around models from OpenAI, Anthropic, or Google. This allows for greater control over the UI/UX but requires more robust internal middleware to manage data flow.
- Self-Hosted Models: For industries with stringent security or intellectual property requirements, data residency is paramount. These companies are increasingly turning to self-hosted models, ensuring that sensitive factory performance metrics never leave the premises.
“Our own product and engineering work at Cybus utilizes Claude and Copilot,” says Sorowka. “The design commitment for our platform, Connectware, is that it behaves the same regardless of which model sits on the other end. Interoperability is the key to future-proofing your stack.”
The Standards Gap: Connectivity vs. Contextualization
A recurring theme throughout the Industrial AI Summit was the dichotomy between connectivity and context. Connectivity—the act of extracting data from a PLC or a sensor—is largely a solved problem. Technologies like OPC UA and MQTT have become the industry standard for transporting data from the machine to the edge or cloud.
However, contextualization remains a major bottleneck.
“Contextualization is not solved,” Sorowka notes. “We have competing conventions like AAS (Asset Administration Shell), i3X, and UNS (Unified Namespace), and most plants model their data differently even within the same organization. For the human-AI relationship and governance, there is no industrial standard yet.”
The Scaling Dilemma
Scaling an AI initiative requires a portable way to define permissions: who may read, who may act, and what an agent is authorized to do. While MCP is a promising start, it covers the call (the ability to request data) but not necessarily the permission (the granular authorization).
Sorowka’s Practical Advice: Do not wait for a single, perfect industry standard to emerge. Instead, declare your model in a flexible, open format that can be projected into AAS or i3X at a later date. Being standard-agnostic today prevents costly refactoring tomorrow.
The LLM as Orchestrator: Decision Layer vs. Interface
There is a misconception that the LLM should become the "brain" of the factory, making every decision. Sorowka offers a more nuanced view: the LLM should act as the orchestrator, not necessarily the decision-making engine for critical physics-based processes.
“Today, the LLM is the interface and the orchestrator,” says Sorowka. “The industrial value still comes from purpose-built models—those designed for anomaly detection, time-series forecasting, and process optimization. These models do not become obsolete just because we have LLMs.”
The LLM’s true power lies in its reach. It can pull disparate factory data, trigger specialized models, and explain the findings in natural language. It acts as the glue that makes sophisticated data science accessible to non-data scientists. However, when it comes to decisions with physical consequences—such as adjusting a furnace temperature or stopping a production line—the system must be bound by hard-coded rules and safety interlocks. In this paradigm, the LLM proposes an action, and a rigid policy layer decides if that action is allowed.
Toward Autonomous Agents: Identity and Governance
Looking forward, the roadmap points toward agent-to-agent interactions. Currently, most industrial AI interactions are human-in-the-loop: an assistant acts on behalf of a named employee. But what happens when an agent needs to request a service from another agent?
“Agent-to-agent chains sharpen the identity question,” Sorowka explains. “When the caller is another agent, ‘on behalf of a person’ stops being sufficient. We are developing an agent identity framework where each agent has its own accountable owner. Our position is that the governed side—the enforcement of these identities—belongs at the factory boundary.”
This shift toward agentic manufacturing will eventually require a move away from simple key-based access to a more robust, decentralized identity management system.
Implications: Should AI Change the Process or Just the Tools?
The most profound realization from the summit was the parallel to the industrial revolution. When steam power was replaced by electricity, factories initially just swapped a steam engine for an electric motor, keeping the same inefficient floor layout. It took years to realize that electricity could power individual, decoupled components, allowing for entirely new manufacturing workflows.
We are at that same inflection point today.
“Most projects are still doing the ‘steam-to-electricity’ swap,” Sorowka observes. “They are using an agent to replace a dashboard within an unchanged, slow process. The real gain comes when the decision loop shrinks, which means fewer handoffs and fewer manual reports.”
The Future Hybrid Organization
We are approaching a future of hybrid org charts, where humans and agents work in tandem, sometimes with agents overseeing processes and humans overseeing agents. The technical challenge will be the creation of a governance framework for closed-loop agentic manufacturing.
If the security, approval flows, and audit logs aren’t baked into the architecture, the speed gains of AI will be negated by the risks of unmanaged autonomy. For manufacturers, the path forward is clear: start with standardized connectivity, prioritize contextual data modeling, and treat the LLM as a powerful orchestrator within a strictly governed, policy-driven environment.
These insights were gathered from the “From Data Streaming to Autonomous Action” session at the Industrial AI Summit 2026. The session featured expert contributions from Kudzai Manditereza of HiveMQ, David Puron of Barbara, Jonathan Alexander of Albemarle, and Hamish Mackenzie of IIoT World. The event was proudly sponsored by Cybus, Barbara, and HiveMQ.
