At the Industrial AI Summit 2026, attendees asked questions about supplying context to industrial AI agents that the panel did not have time to answer live. Peter Sorowka, CEO of Cybus, provided answers covering how factories should structure context for AI, the role of MCP servers in connecting LLMs to plant data, the transition from human-supplied to agent-acquired context, the place of tacit knowledge in AI foundations, and the question of who owns the knowledge AI captures. This is part one of three articles based on unanswered audience questions from the “From Data Streaming to Autonomous Action” session.
How Do You Supply Better Context to Industrial AI?
Peter Sorowka, CEO of Cybus:
This depends really on what type of context we are talking about. Some context is rather static, e.g. the complete shop floor inventory, machine information, process flow, asset relations. This context should at best be supplied together with the unified namespace, in the form of a clear data ontology, knowledge graph and meta data. This context is independent from any specific use case. Additionally, especially when working with LLMs and agentic AI, use case specific context can just be anything, a text, a PDF, a photograph. What matters is to manage this context in a structured way, which can be as simple as a sharepoint or confluence or any more sophisticated system.
Can an MCP Server with Machine Drawings Help an LLM Understand a Plant?
Peter Sorowka, CEO of Cybus:
It will definitely help, but be aware, that those drawings are often outdated, and may disagree with the plant as built. So the human review and further human provided context will be of high relevance. When that is combined with a declared model of assets, signals and relations that the MCP server exposes, the LLM can answer from the model and can cite where a value came from.
How Do AI Agents Move from Human-Supplied to Self-Acquired Context?
Peter Sorowka, CEO of Cybus:
Humans will be crucial for the transition because so much knowledge and experience is actually in the heads of the people who have worked in the plant for decades. Humans supply context once, not on every prompt. The moment to capture it is commissioning, because the person wiring up the machine knows what it is and what it does. From then on the agent reads the model instead of asking a person. Humans stay in the loop for change — a new line, a new interlock — and out of the loop for retrieval.
Are Data, Context, and Tacit Knowledge the Foundation for Industrial AI?
Peter Sorowka, CEO of Cybus:
Yes. Data and context without it give an agent correct values and wrong conclusions — it knows the temperature, not why 82 °C is fine on this line and a problem on the next one. The usable form of that knowledge is concrete: interlocks, limits, and the conditions that must hold before an action is allowed. Writing them down as declared rules keeps them reviewable, so the operator can see what the system believes and correct it. Harvesting the same knowledge from logs gets you a model nobody can audit.
Is Industrial AI Repeating the Knowledge-Ownership Split?
The original question referenced Taylor’s stopwatch, which turned the worker’s tacit know-how into data owned by the organization. Industrial AI does the same at the scale of the whole process.
Peter Sorowka, CEO of Cybus:
The mechanism is the same: tacit know-how becomes an asset the organisation owns. Two things are different this time — capture is continuous rather than a study, and the system also acts on what it learned. So the split is real and it is the default outcome unless someone designs against it. The design choice is where the knowledge lands: as declared rules with a named owner, the operator remains the author and can read and correct what the system believes; extracted from logs without attribution, it is the stopwatch again. Worth noting the counter-position: some argue that formalising know-how is exactly what protects it when experienced people retire.
These questions were submitted by attendees of the “From Data Streaming to Autonomous Action” session, at the Industrial AI Summit 2026. Answers provided by Peter Sorowka of Cybus. The session also featured Kudzai Manditereza of HiveMQ, David Puron of Barbara, Jonathan Alexander of Albemarle, and Hamish Mackenzie of IIoT World. Session sponsored by Cybus, Barbara, and HiveMQ.
Related from IIoT World
