Building the Data Foundations for Agentic AI in Manufacturing
Manufacturers know that AI is only as good as the data behind it. For agentic AI — systems that act on their own in real time — data silos, latency, and inconsistent architectures are often the biggest roadblocks. During the “Agentic AI in Manufacturing: From Copilots to Autonomous Systems” session at AI Frontiers 2025, organized by IIoT World, audience questions zeroed in on this challenge: how to prepare IT and OT systems so autonomous AI agents can deliver.
Central Repository vs. Direct System Access
For a reference architecture, if a manufacturer has an ERP, MES, and historian all in the cloud, then to implement agentic use cases, do you need to move this data into a central repository, or can agents reach into each system directly?
Peter Sorowka, Cybus GmbH: Both approaches exist, but agentic AI needs more than static data lakes. Think of it less as a parking lot and more as a real-time data highway. Unified access, low latency, and consistent data models are non-negotiable.
In greenfield industries like battery manufacturing, new equipment is now expected to connect not just to power but also to a unified namespace from day one. For legacy plants, integration and standardization become the harder — but essential — first steps.
Breaking Down OT/IT Silos
Agentic AI relies on resolving conflicts across OT/IT data silos. What proven strategies or architectures have you seen work best?
Peter Sorowka, Cybus GmbH: Unified namespace and Model Context Protocol (MCP) are among the most effective frameworks. But the bigger hurdle is organizational: if procurement doesn’t require data-friendly interfaces when buying machines, integration later becomes a nightmare.
Sebastian Trolli, Frost & Sullivan: Even when silos are bridged, context is critical. Real-time pipelines alone are not enough. Without metadata and models that explain what the data represents, agents cannot act reliably.
Mapping Agentic AI to ISA95 and ISA88
What should a reference architecture for Agentic AI look like, and how does it map to ISA95/ISA88?
Sebastian Trolli, Frost & Sullivan: There’s no single blueprint yet, but ISA95 provides a useful starting point. Agentic AI acts as a cognitive layer across the pyramid:
- Levels 0–1 (sensors, PLCs): Monitoring and advisory roles, not direct control.
- Level 2 (SCADA/HMI): Agents aggregate views across HMIs, assist operators, and exchange commands.
- Level 3 (MES): The “sweet spot” where agents orchestrate workflows, schedule tasks, and resolve exceptions.
- Level 4 (ERP, supply chain): Aligning enterprise goals with shop-floor execution.
Rather than replacing ISA95 or ISA88, agentic AI extends them, linking operational layers to business objectives with real-time intelligence.
What All This Means
Agentic AI cannot succeed without clean, contextual, and accessible data. That requires both technology and culture shifts: unified namespaces to remove silos, metadata to give meaning to streams, and procurement strategies that treat interoperability as non-negotiable.
The lesson is simple: factories don’t need more data lakes — they need living, connected data ecosystems where agents can think and act with confidence.
Speakers contributing to these answers:
- Peter Sorowka, CEO, Cybus GmbH
- Sebastian Trolli, Head of Industrial Automation & Software, Frost & Sullivan
Related articles:
