Agentic AI Rewrites Factory Data Architecture Requirements

Agentic AI pushes the number of consuming systems on the shop floor from the tens (10–50) typical under ISA-95 to thousands of individualized consumers, creating an estimated ~100x increase in edge-based data consumers. IIoT World cites an IDC datapoint that 56.6% of industrial organizations are planning, piloting, or in early deployment stages for AI agents, indicating this shift is already underway.

Historical industrial architectures (Industry 3.0 and the ISA-95 model) assumed a small set of consumers — SCADA, MES, historians — and later cloud/analytics stacks expanded that to a few dozen consumers. Agentic AI changes the consumer profile: agents are goal-oriented, specialized, and require richly contextualized inputs (for example, a maintenance agent needs pressure telemetry, service history, vendor data, and batch context; a quality agent needs product specs, batch IDs, regulatory thresholds, and alarm history). Raw PLC tags without joined context are insufficient for agent decision-making.

The article argues ISA-95's layered transfer model wasn't designed for many individualized consumers and endorses a hub-and-spoke integration model that connects systems and delivers curated context on demand. It introduces Model Context Protocol (MCP) as an open protocol to aggregate and expose contextualized data for agent discovery and use. MCP is explicitly not a replacement for OPC UA, MQTT, SQL, or REST — those remain data sources; MCP layers semantics and context over them. Practical governance guidance: agents perform best when constrained to a small, focused set of MCP tools (5–10) to avoid degraded decision quality and hallucinations.

This is an operational engineering problem with immediate consequences for OT/IT convergence, edge compute sizing, data cataloging, and runtime governance. Without redesign, factories will face data plumbing bottlenecks, inconsistent context provisioning, and agent failure modes caused by poor or ambiguous inputs. Planning must include data cataloging/context engineering, edge aggregation layers (hub-and-spoke), protocol bridges, and strict tool governance per agent.

adoption and specification activity around MCP or similar context-layer standards; pilots that measure per-agent data volume and latency; integration patterns that combine OPC UA/MQTT sources with semantic/contextual layers; and governance frameworks that limit agent toolsets to 5–10 curated capabilities.