New industrial AI stack challenges automation pyramid

A new industrial AI technology stack is taking shape across manufacturing software and robotics, according to IoT Analytics, based on the firm's analysis of activity at IMTS 2026.

The Hamburg-based research company said the shift marks a move away from the long-established ISA-95 automation pyramid, which for more than two decades has served as a framework for organising industrial software from shop-floor sensors to enterprise resource planning systems.

Its latest analysis outlines eight layers in what it describes as an emerging AI-led stack: physical processes, software-defined edge, industrial DataOps, systems of record, knowledge graph and ontology, interfaces, agents, and applications with human oversight.

The argument reflects a broader debate in industrial technology over how manufacturers should organise data and software for AI. Rather than relying mainly on centralised cloud repositories, vendors are increasingly presenting architectures that allow AI systems to query operational data where it sits and act on it through software interfaces.

Data context

One of the clearest signs of that shift, IoT Analytics said, is the growing prominence of knowledge graphs. Manufacturers had previously been encouraged by major cloud providers to gather plant data into cloud data lakes, but discussions and demonstrations at IMTS pointed to a different approach: connecting and contextualising data where it already resides.

IoT Analytics cited AWS as making that case at the show and highlighted MachineMetrics as one of the vendors treating the knowledge graph as a primary point of differentiation. In practice, the approach is intended to preserve relationships between machines, processes, maintenance records and production data, giving AI tools more structured context than a large pool of raw information alone.

That matters because industrial environments often run on fragmented systems built up over many years. Machine telemetry, maintenance logs, quality systems and production planning tools may all hold relevant information, but AI applications can struggle when those sources are disconnected or lack a shared model.

MCP focus

IoT Analytics also pointed to the spread of MCP servers as another notable development. These servers are designed to let AI agents query live operational data and, in some cases, trigger actions inside industrial software.

Among the companies identified as presenting new MCP servers were MachineMetrics, PTC, Rockwell Automation, Limble and Vention. The research firm noted, however, that not all MCP servers offer the same level of access. It drew a distinction between native servers that expose application logic and wrapper-based approaches built on closed application programming interfaces, which may be limited to read-only access.

That distinction could prove important as manufacturers test how far to trust AI agents in factory settings. Read-only access may support monitoring, search and recommendation tasks, while deeper integration into software logic would be needed for agents to carry out workflow steps or operational changes directly.

The analysis suggests that vendors with direct control over operational systems and the surrounding data models may have an advantage if AI becomes more deeply embedded in industrial processes. It also points to pressure on companies whose role is limited to moving data between systems without adding context or control points.

Hardware layer

Beyond software, attention is now turning to hardware, IoT Analytics said. It highlighted Anthropic's Model Hardware Standard, or MHS, as an early effort to create a specification that allows AI agents to discover and operate physical equipment such as robot arms, sensors and test instruments.

Universal Robots has already tested MHS with four collaborative robots in a proof of concept, according to the firm, while Doosan Robotics and AWS are among the companies adding support. If adopted more broadly, such standards could extend AI interaction beyond software dashboards and business systems into direct engagement with industrial machinery.

That prospect raises technical and commercial questions for manufacturers and suppliers alike. Standards that let AI agents interact with machines could simplify integration, but they also place greater emphasis on safety controls, permissions and accountability when digital systems are linked to physical operations.

Knud Lasse Lueth, Chief Executive Officer, IoT Analytics, said: "On the software and robotics side of IMTS, AI was everywhere this year. What struck me most is that a new industrial AI stack is forming. Knowledge graphs now provide the context, MCP servers open industrial applications to AI agents, and Anthropic's new MHS could do the same for machines. The vendors that own shop-floor data and its context are in the strongest position. The vendors that only move data from A to B will likely come under pressure."