Editor’s note: This is the third of a series from our ARC Advisory Group colleague, Colin Masson, adapted from an article he wrote for ARC. Colin is a valued industry SME who will occasionally help Smart Industry translate what is happening in the fast-changing manufacturing technology landscape and conversations in the community.

See Colin’s two other recent pieces:

Industrial AI has transitioned to the ‘application phase’

Navigating financial anxiety around paying the bill for industrial AI

A profound realization is sweeping through the industrial software engineering community: a generalized large language model understands sentences, but it has no native grasp of thermodynamics, kinematics, or manufacturing geometry.

Capturing real value requires anchoring neural networks within the strict boundaries of physical sciences, transforming the digital twin from a passive system of record into an active, governed foundry for human-agent collaboration.

Under this paradigm, generative architectures are utilized as high-velocity upfront acceleration filters. Rather than forcing engineering teams to spend weeks manually testing infinite permutations in a design space, generative models are deployed to compress and narrow down design boundaries from millions of options to a handful of high-probability candidates.

These candidates are then immediately handed off to high-fidelity, deterministic physical simulation models for absolute mathematical validation. The generative brain creates the concept, but the deterministic twin enforces physical reality.

Eradication of the ‘data tax’

The single highest cost factor in any advanced industrial data initiative is the manual data engineering tax—the tedious, repetitive process of extracting tag streams from disparate plant historians and manually mapping them to asset descriptions.

Advanced software orchestration environments (such as Siemens’ Intelligence Center X, mapping to Archetype 6 of the ARC Industrial AI Taxonomy) target this bottleneck by functioning as automated knowledge graph generation engines.

The primary breakthrough is the introduction of out-of-the-box, pre-populated industrial ontologies built directly into the core platform semantic middleware layer. By structuring ingestion feeds natively into shared lifecycle intelligence, these data models pre-code industrial logic across distinct operational spheres:

• Siemens Designcenter X and Teamcenter X ontologies: Natively pre-populated with engineering lifecycle semantics, understanding the precise relational attributes between 3D CAD topologies, parts lists, geometric configurations, material data sheets, and engineering bills of materials.
• Siemens Opcenter X ontologies: Natively pre-populated with manufacturing operations semantics, mapping real-time station routings, production logs, quality tolerances, machine toolpaths, and standardized equipment downtime fault signatures.

Because these ontologies are pre-coded with industrial context, incoming autonomous execution agents do not have to spend cycles trying to "learn" how a factory data model is organized.

They connect to a pre-structured representation of physical reality, allowing an agent to immediately cross-reference a real-time edge telemetry anomaly directly back to the original CAD design specification or materials log without human intervention.

Brazil’s leading flat-glass manufacturer, Vivix Vidros Planos, grounded this architecture by constructing nearly 30 distinct Mendix applications that bridge data layers across their SAP S/4HANA ERP core, industrial edge nodes, and a centralized Snowflake data warehouse.

When they stood up their AI-powered virtual assistant, they utilized the low-code platform as the secure application interface to govern the agent's actions, driving an 85% reduction in production issue resolution latency while capturing 6,000 hours of manual labor in a single operational year.

Design surrogates vs. lifecycle frontier models

To maximize commercial value, executives must look past unified marketing definitions to identify a sharp technical bifurcation in modern industrial codebases. Software portfolios are advancing along two entirely distinct, parallel algorithmic tracks:

Track A: Customer-specific geometric deep learning surrogates

Designed for high-speed physics approximation, this track transforms the passive CAD/PLM repository into an active predictive environment (such as Siemens’ Simcenter PhysicsAI), achieving 500 times to 1,000 times simulation acceleration metrics.

See also: Autonomy is a journey, not a switch to be turned on and off

This capability does not rely on an out-of-the-box, generalized foundation model; it’s trained strictly on a specific customer's own historical simulation data assets. To build a functional surrogate, an enterprise feeds Graph Neural Networks (GNNs) thousands of its own historical, high-fidelity Computer-Aided Engineering (CAE) meshes for a highly specific product architecture (e.g., past automotive chassis or aerospace wing designs).

Track B: The Industrial Foundation Model initiative

Entirely separate from localized design surrogates is the multiyear campaign to construct the sector's true cross-domain Design-to-Manufacturing Frontier Model. Leveraging massive, global ecosystem datasets (such as a 150-petabyte dataset), the objective of the IFM isn't to calculate localized mechanical stress.

It’s engineered to interpret a 3D CAD model, understand its component dependencies, and programmatically generate the corresponding manufacturing bills of materials, factory floor routing rules, and manufacturing execution system configuration states required to build the physical asset.

Asynchronous routing vs. the deterministic control loop

To properly guide technology deployment, a critical technical nuance must be enforced regarding open connectivity frameworks like Anthropic's Model Context Protocol. MCP represents a profound advancement for data integration, operating as an open-standard semantic routing and information discovery engine across multi-vendor enterprise software estates.

MCP allows an intelligent agent sitting inside a product lifecycle management suite to seamlessly query, locate, and retrieve context from an asset file locked inside an external ERP or third-party database without requiring a customized, proprietary API connector.

See also: Report sees manufacturers split on whether to prioritize industrial AI

However, an analytical hard-stop must be drawn regarding control architectures. While MCP is an exceptional tool for cross-system querying, it remains an asynchronous information protocol. It natively lacks the deterministic, sub-millisecond capability required to control a spinning kinetic asset safely.

Hard shop-floor control loop optimization can never run via an open internet query framework. It demands edge-native orchestration, utilizing containerized virtual PLCs and ruggedized edge hardware acting directly next to the machine, entirely insulated from high-level enterprise agent traffic.

Actionable takeaways for industrial operators

• Reject the low-code obsolescence fallacy: Maintain and strengthen low-code application layers (like Mendix) to serve as mandatory visual safety rails, risk sandboxes, and human-in-the-loop validation canvases for automated agent outputs.
• Isolate control lines from semantic graphs: Rigidly enforce an architecture that isolates asynchronous information routing protocols (like MCP and i3X) from sub-millisecond edge networks running virtualized PLCs.
• Decouple custom MES extensions: Resolve the 30-year upgrade deadlock of custom-code hell by deploying plant-specific routing configurations as extended apps within managed sandboxed containers, keeping the core MES source code standardized.

Read the original article at ARC Advisory Group or continue the conversation with the author on his LinkedIn.