Editor’s note: This is the second 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 interpret what is happening in the fast-changing manufacturing technology landscape and conversations in the community.

If you evaluate the major technology disclosures of the industrial market in isolation, you see a routine series of vendor software refreshes and hardware product launches.

But when you evaluate these consecutive developments as an assembled sequence, a far more compelling macroeconomic reality emerges: the industrial market has transitioned out of the “infrastructure phase” of AI and into the “application phase.”

See also: Report: Manufacturers split on whether to prioritize industrial AI

We have reached a powerful catalyst—a structural convergence where accelerated hardware primitives, enterprise software canvases, and physical context are beginning to interact, dividing the market along a noticeable operational fault line.

The cohorts of the industrial digital divide

To understand how this operational divide is playing out across actual manufacturing and process operations, we must look directly at the quantitative metrics gathered during ARC Advisory Group's global research initiatives.

In analyzing the final cuts of our Industrial AI, Energy, and Robotics Survey, the data proves that the traditional “fast-follower” strategy—a methodology where conservative industrial organizations could safely wait out a technology cycle before purchasing commoditized software—is effectively dead for productive AI adoption.

Waiting out the technology cycle is an operational dead end because productive AI scales through continuous, compounding context.

While a fast follower hesitates, pacesetters are actively decoupling their core data to run autonomous optimization loops and building massive semantic knowledge graphs.

By the time a follower attempts to buy a commoditized solution off the shelf, their underlying product models and schematics remain trapped in legacy formats, leaving them structurally blind and permanently separated by an unbridgeable digital divide.

Our survey research shows that the global industrial market has fractured into three distinct operational cohorts:

• The pacesetters (12.9%): Elite industry leaders that have stopped treating AI as a siloed IT experiment, successfully decoupling their core data from software applications to deploy autonomous optimization loops across their facilities.
• The mainstream (55.3%): The traveling majority seeking solid operational traction yet remaining largely trapped in basic conversational copilots and retrieval-augmented generation (RAG) document summaries that fail to deliver meaningful value.
• The laggards (31.8%): A trailing cohort feeling increasingly left behind, stalled by legacy technical debt, custom codebase traps, and fragmented data silos.

The innovation paradox and the file system crisis

Moving out of pilot purgatory requires a fundamental change in corporate innovation culture. Laggards manage technology adoption through risk-averse procurement structures, targeting a 0% codebase failure rate, which inevitably strands them on obsolete software branches.

Pacesetters treat AI deployment as a continuous, high-velocity R&D campaign. They carry an intentional 50% project scrap rate on exponential budgets, establishing strict review windows to aggressively kill unscalable pilots early.

See also: Smart manufacturing: Accelerator for semiconductor plant production

This operational variance highlights a critical structural bottleneck on the modern engineering desktop: The File System Crisis.

As disclosed by Siemens at its Realize LIVE Americas conference in Detroit earlier this month, a striking technical metric revealed that 50% of active CAD and product lifecycle engineering users still operate directly off standard local desktop file systems.

If your engineering technology division is attempting to deploy advanced multi-agent workflows or train neural networks while your underlying product models, configuration states, and schematics are locked in unversioned local folders, your automation strategy is structurally blind.

Moving critical lifecycle data off legacy desktop file systems into cloud-managed digital threads is the non-negotiable prerequisite to scale physical intelligence.

The multi-model blueprint of the autonomous factory

Manufacturers that resolve these data bottlenecks are unlocking a highly sophisticated software design pattern that systematically disassembles the rigid layers of the traditional Purdue Model (ISA-95).

Winning industrial AI architectures reject the assumption that a single horizontal large language model can safely manage production operations. Pacesetting organizations are abandoning monolithic systems in favor of open, graph-aware data fabrics that coordinate decentralized networks of specialized, sequential multi-model pipelines:

• Natural Language Operational Intent: Captured at the user interface layer to define the scope of work.
• Generative Foundational Layer (Upfront Filter): Explores high-level concepts and design permutations, compressing the design space from millions of options down to a handful of high-probability candidates.
• Geometric Deep Learning/Physics-Informed Neural Networks (PINNs): Runs a first-principles physics validation engine, mathematically enforcing differential equations to guarantee safety and compliance.
• Edge Virtual PLCs (vPLCs): Translates verified instructions into real-time, closed-loop kinetic actuation directly on the physical machinery line.

The true competitive barrier separating these pacesetting environments from generic IT pilots is the sheer scale of the semantic context layer running behind the model.

Podcast: Ensuring success for your OT-IT convergence

To make AI effective, you cannot rely on manual tag-mapping. It requires a living knowledge graph that auto-discovers assets and contextualizes legacy operational data into relational intelligence.

A general-purpose cloud model understands text strings, but it cannot duplicate a 15-billion-node semantic mesh that tracks the continuous, real-time relationships between physical components, operational tolerances, and enterprise business logs across a global footprint. Actionable takeaways for industrial operators:

• Eradicate legacy local file systems: Make cloud-connected digital thread adoption a board-level KPI; eliminate unversioned local engineering folder trees to provide the clean historical meshes required to train Graph Neural Networks.
• Adopt the pacesetter R&D mindset: Restructure AI funding lines to mimic an intentional 50% project scrap rate, establishing strict evaluation windows to kill unscalable, bespoke pilots within 90 days.
• Mandate data decoupling: Rigidly enforce a strategy of strict data decoupling—supported by 63% of the industrial market—to separate your underlying plant schemas from proprietary software application layers, preserving the flexibility to hot-swap intelligence engines.

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