Editor’s note: This is the fifth of an occasional 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 other recent articles:

What to do when ‘rip-and-replace’ is financially impossible

How industrial software operationalizes physics-informed digital twins

Industrial AI has transitioned to the ‘application phase’

Navigating financial anxiety around paying the bill for industrial AI

The initial hype cycle of corporate generative AI was dominated by passive conversational applications. For industrial organizations, treating large language models as basic text summarizers or document search tools has hit an immediate economic and operational wall. The paradigm of simply bolting a conversational "Co-pilot button" onto standalone enterprise applications to write basic code snippets or draft emails is officially dead.

See also: New IFS product grabs industrial data for a single, AI-powered interface

This tension is clearly reflected in the numbers: Across the manufacturing base, widespread penetration of paid premium corporate assistant seats remains critically low, estimated at about 3.3% of the commercial base.

Early rollouts stalled because enterprises managed adoption as a linear product feature launch, attempting to drive feature clicks without redesigning the actual workflows the tools inhabit.

To capture real value, the enterprise software ecosystem is being remodeled from isolated application silos into an integrated, multi-agent canvas engineered to observe, contain, and execute automated intent across secure boundaries.

To give these autonomous software agents a stable foundation for logical reasoning, manufacturers must move past disjointed data pipelines and establish a continuous, stateful understanding of organizational context. This is achieved by deploying unified semantic fabrics (such as Microsoft IQ) directly on top of central cloud data repositories like Microsoft Fabric OneLake, mapping operations across four distinct layers:

• Work IQ: Programmatically abstracts and tracks fluid human collaboration, asset scheduling, and business communications across enterprise application workspaces.
• Fabric IQ: Models structural data patterns of physical business operations, real-time machine telemetry streams, industrial IoT signatures, and active digital twin topologies, eliminating historical cross-silo data engineering friction.
• Foundry IQ: Functions as the centralized knowledge registry and routing engine, allowing autonomous execution agents to programmatically discover, evaluate, and orchestrate workflows across diverse foundation models.
• Web IQ: An open-source, model-agnostic, and Model Context Protocol (MCP)-native web search stack engineered to process real-world payloads 2.5x faster to completely eliminate agent query latency.

Operating system-level execution boundaries

For risk-averse OT guardians, deploying an uncoordinated swarm of non-deterministic software agents inside a plant control network introduces severe operational hazards, creating an unmanaged “agent lake.”

Resolving this risk requires weaponizing the legacy operating system footprint powering industrial PCs, human-machine interfaces, and SCADA networks to enforce strict execution boundaries.

See also: How manufacturers are using digital twins to make real-time decisions

This security architecture is driven by native, OS-enforced sandboxing, known as Windows Execution Containers. Operating systems introduce local process isolation directly at the ground level. IT and OT personnel configure containment parameters once; the operating system then strictly enforces local memory limits, CPU core restrictions, and explicit network folder permissions around open-source agent frameworks.

This structurally bars agents from executing unmanaged network traversals or making unauthorized alterations to adjacent plant control systems.

This edge architecture is governed by a centralized administrative cockpit, such as the Agent 365 Control Plane, which extends standard identity registries (Entra) and security loggers (Purview, Defender) to assign a unique digital identity to every agent, providing full lifecycle management and the critical ability to auto-discover "shadow agents" deployed without central IT clearance.

Actionable takeaways for industrial operators

• Redesign workflows upstream: Shift organizational focus away from individual feature clicks and feature-adoption metrics; permanently re-engineer business workflows around autonomous agent capabilities to escape the corporate transformation paradox.
• Enforce local sandboxing: Mandate that all open-source agent frameworks deployed within plant networks execute inside OS-level isolated containers (like MXC) with hard network folder fences to prevent unauthorized machine writes.
• Audit the agent fleet: Leverage centralized agent control planes to maintain a real-time, comprehensive directory of all operational agents, systematically logging and tracking token utilization metrics to eliminate localized shadow agent sprawl.