Seven Principles for Trustworthy AI in Industrial Operations

In June 2026, GE Vernova published something unusual for a process automation vendor: a long-form piece aimed at system integrators, not executives shopping for software licenses.

The unusual part was that it was about how SIs need to embrace AI not only to make themselves more efficient, but to make their clients more efficient.

The article quotes Rain Engineering, InflexionPoint, INS, Kline Process Systems, and Automatech on how AI is changing SI economics, governance, and recurring services.

They are mostly right. AI is not replacing skilled integrators but it is changing what buyers expect from them.

Data readiness is the bottleneck, not a footnote.

Governance matters as much as model accuracy.

What the article does not say (because it cannot, given Proficy's place in the stack) is that most of these outcomes assume a field and control layer that is already coherent, separable, and trustworthy. When L1 through L3 is one vendor's monolith, you can still sell AI-on-MES services. You just cannot enforce the architecture those services require.

Consider the baseline: a TATA/AWS survey found only 21 percent of manufacturers are fully prepared for AI adoption (despite 75 percent expecting AI to be a top three driver of operating margin this year). The ambition is real. The foundation is not.

This piece offers seven principles for buyers, digital sponsors, and system integrators evaluating industrial AI (with third-party evidence, not vendor slogans) so that you can make better decisions.

Through-line: You cannot run trustworthy AI on a DCS you never made coherent and separable.

1. Ground truth before models

Bob Clark of INS puts the bottleneck plainly:

"If you want to use AI, you not only need a ton of data — you need it organized and contextualized. That's a big opportunity for SIs."

GE's own narrative reinforces it: AI does not solve siloed systems, inconsistent models, and poor data quality. It exposes them.

Third-party research adds harder numbers.

McKinsey reports that many AI use cases in manufacturing are shelved for years while organizations wait for data quality problems to resolve themselves. KPMG's 2026 Global Tech Report finds a paradox: 83% of leaders say their AI data foundations are strong, yet 76% cite insufficiently reliable data as a top AI risk over the next two years. RSM's middle-market survey puts 51% of manufacturers naming data quality as a primary barrier. And Gartner identifies poor data quality as the single most common reason AI projects fail.

The next phase is not collecting more tags. As process industries move toward causal and explainable models, the goal is trusted operational intelligence (data an engineer can validate, not a dashboard that merely correlates).

That work starts in the field. Conforming semantics at source (on an O-PAS-aligned, separable control stack) beats perpetual tag-mapping projects every time an enterprise AI initiative spins up.

2. Plant before "AI on MES"

GE and its SI contributors align on embedded AI: the highest-value use cases sit inside operational workflows (MES, SCADA, maintenance), not in standalone analytics portals. Don Rahrig of Rain Engineering describes designing "AI on top of MES" services and building recurring revenue around monitoring, tuning, and continuous improvement.

The delivery model makes sense. The implied dependency on MES vendor lock-in does not.

Bain's hourglass framing is useful: enterprise AI and smart field devices both need a control middle that feeds them honestly. In process industries, prediction alone is insufficient (engineers need causal understanding they can defend in a safety review). AI embedded in MES on top of a closed DCS still inherits that DCS's alarm model, tag semantics, and lifecycle constraints.

Open L1-L2 with a canonical field model lets integrators deliver the same recurring optimization services without being captive to one MES estate. The SI's domain judgment becomes the moat, not the vendor runtime underneath.

3. Governance is architecture, not prompt hygiene

Chris Monchinski's governance warning is worth reading carefully:

"AI-generated output can be fast, but not necessarily correct, safe, or compliant ... governance becomes a critical differentiator."

He adds the triad every OT leader should keep in mind: firms that resist AI will struggle; firms that embrace it without governance will create risk; firms that succeed will blend AI acceleration with engineering discipline.

GE frames much of this as validating model outputs. That is necessary. It is not sufficient.

Sunil Doddi, writing for ISA and Control Global in 2026, argues that process plants cannot "just give full license and autonomy to AI agents." Safety, cybersecurity guardrails, and predefined governance must exist before agents touch live processes. Wrong LLM output in a process plant can damage products and equipment, not just embarrass a brand team.

The field exposure is not hypothetical. The Dragos 2025 OT cybersecurity review found 75 percent of OT attacks begin as IT breaches, and 70 percent of identified vulnerabilities sit deep within OT networks (making them difficult to patch without disrupting operations). Only 20 percent of organizations have a fully collaborative IT/OT security posture, according to Cisco's 2026 State of Industrial AI report. Most are trying to bolt security onto a layer that was never designed to be separated.

The Purdue Model (levels 0 through 5, with a DMZ between OT and enterprise IT) has taught segmentation for decades. IEC 62443 and O-PAS SEC-F-001 translate that into implementable security requirements. The problem is architectural: when your DCS is one vendor's black box, Purdue becomes a diagram in a proposal, not boundaries you can enforce at runtime.

CISA's June 2026 advisory on internet-exposed automatic tank gauge systems makes it concrete. Attackers executed commands as if they were at the operator console, not in a corporate LLM chat window. Bolt-on network security failed because the architecture exposed L0 before AI entered the picture.

You cannot segment what you cannot separate.

Separation is the foundation of OT security covers that tank-gauge failure in detail. The lesson for AI is the same: governance starts in the stack, not in the prompt.

4. The SI inflection: orchestrator, not assembler

GE Vernova's bottom line is blunt: system integrators are evolving from builders to orchestrators of intelligence. Chris Monchinski of InflexionPoint puts it directly: the role shifts from manually generating everything to "supervising, governing, validating, and orchestrating increasingly autonomous systems."

Brian Beitler of Kline Process Systems is equally practical. He is not worried about AI replacing SIs until it can ingest a P&ID and create PLC code and HMI screens on its own. Don't worry, someone still has to generate the design.

That tracks with Bain & Company's 2026 research on industrial automation, which describes an hourglass economy: value migrating toward software, data, and AI at the top and smart field devices at the bottom, with the control middle still essential but differentiated by who can orchestrate intelligence across the stack. Integrators stay in the winning ecosystem. But the stack they orchestrate determines whether they are governing an open plant or babysitting for a closed runtime.

The question for your next RFP: If the SI role is upgrading, what foundations upgrade with it?

5. Efficiency is table stakes

Bob Clark captures the SI economics shift in one line:

"We can be more efficient — but so can our customers."

Short term, AI-assisted engineering is a competitive advantage. Medium term, buyers expect faster delivery as standard. Bill Kapusta of Automatech notes that AI frees engineers to spend more time on architecture, which is exactly where differentiation moves once Copilot-class tools are everywhere.

The scaling gap tells you something important: IBM's 2025 survey found only 16 percent of AI initiatives scaled enterprise wide. Speed of documentation is not the bottleneck. Foundation is.

When every team has the same generative AI assistants, what remains is a trustworthy OT data model, multi-vendor lifecycle, and certified open control (things a closed-stack SI cannot replicate by prompting faster).

Do not lead with "we have AI too." Lead with what your stack still allows you to change in ten years.

6. Complement enterprise AI, do not fight the corporate layer

Enterprise AI spend is real and accelerating. Bain projects that by 2030 roughly half of industrial automation revenue could rely on AI-enabled offerings, with substantial value migrating to the intelligence layer. RSM finds 88% of manufacturers say AI is partially or fully integrated into operations.

Your CIO's analytics platform, corporate copilots, and MES AI modules are not the enemy. They are downstream consumers. KPMG calls standardizing, connecting, and governing OT/IT device data a no-regrets initiative that supports AI use cases, not a science project.

Position the field as foundation: enterprise agents need ground truth from a coherent OT layer, not another middleware layer that normalizes vendor-shaped tags every quarter. Proficy, C3, and cloud analytics complement open field architecture when that architecture exists. They do not replace separable control.

7. Full stack open, or lock-in moves up a floor

GE's AI story routes value through Proficy and existing MES estates (implicitly, a stack that is already chosen).

The market is now adding an open layer at L3.

CESMII's i3X 1.0 release delivers an open common API with conformance tests for manufacturing information platforms (Imperative #3 of smart manufacturing architecture, done with the same seriousness O-PAS brings to control).

That is good news. It is not the whole stack.

i3X opens how apps talk to contextualized manufacturing data. O-PAS opens how the plant controls and connects at L0-L2 (certified profiles, separable hardware and software, security baseline). O-PAS field deployments at major operators show that open control is not theoretical.

These standards are complementary, not either/or:

• O-PAS: how the plant controls and connects (L0-L2)
• i3X: how software queries contextualized manufacturing information (L3+)

If only the API layer opens while L1 remains a single-vendor DCS, portable manufacturing apps still read vendor-shaped tags through a nicer interface. Segmentation stays on a slide. Lock-in moves up to the next level.

The full stack must be open, or you just relocated the prison.

Checklist for your next evaluation

Use these seven principles when you select an SI, write an RFP, or sponsor a digital initiative:

• Ground truth before models: Is data organized and contextualized at the point of collection, or forever integrated downstream?
• Field before AI-on-MES: Are embedded workflows built on separable control, or MES lock-in?
• Governance = architecture: Can you enforce zones at runtime, not just review LLM outputs?
• SI as orchestrator: Does your integrator govern a stack they can still change?
• Efficiency is table stakes: What is the moat after everyone has Copilot?
• Complement enterprise AI: Is OT foundation work funded alongside corporate AI?
• Full stack open: Which layers are open with conformance tests (and which are marketing)?

What to do next

Industrial AI is not a copilot slide and not a Proficy upsell. It is a stack decision.

If you are evaluating open automation, start by asking your SI partner which Purdue layers they can actually separate in the runtime they propose.

Share this with your OT architect and CISO in the same meeting. That is where this conversation belongs, and it is how you stay aligned.