Industrial AI is entering its less glamorous phase, and that is a good sign.
The easy part is getting approval for a pilot. The hard part is making the result survive contact with a live plant. For most manufacturers, AI moves beyond the pilot stage only after five basics are addressed: usable data with asset and process context; a repeatable way to connect brownfield equipment; workflow integration inside maintenance, quality, or operations; clear OT/IT ownership; and cyber and change-control rules from the start. Smart Industry’s recent reporting on a widening AI execution gap and manufacturing infrastructure strain is a useful shorthand for the problem, while Control Engineering, Control Global, and Automation.com point to the same prerequisites from lifecycle, virtualization, and security angles (Smart Industry, Cisco AI execution-gap study, 2026; Control Engineering, lifecycle-ready AI, 2026; Control Global, virtualization prepares for AI, 2026; Automation.com, human side of industrial cybersecurity and ISA/IEC 62443, 2026).
For plant teams, the sequence is usually straightforward:
- Pick one costly loss.
- Map the data path that explains it.
- Fix the connection points that will be reused.
- Put the output into daily work.
- Set security and approval rules before expanding.
The rest is execution.
Start with the data path, not the model
Most stalled deployments run into trouble before the model is the main issue. The plant knows a line is underperforming, quality is drifting, or a machine is failing unpredictably, but the supporting data is scattered across PLC tags, historian records, MES events, quality systems, and maintenance history. Until those pieces line up, the system cannot answer the questions operations actually care about: which asset, under what conditions, in which batch or order, and what action should follow? Control Engineering’s case for lifecycle-ready AI and Machine Design’s reporting on AI tied to MES in MedTech production point to the same lesson: context is what turns a pattern into a decision (Control Engineering, lifecycle-ready AI, 2026; Machine Design, AI and MES in MedTech production, 2026).
That is why so many first AI projects become data-cleanup projects in disguise. An anomaly score without an asset hierarchy is just another alarm. A quality model without recipe or lot context is hard to trust. Operators do not need a lesson in machine learning; they need to know what equipment is affected, how confident the system is, and what they are supposed to do next. Automation World’s look at AI-enabled ERP for mid-sized manufacturers reinforces the same point higher up the stack: the value improves when operational and enterprise data are connected instead of analyzed in isolation (Automation World, AI-enabled ERP for mid-sized manufacturers, 2026).
A useful test is simple: if the plant cannot explain where the data came from, what operating state it represents, and how it maps to the asset or process in question, it is too early to talk about scaling AI.
Brownfield connectivity is what makes AI repeatable
Plants do not need to replace every controller to become AI-ready. They do need a repeatable way to get data and decisions across older PLCs, skids, DCS nodes, and proprietary networks. The first connection is rarely the problem; the fifth one is.
That is why brownfield modernization now belongs inside the AI discussion. Control Global’s recent article on virtualization framed the point well: virtualized engineering and test environments give teams a safer place to validate interfaces, patches, and application changes before they touch live operations. That matters because it lowers the risk of introducing new software into running production and makes the next deployment easier than the last one (Control Global, virtualization prepares for AI, 2026).
This is the real cause-and-effect link between modernization and AI. Gateways, protocol-conversion layers, standardized asset models, and virtualized environments are not side projects. They are what keep each new use case from turning into a custom integration exercise. Smart Industry’s report on AI-related infrastructure strain makes the same problem visible from a broader manufacturing angle: demand for AI is now outrunning the condition of the underlying plant architecture (Smart Industry, Cisco AI execution-gap study, 2026).
If every AI project requires a fresh workaround for legacy connectivity, scaling will stay slow and expensive. If the plant builds reusable connection points, the second and third use cases become much easier to justify.
AI has to land in work, not in another dashboard
A model can be technically impressive and still fail operationally. If the output stops at a dashboard, the plant has learned something, but it may not have changed anything.
Control Engineering’s lifecycle-ready AI article and Control Global’s recent piece on industrial orchestration both argue that industrial value comes from embedding intelligence into engineering, operations, and maintenance workflows rather than bolting it on as a separate analytics layer (Control Engineering, lifecycle-ready AI, 2026; Control Global, industrial orchestration for resilience, 2026). Machine Design’s MedTech example shows the same pattern from a different angle: AI becomes more useful when it is tied to MES and production records that already govern traceability, exceptions, and response steps (Machine Design, AI and MES in MedTech production, 2026).
That is the right test for any plant proposal. What changes on a shift? Does the output create a work order, change an inspection plan, guide an operator check, or speed up root-cause review? If the answer is no, the project may still be interesting, but it is unlikely to become a durable operating tool.
This is also where many “good demos” stall. A model can detect a pattern, but unless someone has decided how that pattern enters standard work, the insight dies between shifts. Plants do not get sustained value from isolated analytics. They get value from changed decisions.
Clear OT/IT ownership determines whether the second use case happens
Many plants can get one AI pilot running through sheer effort. Expanding beyond that is harder because AI sits across controls, networking, data platforms, edge or cloud compute, and business systems at the same time. Smart Industry’s infrastructure-focused reporting and Automation World’s AI-enabled ERP example both show how quickly these projects cross traditional boundaries (Smart Industry, Cisco AI execution-gap study, 2026; Automation World, AI-enabled ERP for mid-sized manufacturers, 2026).
The practical questions are familiar. Where will inference run: at the machine, on-site, or in the cloud? Who owns the data pipeline? Who approves model updates? How is a bad recommendation overridden? Who supports the system at 2 a.m. when a feed breaks or a connector fails?
Until those answers are explicit, the first deployment remains a special project. The second deployment becomes an ownership dispute. Controls teams understand deterministic behavior, process constraints, and uptime risk. IT and data teams understand identity, storage, platforms, and lifecycle management. Industrial AI needs both. Treating either side as an afterthought is one of the fastest ways to turn a promising pilot into an orphaned system.
Cybersecurity and model control are part of the deployment
AI increases exposure in ordinary industrial ways: more data movement, more software, more remote connectivity, and more dependence on identity and access control. Industrial Cyber recently warned that AI is accelerating OT threats, and Automation.com highlighted why many industrial cybersecurity programs still align with ISA/IEC 62443 principles such as segmentation, controlled remote access, and defined accountability (Industrial Cyber, AI accelerates OT threats, 2026; Automation.com, human side of industrial cybersecurity and ISA/IEC 62443, 2026).
AI adds another layer that plants cannot leave vague: model control. Who can use production data to train or tune a model? Who signs off on a change? Which outputs are advisory, and which can influence operations? How is rollback handled if the result is wrong? ISA’s recent announcement of a position paper on industrial AI is a sign that these questions are moving into mainstream automation practice, not staying in research circles (ISA, industrial AI position paper announcement, November 2025).
This is not a reason to slow every project to a crawl. It is a reason to treat AI like any other operational technology that can affect uptime, quality, or safety: with change control, auditability, and a clear human fallback. If operators cannot tell who approved a model change, when it changed, or how to fall back to the previous method, trust will erode quickly.
A practical sequence for plant teams
For most manufacturers, the order of work matters more than the perfect software stack.
1. Start with one measurable loss. Tie the project to a visible problem such as unplanned downtime on a constrained asset, recurring scrap, long troubleshooting cycles, or excessive energy use. A plant problem should come first; the AI tool should come second.
2. Trace the data path end to end. Confirm which signals exist, where context lives, what is missing, and who owns each system. That exercise usually reveals whether the real problem is model performance or data quality and integration.
3. Fix the connection points that can be reused. Invest in the gateways, interfaces, segmentation, virtualization, and asset models that the next use case can share. Reusable plumbing matters more than a flashy first demo (Control Global, virtualization prepares for AI, 2026; Smart Industry, Cisco AI execution-gap study, 2026).
4. Define the workflow before deployment. Decide who sees the output, what action follows, what escalation path exists, and how a recommendation can be challenged or overridden. If nobody owns the response, the project is incomplete (Control Engineering, lifecycle-ready AI, 2026; Control Global, industrial orchestration for resilience, 2026).
5. Lock in security and change control before expansion. Set remote-access rules, approval steps, logging, rollback procedures, and support ownership before the system spreads across lines or sites. AI does not need a separate rulebook, but it does need the same discipline any connected operational system should already have (Automation.com, human side of industrial cybersecurity and ISA/IEC 62443, 2026; Industrial Cyber, AI accelerates OT threats, 2026).
Practical takeaway
Industrial AI has reached the stage where the hard work is less about finding another demo and more about industrializing the basics. That is not disappointing. It is what progress looks like. Smart Industry’s 2026 outlook described AI as moving from promise to production, and the supporting articles on lifecycle integration, virtualization, and OT security all suggest the same thing: plants are now being judged less by whether they can launch an AI pilot and more by whether they can make one reliable, secure, and useful in daily operations (Smart Industry, Crystal Ball 2026, 2026; Control Engineering, lifecycle-ready AI, 2026; Control Global, virtualization prepares for AI, 2026; Automation.com, human side of industrial cybersecurity and ISA/IEC 62443, 2026).
Plants that fix data context, brownfield connectivity, workflow fit, ownership, and cybersecurity will not just run better pilots. They will be in a much better position to turn AI into a repeatable operating capability.
Sources consulted
- Automation World, “AI-Enabled ERP Is Closing the Automation Gap for Mid-Sized Manufacturers” — https://www.automationworld.com/factory/digital-transformation/article/55355640/sap-ai-enabled-erp-is-closing-the-automation-gap-for-mid-sized-manufacturers
- Control Engineering, “Lifecycle-ready AI: unlocking value at every stage of process manufacturing” — https://www.controleng.com/lifecycle-ready-ai-unlocking-value-at-every-stage-of-process-manufacturing/
- Smart Industry, “New Cisco AI study sees widening execution gap, strain on manufacturing infrastructure” — https://www.smartindustry.com/industry-news/news/55361036/new-cisco-ai-study-sees-widening-execution-gap-strain-on-manufacturing-infrastructure
- Smart Industry, “Crystal Ball 2026: The year AI moves from promise to production” — https://www.smartindustry.com/special-reports/article/55337079/crystal-ball-2026-the-year-ai-moves-from-promise-to-production
- ISA, “ISA Explores Industrial AI’s Impact on Automation in a New Position Paper” — https://www.isa.org/news-press-releases/2025/november/isa-explores-industrial-ai-s-impact-on-automation
- Machine Design, “MD&M West 2026 Booth Briefing—How AI and MES Are Transforming MedTech Production” — https://www.machinedesign.com/automation-iiot/article/55353331/critical-manufacturing-mdm-west-2026-booth-briefinghow-ai-and-mes-are-transforming-medtech-production
- Control Global, “Avoid irrecoverable failure with industrial orchestration” — https://www.controlglobal.com/control/ai-ml/article/55361751/reimagining-industrial-ai-the-critical-role-of-orchestration-for-resilience
- Control Global, “Don’t hesitate to jump into virtualization” — https://www.controlglobal.com/manage/systems-integration/article/55357716/hargrove-controls-automation-prepares-for-ai-with-virtualized-tools
- Automation.com, “The Human Side of Industrial Cybersecurity, and Why Consultants Align With ISA/IEC 62443” — https://www.automation.com/article/human-side-industrial-cybersecurity-why-consultants-isa/iec-62443
- Industrial Cyber, “AI accelerates industrial cyber threats, transforms OT attack landscape to challenge traditional defenses” — https://industrialcyber.co/features/ai-accelerates-industrial-cyber-threats-transforms-ot-attack-landscape-to-challenge-traditional-defenses/