Industrial AI is not stalling because manufacturers lack algorithms. It stalls because OT data is still fragmented, stripped of meaning, or lacking clear ownership. Across recent industrial reporting, the message is consistent: AI is moving closer to production use, but only where operational data is trustworthy, usable, and securely accessible (Smart Industry, Dec. 26, 2025; IndustryWeek, Mar. 27, 2026).
For most manufacturers, the order of work is clearer than the marketing suggests:
- Define one operational problem worth solving.
- Make sure the needed OT data carries enough meaning to be understood outside the controller or historian.
- Create a secure OT-to-AI path that does not disturb hard real-time control.
- Put governance around data access, model use, and any action an AI system can trigger.
Miss any one of those steps and the pilot may look good in a demo, then stall when it reaches live production.
Start with the use case, not the tool
Manufacturers often say “AI” when they mean very different things. A generative assistant that summarizes manuals or writes draft text is not the same as an engineering agent built for automation workflows. Agentic AI, in turn, is aimed at coordinating multistep software tasks (Automation.com, Apr. 20, 2026; Control Global, Apr. 15, 2026, “Setting up for agentic AI”).
Plant autonomy is broader than all three; it depends on instrumentation, data flow, and control design as much as on AI itself (IndustryWeek, Mar. 27, 2026).
So the first decision is not vendor selection. It is use-case selection. Pick one job with a measurable outcome: reduce chronic unplanned stops on a line, shorten the time needed to find the right maintenance procedure, improve abnormal-condition detection, or give engineers faster access to approved project knowledge. Those starting points are easier to justify because they tie back to downtime, scrap, energy, or labor rather than to a vague promise of “AI transformation.”
This first step also sets a safe boundary. In most facilities, the early wins sit above the controller—in maintenance, quality, engineering, energy, planning, or supervisory decisions—not in scan-cycle control. Hard real-time control means actions that must happen predictably at the controller. That layer should remain local while manufacturers prove out data quality, validation rules, and human approval paths.
Give the data meaning before you ask AI to interpret it
Raw tags are not enough. AI needs to know what a signal belongs to, when it was captured, what operating state the equipment was in, and whether the signal can be trusted. Recent manufacturing coverage keeps linking AI to data fabric, instrumentation, and unified namespace architectures rather than to models alone (IndustryWeek, Mar. 27, 2026; Machine Design, Mar. 2, 2026).
The labels vary, but they are not interchangeable. A unified namespace, or UNS, is one way to organize and publish plant events and states in a consistent structure. An asset model defines what equipment, tags, states, and relationships those events represent. A data fabric is the broader approach used to make governed information accessible across systems. The context layer is the practical outcome: data that still knows which asset, batch, recipe, SKU, shift, or alarm state it belongs to (Machine Design, Mar. 2, 2026; IndustryWeek, Mar. 27, 2026).
That point sounds abstract until it hits a live line. A reject count without part number, machine mode, and changeover status is only a number. A temperature trend without batch phase, lab result, and maintenance history is only a trace. Process-industry teams are reaching the same conclusion: AI becomes far more useful when manufacturers reduce data friction and preserve operating meaning across systems (Control Global, Apr. 15, 2026, “Coordinate convergence and calm complexity”).
Instrumentation belongs here too. If sensors drift, timestamps do not align, or sampling practices vary, the problem starts before analytics do. IndustryWeek placed instrumentation among the four fundamentals on the path to autonomous manufacturing for exactly that reason (IndustryWeek, Mar. 27, 2026).
Modernize the connections without disturbing stable control
Once the data has meaning, it still has to move. Interoperability is no longer a side discussion for standards groups; it is a production requirement when AI, historians, engineering tools, maintenance systems, and business applications all need consistent information (Automation World, Apr. 9, 2026).
For many facilities—especially those with long-lived PLC, DCS, and SCADA assets—the sensible path is staged modernization. Add edge connectivity where needed. Normalize interfaces. Clean up naming standards and handoff rules. Then extend access to higher-level applications. That is usually safer than disturbing a stable control environment just to satisfy a software roadmap (Automation World, Feb. 23, 2026; ISA, Feb. 24, 2026).
ISA’s new ISA-112 standard for SCADA systems is a useful reminder that lifecycle discipline still matters as older supervisory environments are connected to newer digital workflows (ISA, Feb. 24, 2026). The same principle applies in discrete manufacturing. A robot cell, vision system, and packaging line do not become AI-ready just because they can each export data. They have to describe equipment states and events in a way the rest of the enterprise can actually use.
Security has to be designed into that path from the start. Feeding OT data to AI depends as much on deployment design as on protocol choice. Network segmentation, identity controls, logging, and clear boundaries between control and higher-level applications matter before the first model is put into service (Automation.com, Feb. 16, 2026).
Governance is what turns a pilot into a repeatable practice
Data and integration problems are visible. Governance failures are quieter, but they are often what stop scale-up. Governance is not just a cyber policy or an IT steering committee. It is the set of operating rules around who may expose data, change naming standards, approve models, validate outputs, retain records, and authorize any action that could affect production.
That matters because AI programs cut across controls, operations, engineering, cybersecurity, and enterprise data teams. When IT and OT do not agree on ownership, even a technically sound pilot tends to stall in review, remain trapped in pilot mode, or get used informally without clear accountability (Automation World, Feb. 23, 2026).
At a minimum, a site should know who owns tag naming and asset models, who approves a new OT data feed, and what testing is required before an engineering assistant is used on a live project. It should also know whether prompts, outputs, or generated code are retained, who must approve an AI-generated work order or quality hold, and what the rollback path is if the recommendation is wrong.
Those decisions sit inside a broader cybersecurity challenge. As IT and OT monitoring converge, industrial teams are being asked to expose more operational data while defending older environments that were never designed for this level of connectivity (Industrial Cyber, Feb. 8, 2026). That is another reason human approval should remain the default whenever AI crosses from insight into action.
A quick plant example
Consider a packaging line that wants to use AI to cut recurring stops. The line has PLC alarms, robot-cell events, vision rejects, and maintenance notes, but each sits in a different system. If the team simply points an AI assistant at the historian, it may get a clean summary of alarms, yet it still cannot tell the difference between a feeder fault, a changeover loss, and a bad film lot.
The project becomes useful when the team first standardizes machine states, links reject events to SKU and shift, adds recent maintenance history, and exposes the data through a governed edge layer. Now the model has a much better chance to identify repeatable failure patterns and guide technicians toward the right response.
A process example follows the same logic. A site investigating abnormal conditions on a blending unit needs DCS trends, operating mode, lab results, batch timing, and work-order history tied together before AI advice will mean much to an operator or process engineer.
A practical rollout sequence
Manufacturers that want plant-floor value from AI do not need a grand program first. They need an ordered one:
- Choose one use case and one business metric.
- Map the required OT data sources and the owner of each source.
- Add the missing meaning: asset, state, batch or SKU, time, and maintenance history.
- Create a secure, segmented OT-to-AI path.
- Define governance before scaling: access rights, validation steps, retention, approvals, audit trail, and rollback.
- Expand through staged modernization while keeping hard real-time control local.
That sequence is not flashy. It is, however, what turns AI from a promising demo into something engineers, operators, and maintenance teams can use with confidence.
Industrial AI will not be won by the manufacturer with the loudest AI message. It will be won by the one that can present trusted operational data, clear ownership, and safe deployment rules when a real production problem needs to be solved. That groundwork is less glamorous than a new model announcement, but it is what makes AI useful on a line, in a process unit, and eventually across a site.
Sources consulted
- Automate (A3). “Robot Orders Grow 6.6% in 2025 as General Industries Drive Broader Automation Adoption.” Feb. 6, 2026.
- https://www.automate.org/robotics/news/robot-orders-grow-6-6-in-2025-as-general-industries-drive-broader-automation-adoption
- Automation.com. “Securely Feed Your OT Data to AI.” Feb. 16, 2026.
- https://www.automation.com/article/securely-feed-your-ot-data-ai
- Automation.com. “Siemens Launches the Eigen Engineering Agent, Bringing Purpose-Built AI to Industrial Automation Engineering.” Apr. 20, 2026.
- https://www.automation.com/article/siemens-eigen-engineering-agent-purpose-built-ai-industrial-automation-engineering
- Automation World. “How i3X Looks to Address Industry’s Interoperability Issues.” Apr. 9, 2026.
- https://www.automationworld.com/factory/digital-transformation/article/55368446/how-i3x-looks-to-address-industrys-interoperability-issues
- Automation World. “The Hidden Risks of IT/OT Misalignment.” Feb. 23, 2026.
- https://www.automationworld.com/process/digital-transformation/article/55358958/control-system-integrators-association-csia-the-hidden-risks-of-it-ot-misalignment
- Control Global. “Coordinate convergence and calm complexity.” Apr. 15, 2026.
- https://www.controlglobal.com/control/ai-ml/article/55371091/highbyte-coordinates-convergence-and-calm-complexity-with-amazon-web-services
- Control Global. “Setting up for agentic AI.” Apr. 15, 2026.
- https://www.controlglobal.com/control/ai-ml/article/55371089/setting-up-for-agentic-ai-processing-industries-tackle-artificial-intelligence
- IndustryWeek. “4 Fundamentals on the Path to Autonomous Manufacturing.” Mar. 27, 2026.
- https://www.industryweek.com/technology-and-iiot/automation/article/55366866/4-fundamentals-on-the-path-to-autonomous-manufacturing
- Industrial Cyber. “Rethinking next-generation OT SOC as IT/OT convergence reshapes industrial cyber defense.” Feb. 8, 2026.
- https://industrialcyber.co/features/rethinking-next-generation-ot-soc-as-it-ot-convergence-reshapes-industrial-cyber-defense/
- ISA. “ISA Announces Publication of New Standard for SCADA Systems.” Feb. 24, 2026.
- https://www.isa.org/news-press-releases/2026/february/isa-announces-publication-of-new-standard-for-scad
- Machine Design. “The Future of Smart Manufacturing Relies on a UNS.” Mar. 2, 2026.
- https://www.machinedesign.com/automation-iiot/article/55361007/highbyte-the-future-of-smart-manufacturing-relies-on-a-uns
- Smart Industry. “Crystal Ball 2026: The year AI moves from promise to production.” Dec. 26, 2025.
- https://www.smartindustry.com/special-reports/article/55337079/crystal-ball-2026-the-year-ai-moves-from-promise-to-production