If industrial AI at your site still depends on a pilot team, ad hoc data cleanup, and one-off approvals, the plant is probably not ready to scale it. In many facilities, the limiting factor is not the model. It is the operating environment around it.
Here, scale means moving beyond a supervised demo or one isolated dashboard and into repeatable use across assets, lines, sites, or day-to-day workflows that affect uptime, quality, throughput, or response time. For brownfield plants with years of installed equipment, that does not mean starting over. Automation World recently argued for augmenting existing automation in phases rather than replacing it outright (Automation World, April 28, 2026). IndustryWeek made a related point in January: AI pays off when workflows change, not when companies simply layer on another piece of software (IndustryWeek, January 15, 2026).
Two practical issues sharpen that message. First, can the plant network support AI workloads? Second, is there a controlled layer between AI output and shop-floor action? Automation.com has focused on both questions in recent coverage of network readiness and industrial orchestration (Automation.com, April 29, 2026; Automation.com, May 1, 2026). Control Global, writing from the process side, likewise argued that plants moving toward more advanced AI need clearer system boundaries and digital discipline first (Control Global, April 15, 2026).
Before you try to scale industrial AI, ask five questions:
- Is the use case a repeatable workflow with a clear owner and measurable value?
- Is the needed data trustworthy in its operating context?
- Can the network and edge layer support the application without disrupting control traffic?
- Is there a rules-based handoff between AI output and plant action?
- Are people, governance, and OT cybersecurity ready for daily use?
- Is the use case a repeatable workflow with a clear owner and measurable value?
The first question is whether the use case is really a workflow or just an AI idea. Automation World’s April 28 article urged manufacturers to target repeatable decision loops and expand in phases (Automation World, April 28, 2026). That is a useful filter. If the plant cannot name the task, the owner, and the success measure, it is too early to talk about scale.
That filter also helps separate very different technologies. An assistive tool might search maintenance procedures, summarize alarms, or pull documentation for a technician. A more advanced, agent-style application might collect data from several systems and prepare a work request or troubleshooting packet for approval. Physical AI applications such as inspection, machine vision, and robotics bring a different risk and infrastructure profile again (Control Global, April 15, 2026; Machine Design, May 4, 2026).
Those categories do not belong in one generic AI bucket. A document assistant can tolerate delay and heavy human review. A vision system on a live line or a robot-supported inspection route cannot. Before the plant spends money on scale, it should be able to say exactly which operating decision is being improved and what “better” looks like.
- Is the needed data trustworthy in its operating context?
Data problems in manufacturing are often context problems. An established facility may have historian tags, SCADA points, CMMS history, MES data, and ERP records, yet still lack a clean way to connect them to the same asset, event, batch, or production state, a challenge reflected in current discussions of workflow redesign and IT/SCADA convergence (IndustryWeek, January 15, 2026; Automation World, April 27, 2026).
One practical question cuts through the noise: which system is authoritative for this decision? If an AI tool is going to help prioritize maintenance, for example, the plant needs consistent asset identity, usable work history, and a shared understanding of what counts as a real operating event.
Smart Industry’s May 5 profile of Wolfspeed offers a useful lesson on scale. The story described an enterprise AI hub supporting 22 agents across key teams, which suggests a shared platform, governance rules, and human review rather than a collection of disconnected pilots (Smart Industry, May 5, 2026). That example reaches beyond a single production line, but the manufacturing takeaway is clear. A maintenance assistant, a quality assistant, and a troubleshooting assistant cannot each run on different asset names, status codes, and permission models if the goal is repeatable use.
- Can the network and edge layer support the application without disrupting control traffic?
In many plants, infrastructure becomes a limit before model capability does. Automation.com’s April 29 article put the warning plainly: the shop floor may be ready for AI, but the network may not be (Automation.com, April 29, 2026). That matters when a use case depends on machine vision, continuous inspection, remote diagnostics, search across large document sets, or data pulled across several systems.
This is not automatically an argument for a plant-wide rip-and-replace. It is an argument for better design. Teams need to know where inference will run, how much data has to move, what latency the application can tolerate, and how AI-related traffic will be kept from interfering with time-critical control communications.
Automation World’s April 27 article on IT and SCADA convergence reinforces that point. As plants expose more operational data and services through software layers, they gain flexibility, but they also need clearer boundaries between IT-style applications and control functions (Automation World, April 27, 2026). The demands also vary by use case. A tool that helps technicians search manuals can tolerate delay and manual review. A vision application or AI-assisted inspection workflow tied to robotics may need local compute, reliable coverage, and tighter performance. Machine Design’s tour of Siemens’ electronics factory in Erlangen showed AI, digital twins, and robotics inside a live production site, which is exactly why connectivity and system design become operating issues, not back-office details (Machine Design, May 4, 2026).
- Is there a rules-based handoff between AI output and plant action?
Between “the model produced an answer” and “something happened in the plant,” there needs to be a controlled step. Automation.com’s May 1 article on the industrial orchestration layer focused on that missing middle: the coordinating logic that turns shop-floor intelligence into governed action (Automation.com, May 1, 2026).
In plain English, orchestration is the rules-based handoff between AI output and execution. It determines where the result goes, what system receives it, which approvals are required, and what record is kept. Conventional control logic is designed for predictable response to validated inputs. AI tools are better treated as systems that generate likely answers, rankings, or recommendations. That difference is why most plants should not let raw AI output touch execution systems without a defined checkpoint (Automation.com, May 1, 2026; Control Global, April 15, 2026).
A maintenance example makes the point. An AI model may flag a pump as a likely failure risk. The next step should not be an automatic control action. A better pattern is to send that finding into a CMMS or maintenance workflow, attach the evidence, and require a planner, reliability engineer, or technician to review it. The same logic applies to quality. An AI system can identify suspect product, but a defined workflow should decide whether material is held, reworked, or released.
That human-in-the-loop boundary does not mean endless manual approval. It means the plant has explicitly decided where a person must review the result before it can affect production, quality, compliance, or safety. Control Global’s April article made the same point from the process side: plants that are better prepared for agent-style AI tend to be the ones that already understand their execution flows and system boundaries (Control Global, April 15, 2026).
- Are people, governance, and OT cybersecurity ready for daily use?
A plant is not ready to scale AI if operators, technicians, and engineers do not know when to trust the tool, when to challenge it, or who makes the final call. IndustryWeek tied AI payoff to changes in work and AI literacy (IndustryWeek, January 15, 2026). Smart Industry added a management warning in May, arguing that IT/OT initiatives often falter when executive involvement stops at sponsorship and never turns into day-to-day ownership (Smart Industry, May 4, 2026).
That ownership has to exist at line level, not just in a steering committee. Someone in operations should own the workflow. Someone in controls should own the system impact. Someone in IT or security should own identities, remote access, and change control. The plant also needs application-specific training: what the tool is allowed to do, what it is not allowed to do, and how users should report a bad output or bad recommendation.
Cybersecurity belongs in the same readiness check. ISA’s overview of the ISA/IEC 62443 series is a useful anchor because it frames industrial cybersecurity across asset owners, system integrators, and product suppliers, not as a standalone IT exercise (ISA, undated). For an AI rollout, that means deciding who approves new data paths, how credentials are managed, how updates are tested, how remote access is limited, and what the fallback plan is if the application is unavailable or wrong.
That is not a theoretical concern. Industrial Cyber reported a significant rise in ransomware pressure on manufacturing in 2025, driven by ransomware-as-a-service activity, legacy OT exposure, and supply-chain weaknesses (Industrial Cyber, April 15, 2026). Every new connector, API, or remote service introduced for AI should be evaluated against that reality.
What scale actually looks like in an existing plant
The most useful lesson from current showcase deployments is not that every manufacturer should copy them. It is that durable programs expand in stages. Automation World’s April 28 article argued for augmentation over replacement (Automation World, April 28, 2026). Smart Industry’s Wolfspeed story showed governed growth on a common AI platform rather than scattered one-off tools (Smart Industry, May 5, 2026). Machine Design’s Erlangen factory tour showed AI, digital twins, and robotics presented as part of a working production environment, not as isolated demonstrations (Machine Design, May 4, 2026).
For a typical brownfield plant, scale usually starts smaller: one workflow, one asset class or line, one named owner, one approved handoff into maintenance, quality, planning, or operator support. Only after that step is trusted do the same data rules, network design, approval logic, training, and security controls get reused elsewhere. That is the difference between accumulating pilots and building something operations can live with.
Practical takeaway
If your team cannot answer the five questions above with named owners, system boundaries, and approved next steps, the plant is probably not ready to scale industrial AI yet. The next investment may be less glamorous than another model pilot: data mapping, network segmentation, edge compute, operator training, or better change control. In an existing facility, that is not a detour. It is the work that turns AI from an interesting demo into a tool the plant can trust.
Sources consulted
- Automation World — “Augment, Don’t Replace: The New Playbook for AI in Industrial Automation”
- https://www.automationworld.com/factory/digital-transformation/article/55372960/augment-dont-replace-the-new-playbook-for-ai-in-industrial-automation
- Automation World — “The Convergence of IT and SCADA: A New Model for Industrial Automation Development”
- https://www.automationworld.com/control/article/55373301/control-system-integrators-association-csia-the-convergence-of-it-and-scada-a-new-model-for-industrial-automation-development
- Automation.com — “The Shop Floor Is Ready for AI. Is Your Network Ready, Too?”
- https://www.automation.com/article/shop-floor-ready-ai-is-network-ready
- Automation.com — “The Industrial Orchestration Layer: Empowering Shop Floor Intelligence”
- https://www.automation.com/article/industrial-orchestration-layer-shop-floor-intelligence
- Smart Industry — “Stories of AI adoption: Wolfspeed all-in with 22 agents across key company teams”
- https://www.smartindustry.com/industry-news/news/55375142/stories-of-ai-adoption-wolfspeed-all-in-with-22-agents-across-key-company-teams
- Smart Industry — “Why IT/OT initiatives fail when executive engagement stops at sponsorship”
- https://www.smartindustry.com/benefits-of-transformation/it-ot-convergence/article/55374953/why-it-ot-initiatives-fail-when-executive-engagement-stops-at-sponsorship
- IndustryWeek — “What Must Change for AI to Pay Off?”
- https://www.industryweek.com/technology-and-iiot/emerging-technologies/article/55343848/what-must-change-for-ai-to-pay-off
- Control Global — “Setting up for agentic AI”
- https://www.controlglobal.com/control/ai-ml/article/55371089/setting-up-for-agentic-ai-processing-industries-tackle-artificial-intelligence
- Machine Design — “Factory Tour: Siemens Electronics Factory in Erlangen, Germany”
- https://www.machinedesign.com/automation-iiot/article/55375265/siemens-factory-tour-siemens-electronics-factory-in-erlangen-germany
- Industrial Cyber — “Manufacturing absorbs 56% ransomware surge of global attacks in 2025, as RaaS, legacy OT, supply chains fuel spike”
- https://industrialcyber.co/manufacturing/manufacturing-absorbs-56-ransomware-surge-of-global-attacks-in-2025-as-raas-legacy-ot-supply-chains-fuel-spike/
- ISA — “ISA/IEC 62443 Series of Standards” (undated; accessed May 6, 2026)
- https://www.isa.org/standards-and-publications/isa-standards/isa-iec-62443-series-of-standards