In many plants, industrial AI does not stumble because the model is weak. It stumbles because the data reaching the model is fragmented, short on operating context, and owned by no one in particular.
Automation World’s July 2026 message was blunt: manufacturers need better industrial data governance before AI, analytics, or dashboards can scale. Smart Industry made a similar case for a data ops foundation, and Automation.com described AI-enabled data management as a missing layer in industrial automation (Automation World, July 13, 2026; Smart Industry, May 27, 2026; Automation.com, February 23, 2026).
That does not make data governance a complete answer. Plants still have to solve workflow design, validation, integration effort, and cybersecurity. But governed, contextualized data is a frequent prerequisite for moving from a promising demo to something operators, engineers, and maintenance teams can use with confidence.
Where pilots break down
A common pattern is easy to recognize. A team proves value on one line, one machine, or one unit using a narrow dataset assembled for the pilot. Then the rollout expands. At that point, the application has to work across live production systems, inconsistent naming, and real operating states.
The same asset may carry one identifier in the programmable logic controller (PLC), another in the human-machine interface (HMI), a third in the manufacturing execution system (MES), and a fourth in the computerized maintenance management system (CMMS). Timestamps do not align cleanly. Downtime codes mean different things by department. A model that looked promising in a controlled pilot can still run, but the output becomes harder to trust.
The failure mode looks different by sector. In discrete manufacturing, an AI application may see slower cycle times without knowing whether the line was in changeover, blocked downstream, starved upstream, or deliberately slowed by the operator. In process industries, an anomaly model may detect a temperature or pressure change without knowing whether the unit was in startup, steady state, cleaning, or maintenance bypass.
Two recent examples help make the point. IndustryWeek’s July 2026 article, which used Vivix as a discrete-manufacturing example, framed better plant data as a broad operations enabler, not a side project for analytics alone (IndustryWeek, July 24, 2026). On the process side, Control Engineering’s June 2026 discussion of data fabric software, including AspenTech-related examples, showed why plants often need a connective data layer before advanced analysis can rely on a shared operating picture (Control Engineering, June 3, 2026).
The five traits of AI-ready plant data
“AI-ready” does not mean every tag in the enterprise is pristine. It means the data for a defined decision is reliable and well contextualized enough to support action. In practice, five traits matter most.
1) Consistent asset identity The same machine, unit, skid, or work cell should carry the same identity across OT and IT systems. If Line 2’s filler is named one way in the PLC, another in MES, and another in maintenance records, every AI use case has to solve the same mapping problem again.
2) Operating context and hierarchy Raw values need business and process context. That usually means linking data to line, area, unit, batch, recipe, product, shift, and operating state as needed. A motor current value has limited meaning if the system cannot tell whether the equipment was starting, running normally, cleaning, or idled for changeover.
3) A clear source of record For each important data type, the plant should know which system is authoritative. In plain terms, “source of record” means deciding where the official production count, downtime reason, lab result, or master asset record lives. Without that decision, every discrepancy turns into an argument.
4) Time alignment and event quality Sensor values, alarms, work orders, quality results, and state changes have to line up on a usable timeline. Otherwise the model may identify correlation where operations sees coincidence.
5) Ownership and change control Someone needs to own tag mappings, naming rules, permissions, and updates after equipment or software changes. A model is only as durable as the process that keeps its inputs accurate after the pilot team leaves.
This is close to what Smart Industry called a “data ops foundation”: the repeatable work of cleaning, mapping, monitoring, and updating industrial data as assets and workflows change (Smart Industry, May 27, 2026). It is also where ISA-95 remains useful. ISA’s 2025 update to the standard addresses integration of enterprise and manufacturing control systems, reinforcing the value of clear relationships between plant systems and business systems even when companies use modern architectures rather than a rigid standards implementation (ISA, April 10, 2025).
Minimum viable governance for a real plant
Data governance can sound like committee language. On the plant floor, it is simpler: who decides what a data point means, where it comes from, and what happens when it changes?
A workable minimum for one AI use case usually includes:
- one business owner for the use case and KPI
- one OT or controls owner for tags, state models, and interface changes
- one maintenance or asset-data owner where equipment structure matters
- one IT or data owner for integration, access, and monitoring
- one agreed source of record for each critical input
Call it data operations, master-data discipline, or interface management; the label matters less than the routine.
That is the unglamorous work behind what Automation.com called the missing layer in industrial automation (Automation.com, February 23, 2026). In plant terms, that layer is not just a software product. It is the combination of metadata, mappings, permissions, and monitoring that keeps data usable after the first proof of concept.
Plants do not need a large enterprise program to start. They do need a way to resolve predictable disputes quickly: Which downtime code list is authoritative? Who updates asset mappings after an OEM machine change? When counts disagree between HMI and MES, which system wins? If no one can answer those questions, the AI application inherits the ambiguity.
Start with one KPI and the smallest useful scope
The practical starting point is a KPI-linked use case, not an all-at-once data cleanup program. Control Global’s July 2026 readiness article asked the right question: is the plant ready for AI, not simply interested in it? (Control Global, July 7, 2026).
For a discrete plant, that may mean focusing on chronic micro-stops on one packaging line. The first phase is not “deploy more AI.” It is:
- standardize the asset name across the PLC, HMI, MES, and maintenance system
- normalize operating-state and downtime codes
- align timestamps
- decide which production count is official
- feed model output into an existing maintenance or continuous-improvement workflow
For a process plant, the first phase might center on one unit operation with quality drift or energy-intensity loss. The corresponding work is to connect historian data, lab results, operating-state information, and maintenance context for that unit before trying to scale site-wide.
This use-case-first approach keeps the scope honest. It also creates value even if the AI model changes later, because the plant has still improved data quality, context, and accountability.
Match architecture to the decision, not the hype
Once the data model is clear, architecture choices become less ideological. Control Engineering’s recent coverage of data fabric software is useful because it points to a practical middle path: connect data across systems without waiting for a full rip-and-replace project (Control Engineering, June 3, 2026). In plant terms, a data fabric is a connective layer that helps teams reach the right data from multiple sources without rebuilding every interface for every use case.
Where AI runs should follow the decision it supports.
- Applications tied to near-real-time operator guidance, machine vision, or tight control windows are often evaluated for edge or on-premises deployment.
- Applications built on heavy historical analysis or cross-site comparison may justify cloud resources.
- In both cases, the design question is the same: what data, model, and context need to be available at the point of use?
This matters even more as physical AI and adaptive robotics attract attention. Recent coverage from A3/Automate and Machine Design has emphasized easier deployment and avoiding last-mile failures, which brings the conversation back to interoperability, trusted data, and application fit rather than model novelty alone (A3/Automate, July 7, 2026; Machine Design, January 6, 2026).
A data foundation is necessary, not sufficient
Even with better data, AI projects can still fail for reasons industrial teams already know well: unclear ownership, poor workflow fit, weak operator trust, unmanaged changes to the machine or process, or new integrations that never pass a proper review.
That is why early deployments should favor decision support over unattended action. A maintenance planner can review an AI-prioritized asset list. An operator can validate an abnormal-condition alert. A process engineer can compare a recommendation against known constraints. A probability score with no operating explanation will not build confidence on the plant floor.
The same caution applies to AI agents that summarize downtime, propose work orders, or surface root-cause leads; without clean event context and a clear approver, they simply move ambiguity faster.
ISA’s November 2025 announcement of a position paper on industrial AI was a reminder that AI is moving into mainstream automation practice, not staying a software side experiment (ISA, November 6, 2025). Once it enters production, it deserves the same change-management, validation, and access-review discipline as any other system that affects operations.
The practical takeaway
Manufacturers do not need perfect enterprise data before they begin using AI. They do need enough structure and accountability to make one real use case dependable.
A solid first-phase plan for the next quarter looks like this:
- choose one KPI-linked use case
- identify the minimum data sources required to improve it
- standardize asset names and operating-state definitions in the pilot area
- assign a source of record for each critical input
- name the owners responsible for updates and exceptions
- decide where the application should run based on latency, access, and workflow
- validate outputs inside an existing operating process before scaling further
When that groundwork is in place, AI starts to behave less like a standalone demo and more like an operating tool inside the automation stack.
Sources consulted
- Automation World — Before AI, Analytics or Dashboards: Manufacturers Need Better Industrial Data Governance (July 13, 2026): https://www.automationworld.com/analytics/blog/55390347/before-ai-analytics-or-dashboards-manufacturers-need-better-industrial-data-governance
- Smart Industry — Why industrial AI requires a data ops foundation to scale (May 27, 2026): https://www.smartindustry.com/artificial-intelligence/article/55379550/why-industrial-ai-requires-a-data-ops-foundation-to-scale
- Automation.com — AI-Enabled Data Management: The Missing Layer in Industrial Automation (February 23, 2026): https://www.automation.com/article/ai-enabled-data-management-missing-layer-industrial-automation
- IndustryWeek — Improving Plant Data Improves Everything (July 24, 2026): https://www.industryweek.com/technology-and-iiot/article/55393207/improving-plant-data-improves-everything
- Control Engineering — Answers, examples: How to connect data with data fabric software (June 3, 2026): https://www.controleng.com/answers-examples-how-to-connect-data-with-data-fabric-software/
- ISA — Update to ISA-95 Standard Addresses Integration of Enterprise and Manufacturing Control Systems (April 10, 2025): https://www.isa.org/news-press-releases/2025/april/update-to-isa-95-standard-addresses-integration-of
- Control Global — How to ensure your plant is ready for artificial intelligence adoption (July 7, 2026): https://www.controlglobal.com/control/ai-ml/article/55389223/how-to-ensure-your-plant-is-ready-for-artificial-intelligence-adoption
- A3 / Automate — How Physical AI Could Make Industrial Robots Easier to Deploy (July 7, 2026): https://www.automate.org/robotics/industry-insights/how-physical-ai-could-make-industrial-robots-easier-to-deploy
- Machine Design — AI Adoption in Manufacturing: Future Tech’s Matt Scavetta on Avoiding Last-Mile Failures (January 6, 2026): https://www.machinedesign.com/automation-iiot/article/55340740/ai-adoption-in-manufacturing-future-techs-matt-scavetta-on-avoiding-last-mile-failures
- ISA — ISA Explores Industrial AI’s Impact on Automation in a New Position Paper (November 6, 2025): https://www.isa.org/news-press-releases/2025/november/isa-explores-industrial-ai-s-impact-on-automation