For years, automation engineers have been measured by what they control: loops tuned, programmable logic controllers (PLCs) programmed, human-machine interfaces (HMIs) built. In 2026, the bottleneck has moved. AI analytics, digital twins, and dashboards are widely available, but data that a machine, or a colleague, can interpret without guessing is scarce. Consider the value “50”: In a plant, it could mean 50°C, 50 Hz, or 50 psi.
This issue of Automation.com Monthly focuses on the industrial internet of things (IIoT) and digital transformation enabled by sound data strategies. With AI agents popping up to assist in every aspect of production, you need a data model control strategy to provide the data the agents need to produce reliable analytics.
In A Solution Architect’s Guide to Driving Manufacturing Insights from AI, authors John S. Rinaldi and David Schultz put it plainly: “For an AI model or any other application to properly consume and utilize data, it must have clear, consistent, and standardized data models.”
Humans infer meaning from context. AI systems do too, and inference introduces error, drift, and unreliable predictions. AI does not create manufacturing insight; it consumes structured information and returns an analysis. If the data is inconsistent or implicitly defined, the results will be unreliable regardless of how advanced the algorithm appears.
Data science fundamentals are now a control skill
You don’t need to become a data scientist, but you do need to understand what makes a model trustworthy, say the authors. That means knowing how units, ranges, quality flags, timestamps, and version history affect the outcome. It also means understanding how training data becomes a prediction and where that prediction can fail.
The best starting point is the decision, not the tag list. The guidance is to begin with the decisions operations must make, then work backward to the values those decisions require. A pump maintenance model, for example, may need cycle counts, run times, energy use, asset identity, and location. An engineer who can frame that model is more valuable than one who only wires I/O.
Structured data strategy pays off across verticals
The strongest case for structured data comes from industries that have already tried to automate without it, and examples are throughout the issue:
Life sciences. Soumen Das argues that intelligent automation is only as reliable as the reference records it validates against. In commercial operations, AI that validates against inconsistent customer or prescriber data “will flag noise or, worse, confidently confirm errors.” The fix is a governed golden record, built by standardizing records first, then deterministically matching records that share authoritative identifiers, and, finally, applying probabilistic matching. The key is creating reliable data for AI to access. The analysis it provides is only as reliable as the records it uses for validation.
Food manufacturing. Andreas Eschbach describes an allergen changeover in which verification tasks, deadlines, and escalation paths live in separate systems, such as spreadsheets, whiteboards, and shift logs. When an overdue verification is found, someone has to notice it and make a phone call. Connecting each task to a line, shift, asset, and standard work turns that gap into a tracked workflow. The same structure enables verifying a corrective action across shifts and developing a repeatable practice that doesn’t disappear when experienced staff moves on. ISA’s new ANSI/ISA-112.00.01-2025 SCADA lifecycle standard supports this kind of modernization.
Process and energy infrastructure. Tom Goff describes a large company building liquefied natural gas (LNG) terminals worldwide. Due to the design of early projects, a site turnover delivered an “enormous data dump” of data in incompatible formats. In a later project, the firm contractually specified data formats and design tools in its requests for quotes (RFQs) and contracts; the result was an integrated platform that is now the company standard for capital projects. Taking data ownership is a procurement step that begins with “explicitly specifying the form and format of the information.”
Standardization beats customization
Naming and units are where good intentions fail. In their guide, Rinaldi and Schultz give the example of “Temp21,” which becomes “Oven 21 Current Temperature” in Fahrenheit, stored as an integer. The guidance is equally specific about naming: “Oven21Temp” and “Oven_21_Temperature” are different tags to a human and an AI model. Plant-to-plant comparison fails when every site calculates overall equipment effectiveness (OEE) differently, and the same logic applies to time conventions and units.
Standards reduce this burden. ISA-95, ISA-88, PackML, OpenAPI, and AsyncAPI give teams a shared vocabulary. The architectural rule is also important: Master data models define meaning, and PLC user-defined types should conform to them. When PLC structures become the de facto master, enterprise scalability breaks down.
Manufacturing’s next gains will not come from the most sophisticated model. They will come from plants that can explain what their data means, who owns it, and how it changed to the AI system asking for it. You don’t need to become a data scientist to do that, but you may need to start thinking like one.
