Data Architecture First: The key to successful AI implementation in industry

No More Fiddling with Excel

Many industrial companies sit on a treasure trove of data that remains largely untapped in practice. Most factories have been collecting data for years via machines, sensors, ERP systems, and quality measurements, but turning this into actionable insights often fails.

“Many of our clients are gathering data from their factories and various systems to improve production efficiency,” explains Mark van Lanen, Sales Director at ICT Group. “It’s great to have lots of data—but can people analyse it properly? Is there clear governance? A clear structure? These are recurring questions we hear.”

Carmen Berends, Digital Transformation Business Consultant, identifies two main causes: “On one hand, data may exist, but if it’s siloed within a machine, it takes time to access it. On the other hand, working with large datasets requires different skills than most product engineers are used to. You simply can’t manage it all in Excel anymore.”

The foundation: The right data architecture

Meanwhile, industries are being flooded with potential AI applications—raising a new question: what should you actually do with it? Leendert Mijnders, Manager Business Consultancy Industry, sees a clear shift in the market. “Over the past few years, many data challenges were approached via pilots—often dashboards. With the rise of AI, a more fundamental question is emerging: what data architecture is needed to truly embed AI into the organisation?”

In essence, AI is becoming a catalyst for something more foundational: getting the data architecture right before scaling. “We’ve seen a lot of experimentation with specific applications, but the pilot phase is over,” Mijnders states. “Now it’s about securing those successes, which requires solid architecture.”

Unified Namespace and ISA 95

One of the biggest challenges is unlocking data from legacy systems. “We increasingly work with a Unified Namespace,” says Berends, “a kind of phonebook that tracks where data resides—so you no longer need to connect every system separately.”

Industry is grappling with a patchwork of systems: old measurement systems with incompatible data models, alongside newer MES and ERP solutions. “To integrate them, you ideally work with a generic data model,” explains Mijnders. “We use the Asset Administration Shell, a Smart Industry initiative from Germany. This translates data into an ISA 95 structure based on industrial standards.”

The result? A hierarchical structure that maps out the business organisation. “That context enables easy connections to data applications such as dashboards or AI models.”

Balancing edge and cloud

There is no one-size-fits-all solution for technical infrastructure, say the ICT Group experts. “For me, ensuring data quality is the most important thing,” Berends stresses. “With massive datasets, it’s vital to know that what you think you have is actually what you do have. Whether it’s on-premise or in the cloud is secondary.”

In practice, a hybrid approach tends to work best. “You often see a ‘first layer’ on-site,” says Mijnders. “A first dataset is collected locally, with tools for dashboards and analytics at the edge. Then, for heavier calculations, there's typically a sync to the cloud once an hour.”

Once the data architecture is in place, attention can shift to actual AI applications. “You’ll often see that a first use case works well,” Mijnders notes. “But when trying to scale it and create broader value, it turns into an expensive, complex project—because the foundation is missing. That’s when IT departments start getting headaches.”

Smarter selection and search with AI

Still, there are already successful use cases in industry. “Our Applied Solutions unit has been working with AI and vision technology for some time,” says Mijnders. “At Averis Seeds, smart cameras inspect the quality of potatoes. These visual inspections are now fully automated using an AI model.”

The Averis Seeds case illustrates how image recognition has transformed the selection of new potato varieties. “Only about 1 in 10,000 seedlings makes it to market. Using image recognition instead of human visual inspection has made the process more objective and accurate.”

Another example is intelligent search technology. At Royal IHC, ICT Group developed an AI-powered search tool capable of scanning over a million documents in under two seconds. “Engineers no longer need to spend hours cross-checking systems, which has cut search time by 60%.”

AI assistants for operators

Are there any concrete AI applications that will become indispensable in the next two or three years? “An AI assistant for operators will be commonplace within two years,” predicts Mijnders. “Just like you have a Copilot in Office, there will be an industrial version. It’s already on the roadmap of every major vendor—from Schneider and AVEVA to Siemens.”

Fully autonomous line control, however, remains a challenge. “It’s questionable whether we’ll see major progress soon. If you look at depreciation periods for existing factories, wide-scale adoption isn’t likely in the short term. It might take off with new builds, though.”

Data becomes a common language

The shift to data-driven decision-making may represent the most important cultural transformation of all. “Traditionally, businesses base decisions on experience—a gut feeling,” observes Berends. “But as companies globalise and run more factories, that no longer works. The gut feeling in an Italian factory may differ from the one in the Netherlands. Data becomes the shared language.”

This raises a fundamental question: do you follow what the data says, or what experience suggests? “With data-driven decisions, you’re less dependent on operators who’ve been running the process for decades,” says Mijnders. “These people are becoming increasingly scarce, which makes organisations vulnerable. If set up correctly, data doesn’t lie.”

Taking the First Steps

For companies looking to make better use of their data, the first step is to clarify the business case. “A fully autonomous factory comes with a price tag,” Mijnders warns. “You have to ask: what product are you making, and do you have the margin to recoup that investment? A manually operated machine may be more cost-effective in the short term.”

Van Lanen recommends starting with an MVP (Minimum Viable Product). “Invest as little as possible initially to optimise your process. You don’t need to do everything at once—but make sure any machine you buy is future-ready. For example, choose interfaces that will support later automation.”

“If I could give one piece of advice,” concludes Mijnders, “it’s this: ensure your data architecture is future-proof before investing in AI. Only with the right foundation can you turn pilots into sustainable, scalable solutions that generate real business value.”