The graveyard of smart manufacturing projects is filled with perfect, scalable, future-proof monoliths. They are elegant technical cathedrals, engineered on whiteboards to last a decade. The problem is, the ground beneath a manufacturing operation is not solid. It is shifting sand—a living system of legacy machines, undocumented workarounds, and relentless production pressure. Your beautiful architecture will crack unless it is designed for a reality it cannot fully know.
The single most overrated concept in industrial data is static, upfront scalability. The most critical capability is versatility.
The Myth of the Identical Machine
No two production lines are the same. Not even two machines of the same model, purchased on the same day, are truly identical. One has a sensor recalibrated by a technician five years ago. Another runs a different batch material every Thursday, altering its thermal profile. A third has a PLC that was subtly rewired to solve a ghost fault.
Building a data infrastructure that demands uniformity is a declaration of war against reality. The winning approach does not try to force this chaotic world into a rigid schema. Instead, it builds a system that embraces and manages variability as a first principle. It asks not, “What is our standard data model?” but “How do we onboard a new, bizarre data source by tomorrow?”
The “One-Off” is Your Most Important Use Case
Pilots fail at scale not because the technology can’t handle the volume, but because it can’t handle the exceptions. The project that worked flawlessly on Line 1, with its modern PLCs, collapses when it meets the 20-year-old device on Line 2 speaking a proprietary protocol. The team spends months on the “one-off,” blowing timelines and budgets.
This is a strategic error. The “one-off” is not an anomaly; it is a test. It is the system’s first true encounter with the chaos of real operations. If your architecture cannot absorb this shock elegantly—through tools that handle diverse connectivity, or a data model flexible enough to incorporate new asset types on the fly—it will never scale. True scale is not about handling more of the same; it’s about handling relentless difference.
Design for Obsolescence, Not Permanence
The tools and AI capabilities that will be critical in three years do not exist today. Locking yourself into a proprietary, closed ecosystem—a “walled garden”—in the name of stability is a long-term trap. It makes your data inaccessible to the next generation of tools and your team dependent on a single vendor’s roadmap.
The resilient infrastructure is built from interchangeable, decoupled components connected by open standards. It uses open APIs, standard protocols like OPC UA, and common file formats. This philosophy accepts that parts of the stack will become obsolete. It ensures you can swap out a data visualization layer, an analytics engine, or a machine learning framework without gutting the entire pipeline. Your foundation is not a single cathedral, but a nimble village where buildings can be upgraded one at a time.
The Strategic Pivot: From Project to Platform
The end goal is not to complete a predictive maintenance project. The goal is to establish a data product platform. This is a fundamental shift in mindset. A project has a defined end. A platform is a living capability.
This platform is the versatile backbone that allows your team to rapidly answer questions you haven’t even thought to ask yet. It is the reason you can pilot a new AI use case next quarter in weeks, not years, because the data is already accessible, contextualized, and ready. It turns your data from a project deliverable into a permanent, productive asset.
In an environment of constant change, the most scalable design is the one that is the most adaptable. Stop building for the factory you wish you had. Start engineering for the complex, messy, wonderfully unpredictable one you actually run.
Sponsored by InfluxData
This article is based on the IIoT World Manufacturing Day session, “Building Data Infrastructure for Predictive Operations,” sponsored by InfluxData. Thank you to the speakers: Benjamin Corbett of InfluxData, Sam Elsner of Litmus, and Calvin Hamus of SkyIO.