Why Industrial AI That Works at One Plant Fails at Twenty

An IIoT World article by Paul Viorel, dated October 29, 2026, summarizes a panel at the Industrial AI Summit 2026 on why industrial AI projects that work at one plant stall when extended to dozens of sites. The panelists were Gaspard Duthilleul (Galyos), Sanjay Mahato (AWS), Ian Clark (InfluxData) and Tauseef Ahmed (Grayhawk Consulting Group). The thesis: the problem is mainly data architecture, not the model.

One necessary caveat: the article is labeled as sponsored by InfluxData ("Editorially Independent"), a time-series database vendor. Its conclusions should be read with that framing in mind.

What breaks first beyond a single site

According to the panel, the most common breaking points are:

• Tag name drift: "compressor one" can mean different things at different sites.
• Clock skew between pipelines, which introduces timestamp errors into training data.
• Data ownership and the definition of "quality data" diverging from site to site.
• Homegrown workarounds (manual fixes, tag spreadsheets, custom integrations) that become a burden at scale.
• Per-tag historian licensing: cost grows with every sensor added, while AI tends to work better with more sensors connected.

Five requirements for an "AI-ready" data foundation

The article lists five characteristics:

• durability with store-and-forward, to buffer data during network outages;
• nanosecond timestamp precision;
• open standards, to avoid vendor lock-in;
• support for both training and inference;
• a semantic layer that gives data the same meaning across all sites.

Context matters more than volume

A reading is useful to a model only if it comes with asset identity, product being processed, operating conditions, maintenance history and quality outcomes. Then there is resolution: dashboards sample at 1–5 seconds, while anomaly detection models often need signals in the kilohertz range. Training a model on dashboard-resolution data flattens exactly the signals it is supposed to recognize.

The Volkswagen example

The panel cited Volkswagen's Digital Production Platform, running across 43 plants. Available data is collected at the edge without prior selection, with metadata traveling alongside the data. It stays at the plant of origin, while a semantic layer keeps meaning consistent across sites. IT, OT and engineering data form a single digital thread. When one plant solves a quality problem with AI, the other 42 benefit. As reported, VW is extending the architecture to another 70 plants.

How to justify the investment

The advice is to tie the project to measurable business problems: unplanned downtime, first-pass yield, maintenance and warranty costs, labor productivity. The platform should be built as a reusable enterprise capability, and data treated as a product, with an owner and an ongoing budget. In a live poll, 33% of attendees had already started scaling their data infrastructure, while 41% considered it a low priority. An MIT survey cited in the session estimates that about 95% of the $35–40 billion invested in AI each year fails to scale; the figure is secondhand and has not been verified.

What IT and engineering leaders can take from this

The practical message is that scaling industrial AI depends on data consistency, time synchronization, context and edge architecture more than on the choice of model. Before replicating a use case at a second site, it is worth checking a few points:

• Is there a shared naming convention, or at least a semantic mapping between sites?
• Are PLC, gateway and pipeline clocks synchronized and monitored?
• Is the acquisition resolution adequate for the planned models, not just for dashboards?
• Does the data licensing model penalize adding sensors?
• Is someone responsible for data definitions and quality?

These are industrial software integration and engineering questions, and resolving them before the second installation generally costs less than fixing them at the twentieth.