From One Plant to 20 Sites: Where the Data Architecture for AI Breaks

On October 29, 2026, IIoT World published a report on a panel at the Industrial AI Summit 2026 that tackled a practical question: why does the data architecture fail first when a company tries to extend AI across the whole organization? The speakers were Gaspard Duthilleul (Galyos), Sanjay Mahato (AWS), Ian Clark (InfluxData) and Tauseef Ahmed (Grayhawk Consulting Group).

Why the pilot works and scale doesn't

According to the article, a single-site AI project succeeds because its scope is narrow: one line, one team, well-understood data. When the effort grows to around 20 plants, the shortcuts taken at the start come to light. The critical points identified are:

• Inconsistent tag names: "compressor one" can mean different things at different sites; the article explicitly cites inconsistent tags in MES systems.
• Clock skew: timestamps from different sources drift out of alignment, introducing errors into the training data that the original team does not see.
• No shared definition of quality data.
• Manual workarounds: spreadsheets of tag schemas and custom integrations become dead weight with every site added.
• Per-tag historian licensing: cost grows with every sensor, while AI models tend to perform better with more sensors.

Five capabilities of an AI-ready data foundation

The article lists five requirements:

• store-and-forward to withstand network interruptions;
• nanosecond timestamp precision;
• open standards, to avoid vendor lock-in;
• support for both training and inference;
• a semantic layer that enforces consistent meaning across sites.

The Volkswagen example

The case cited is the Volkswagen Group's Digital Production Platform, which spans 43 plants. Data is collected at the edge without deciding in advance what to keep, with metadata (product, maintenance status, operating conditions) travelling alongside the measurements. Data stays within each plant, and a semantic layer keeps its meaning uniform; IT, OT and engineering data form a single digital thread from order to production. Also according to the article, when one plant solves a quality problem with AI, the other 42 benefit immediately, and the architecture will be extended to 70 additional plants. These are claims made in the article and have not been independently verified.

Context before volume, and the right resolution

Manufacturers, the piece notes, do not lack data: MES, historians, ERP and quality systems have been collecting it for years. What is missing is usable data, which carries asset identity, product, maintenance history and quality outcomes. Then there is the matter of resolution: dashboards show values at 1–5 second intervals, while anomaly detection models often require signals in the kilohertz range. Training them on dashboard-resolution data smooths out the very phenomena they are meant to detect.

The problem is not the quantity of data, but how much context it carries.

How to justify the investment

The advice is to tie the data platform to specific business problems: unplanned downtime and productivity, first-pass yield and defects, maintenance and warranty costs, and labor productivity. And to build it as a reusable capability across multiple plants, not as a single-scope project.

What it means for SMEs

The article is about large groups, so its lessons should be applied with caution to smaller companies. The principle, however, holds for those with two or three plants or several lines: before an AI project, it pays to standardize MES data. In practice:

• adopt a single, documented and governed naming convention for tags and assets;
• synchronize the clocks of machines, PLCs and systems;
• attach context to every reading: asset, product, order, maintenance status;
• assess the historian licensing model early, against the growth in sensor count;
• check that the acquisition rate is adequate for the phenomenon being modeled.

None of these steps is spectacular, but according to the panel they are what decides whether a good pilot becomes a repeatable capability or remains an island.