Why AI That Works in One Factory Stalls in the Second
A panel at the Industrial AI Summit 2026, reported by IIoT World on 18 November 2026, argues that what fails to travel between plants is not clean data but operational context. A useful read for anyone planning multi-site MES and monitoring rollouts.
On 18 November 2026, IIoT World published "Why AI Stalls After the First Factory", a summary of a panel at the Industrial AI Summit 2026. The speakers were Tim Burke (Arch Systems), Abby Eon (Velotic), Bryce Barnes (LG CNS) and Sahil Yadav (AOI), moderated by Jeff Winter. Note that the article is sponsored content from Arch Systems and that the views expressed are those of vendors: read them as food for thought, not as independent evidence.
The problem: clean data is not enough
The article cites two figures from the Manufacturers Alliance's Great Acceleration 2026 research: 63% of manufacturers had to clean and map their data before starting AI pilots, and 21% name inconsistent data formats across plants as the single biggest obstacle. We have not verified the original report. The panel's thesis is that clean data is a prerequisite, but what does not transfer from one site to another is the operational context.
Ready by use case, not by plant
Readiness should be assessed use case by use case: a plant may be ready for downtime attribution but lack the data needed for quality prediction. The proposed test is simple: can people from another site, who have never worked at the original one, configure the use case using only what is documented? If not, the pilot relied on tacit knowledge that does not transfer.
To survive the move to another site, a use case must meet four conditions:
- the data can be linked;
- operational context is available beyond asset data;
- the data flow is reliable and flags failures at any hour;
- the output feeds an existing workflow or system.
UNS and a data foundation are not the same thing
According to the panel, the most common mistake in AI projects is confusing a Unified Namespace with a unified data foundation. The UNS sits beneath the operational systems (MES, SCADA, APS, historian): it moves data and solves access and transport. The data foundation sits above those systems and captures what the data means, how it relates to the production cycle and what state the machine was in when it was recorded. You need both layers.
How the second plant fails
At the second site, tags have different names and slightly different meanings. The model degrades silently, without alarms, and operators lose trust before anyone identifies the cause.
"The pipeline is the asset, not the model." (Sahil Yadav)
The countermeasures suggested: standardize data contracts and deployment patterns internally without waiting for a standards body; describe relationships with an ontology and a graph model; attach provenance to every AI recommendation, meaning the data used and the model version.
The context that resists automation
The article's example: an operator running at 70% utilization looks underperforming on every dashboard. But they may have chosen that pace because spare parts arrive in six weeks and a breakdown would cost more than the lost output. That judgment lives in people's heads, which is why the operational context model should be treated as core infrastructure.
What this means for multi-plant operators
The article does not discuss MES or specific industries; what follows is our interpretation. In multi-site MES and monitoring rollouts, moving data is not enough: you need consistent semantics, such as shared tag names, machine states defined the same way and production-cycle context. A practical takeaway: before scaling AI, define a common glossary of tags and states, document it and check it against the panel's test, that is, whether another site can replicate the use case from the documentation alone.