Factory Data and AI: Context Is the Real Bottleneck

On 9 October 2026, IIoT World published a summary of a panel from the Industrial AI Summit 2026 that asked a simple question: why isn't factory data reaching artificial intelligence? The speakers were Jonathan Alexander (Albemarle Corporation), Peter Sorowka (Cybus) and Kudzai Manditereza (HiveMQ). Their shared answer: the problem is no longer data collection, but contextualization.

Collection and storage are (mostly) solved

According to the panel, manufacturers have spent 10-20 years installing sensors and connecting machines, so data acquisition is largely a solved problem. Storage is also widespread, since data is routinely forwarded to historians. What is missing is the step in which raw readings acquire process knowledge, equipment relationships and operating conditions.

The speakers noted that this knowledge still lives mainly in the heads of experienced operators and engineers, not in systems. The panel identified it as the main obstacle to adopting AI at scale.

Historians store values and timestamps, but not relationships, operating context or access governance.

Two integration models

The article contrasts two approaches:

• Application-centric: each project gets its own point-to-point integration. Predictive maintenance first, then energy, then quality, each with its own connection. The result is a fragile tangle that keeps companies stuck in isolated pilots.
• Infrastructure-centric: a shared data platform connects once to operational sources and serves any application through standardized interfaces, allowing new use cases to be added incrementally.

The LLM debate

The speakers disagreed on the role of large language models. Alexander argued that plant-specific data, such as process parameters and P&IDs, is not part of the training sets of public models; in his test, a general-purpose AI got the instrumentation of a specific plant wrong.

Sorowka replied that LLMs can handle messy data, unlike traditional machine learning, where roughly 90% of the work was labeling and cleaning (his own estimate). Context can be supplied at inference time, even from a poorly scanned PDF, and waiting for perfect data is not a good reason to postpone.

Manditereza stressed meaning: knowledge graphs link work orders, equipment and downstream impacts, something an LLM looking at isolated documents cannot replicate. All three did agree on one point: operational data must be exposed to AI through a proper, governed data infrastructure.

Governance: AI as an actor with its own identity

According to the panel, AI systems on the shop floor need their own identity within the security architecture, plus explicit access rules on what they can read, which systems they can influence and which decisions they can trigger. Audit trails are also needed for what the AI has consulted, recommended and done.

Key takeaways

For plant managers and IT/OT leaders, the message is that the next step is not more sensors. It makes more sense to add context to the data already collected, such as asset, line, batch, order and operating conditions, and to connect sources once rather than project by project. From this perspective, an MES is a natural source of production context, because it knows which order and which batch are being processed on a given line. The same applies, by extension, to monitoring processing lines such as poultry lines: this is our own reading, not something stated in the article.

Limitations to keep in mind

• The article is sponsored by HiveMQ and Cybus, although the site describes it as "editorially independent", and two of the speakers are vendors: the infrastructure-centric approach is a position aligned with their market.
• Quantitative data is scarce, apart from the 10-20 years and the 90% estimate, which is Sorowka's own claim.
• The LLM question remains open: the panel did not resolve it.