From shop floor signals to industrial intelligence: what it takes to make AI work in Industry

Artificial Intelligence (AI) is becoming part of industrial operations, call it optimization, prediction or automation. Yet, in many organizations with factories, plants, and operational sites, the limiting factor is the ability to translate operational reality into data an operator can act on before the next break: data that reaches the operator in time, with context, and that the operator actually trusts. Until that gap closes, most AI initiatives die as pilots or as local wins that never survive real production constraints.

This challenge is not specific to manufacturing alone. It affects any organization with a strong industrial or operational footprint: manufacturing, retail, food and beverage, pharma, energy, logistics… Wherever there is heavy equipment, safety constraints and a daily trade-off between throughput and quality, the pattern repeats. Regardless of the sector or processes, data realities are remarkably similar.

Clear objectives, difficult execution

In most industrial organizations, priorities are well understood: improve Overall Equipment Effectiveness (OEE), reduce scrap, stabilize quality, increase throughput, and shorten the time required to diagnose and resolve issues. Defining the goals is the easy part; the hard part is acting on them shift after shift, with enough confidence to stop “firefighting”.

On the shop floor, these ambitions translate into tangible challenges:

• Stops occur too frequently.
• Restarting takes time and manual intervention.
• Defects emerge and propagate before they are detected.
• Production parameters are continuously adjusted to compensate for variability, often at the expense of performance.

Each issue may look isolated, but together they create a structural drag on efficiency and quality, and they become exactly the type of situations where organizations expect AI to help.

A real example: film breakage and what “data readiness” means in practice

A concrete illustration comes from a real project we delivered at Orange for an industrial manufacturer operating a plastic film production line. The business objective was to improve OEE, and one of the major drivers of performance loss was film breakage. Each break triggered an unplanned stop, required manual re-threading, generated material waste, and created downstream synchronization issues. Beyond availability, quality and performance were also affected, because operators continually adjusted speed and process parameters to stabilize the line and reduce the probability of further breaks.

What made this problem difficult was not the symptom, which everyone understood, but the analysis and the response. A break is almost never caused by one signal: it comes from a combination of conditions across several steps of the line. In this project, we started by working closely with production teams to understand the process end to end, then identified and mapped a large set of production signals across the line. The work involved dealing with hundreds of data points and making them consistent and meaningful across the production sequence.

Before touching a model, we had to build a reliable operational data foundation. We moved away from the typical pattern of point-to-point connections and isolated exports, and instead built a structured real-time data layer that unified OT and IT consumption. Practically, this meant taking raw OT signals and organizing them so they reflected how the production line actually runs, from the site level down to each unit, making it available close to the line for low-latency monitoring and alerting; and making the same contextualized data available centrally for analytics and AI work. The result was a shared backbone that could support operator-facing real-time views, and also enable central teams to use the same data for investigation and future optimization.

This is why we often describe “AI in industry” as a journey that starts earlier than AI itself. If you cannot produce consistent, contextualized, real-time operational data, you will struggle to build AI that is trusted, maintainable, and repeatable across lines and plants.

One problem, two viewpoints

On the shop floor, the frustration is straightforward: visibility is limited, some KPIs are still captured manually, data arrives when the damage is done… Too often, issues are discovered once they have already become downtime, scrap, or rework. Even experienced operators are forced to rely on intuition where data should provide support.

From the IT and engineering side, the picture looks different but leads to the same outcome. OT has real security constraints: segmented networks, tight access control, and good reasons for both. Core systems remain separated, with ERP, MES, SCADA, historians, and maintenance platforms typically operating with different semantics and cadence. Integration is frequently achieved through custom, point-to-point connections that work initially, but become fragile over time and expensive to evolve.

These explanations describe the same situation from two angles. Operations feel the lack of speed and clarity, while IT manages structural complexity and risk. When organizations attempt to introduce AI without addressing both dimensions, the outcome is usually solutions that work locally but cannot be replicated across lines, plants, or use cases.

AI use cases exist, but they share the same prerequisites

Identifying the right use cases is not usually the problem. Predictive maintenance, OEE, quality analytics, troubleshooting copilots, safety… every plant can list them in five minutes. What they all need is the same thing: operational data that is reliable, on time, and trusted.

In many environments, the real operational knowledge still lives in people’s heads. Meanwhile, the data that could capture this knowledge is diluted across the production process, locked inside legacy machines, and fragmented by departmental boundaries. When that happens, industrial intelligence does not accumulate, it stays local, and every new project interviews the same operators to rebuild what the previous one already learnt.

On the film line, the model was never the hard part. The hard part was knowing which of the hundreds of available signals actually explained a break, and making sure those signals arrived in the same shape, at the same cadence, every shift. Without that, any model ends up disconnected from what is happening on the line.

Data with context, shared in real time

A single line can generate hundreds of signals across machines and steps, and those signals rarely come with a shared meaning. Making them useful requires understanding the process, mapping data points to operational concepts, and aligning them with how the plant truly runs.

This is where the notion of a Unified Namespace (UNS) becomes powerful. Instead of moving data through a web of point-to-point integrations, data is published into a shared, hierarchical structure aligned with the enterprise’s operational model, such as site, area, line, and production unit. Data becomes contextualized, consistent, and reusable, both close to the line and at enterprise level. This shift directly addresses the classic limitations of point-to-point approaches, including lack of shared meaning, inconsistent formats, high maintenance, poor scalability, and delayed decision-making.

The impact is tangible:

• Near the production line, this enables low-latency analytics and alerts, where seconds can determine whether an operator limits quality impact or avoids a stop.
• Centrally, it creates a common operational view that supports analysis, optimization, and machine learning, and it becomes far easier to combine OT data with IT sources such as ERP, planning, procurement, inventory, or quality systems.

If it breaks, nobody uses it

Plants do not forgive fragile systems. If the data is missing or late once, people stop looking at the screen. Twice, and they never come back.

For this reason, reliability and security are what make the difference between a useful system and a nice demo. If the data drops, arrives late, or cannot be trusted, teams go back to manual work. Whatever we build must remain stable while the plant runs, and it has to be maintained without putting operations at risk.

They must also respect OT security constraints. Segmented networks, strict access control, and minimal exposure to external systems are realities to design with. A viable solution allows data to flow in a controlled, auditable way without compromising the integrity of the industrial environment.

These characteristics determine whether advanced scenarios become realistic: real-time operations insights, local inference close to equipment, distributed models when connectivity is limited, and decisions that cannot wait for batch processing.

Start from the line, but build for repeatability

This work starts in the plant. You cannot understand a process, or figure out which signals actually explain performance, from a meeting room in Madrid city center. In practice, progress often starts by mapping a large number of data points into a simpler operational representation, then refining it until it matches reality.

At the same time, choices must remain compatible with the broader organization. Industrial landscapes are heterogeneous:

• Some environments are cloud-first, others require on-premise or edge execution.
• Some decisions must happen locally because latency is critical, while aggregation and advanced analytics naturally belong centrally.

What matters is coherence: a shared semantic layer that keeps local and central usage aligned.

Agentic systems will raise the bar

Agentic systems will reshape industrial operations. Systems capable of planning, coordinating actions, and interacting with tools will accelerate troubleshooting, optimization, and operational support. Their effectiveness, however, relies on a basic condition: access to reliable, contextualized, secure data in near real time.

Agentic systems will not reduce the importance of industrial data foundations. They will amplify it. Plants that already have a solid data foundation will adopt these capabilities faster and with less friction. Others will discover that the limiting factor is not model sophistication, but data readiness and integration with reality.

The strategic question is therefore not whether to adopt AI: it is whether organizations are building the conditions that allow AI, predictive, generative, or agentic, to connect safely and sustainably to production. In industrial environments, that connection is what separates experimentation from transformation.