AI is only as good as the data behind it.
AI is only as good as the data behind it.
Every week, a new AI tool promises to transform operations: predicting breakdowns before they occur, automatically optimising energy consumption, reducing maintenance costs and minimising downtime… Most of these tools are genuinely effective. The question is: what data are they fed?
According to an Equans study with L’Usine Nouvelle / L’Usine Digitale, 91% of respondents say they have data governance rules, yet only 45% say they use more than half of the data they collect. For companies investing in AI, this figure is worth considering, as the reliability of an AI model depends heavily on the quality of the operational data on which it is based; yet, for most companies, this data is either siloed, unstructured, or simply never collected in a usable form.
Key takeaways
- Most organisations collect far more operational data than they actually use: only 45% of them make use of more than half the data they collect.
- The reliability of AI models depends entirely on the quality of the data that feeds into them: incomplete input data produces results that appear correct but are not.
- In most programmes, the most important step is not the model itself, but linking and structuring what already exists.
- Equans Digital starts with the operational reality: the equipment, the systems and the data source.
The data problem nobody talks about
Data exists. Usable data is rarer.
Buildings, industrial sites and infrastructure generate enormous volumes of operational data around the clock: sensor readings, equipment states, energy flows, process parameters. But the problem is that most of it never reaches anything that could act on it.
OT systems (the hardware and software controlling physical processes) were designed for reliability, not connectivity. A SCADA system managing a production line, a BMS regulating a hospital's HVAC, a historian recording a decade of equipment behaviour: all of them produce data continuously. Yet much of this operational data can remain disconnected from the analytics environments that could turn it into useful insight.
The result is an organisation that has invested in AI capability and cannot understand why it is underdelivering. In many AI programmes, the model itself is only part of the challenge. The quality, structure and accessibility of the input data matter just as much.
When the data is wrong, the model looks right
An AI model trained on incomplete or poorly structured data does not crash. It produces outputs that look plausible (predictions, recommendations, anomaly alerts) built on foundations that do not hold up. In operational environments, that matters more than in most. In predictive maintenance, incomplete or unreliable data can reduce the model’s ability to detect failures early, limiting the operational value it can deliver. An energy optimisation algorithm working from partial data can entrench inefficiency rather than reduce it.
Bad data does not produce visible errors, it produces decisions made on false confidence.
What good data actually requires
From operational data to actionable insight: the four steps
Connection before intelligence
The first step has nothing to do with AI. It is connecting the systems that generate operational data (sensors, control systems, historians, BMS) to an environment where that data can be structured, governed and used.
For many organisations, this means bridging the longstanding separation between OT and IT: two environments that have traditionally operated in parallel and are not always seamlessly connected.
Structure before analysis
Raw data from operational systems is not ready for analysis as it comes. Sensor readings are transmitted in a variety of formats, come from different manufacturers, use incompatible protocols, follow inconsistent naming conventions and vary in quality. The structuring, standardisation and governance of this data are rarely highlighted during demonstrations of AI products, even though this is precisely where the return on investment lies.
History before prediction
Predictive models require historical data: a machine learning algorithm capable of anticipating equipment failures relies on thousands of operating cycles, anomalies and maintenance interventions recorded over months and years. Many facilities already hold valuable historical data in historians and control-system logs. The challenge is making this data accessible, structured and connected to the systems that can analyse it.
What is a data historian?
A specialised database that records time-stamped operational measurements at high frequency and over long periods: temperatures, pressures, machine statuses and energy flows. For AI programmes, historical operational data can be a particularly valuable source of insight, especially when it is structured and connected with other available data.
What changes when the data is right
Safran, world leader in landing and braking systems, deployed 40 IoT sensors on critical production equipment to capture continuous vibration data transmitted to a predictive-analytics platform. Integrated into an Industry 4.0 platform, this data feeds predictive alerts that detect anomalies before they cause a breakdown, reducing unplanned downtime and optimising maintenance costs.
Storengy, a major gas storage operator, needed to centralise and improve the reliability of technical data across its 14 gas storage sites. Equans Digital deployed a common data environment and, together with SAMP, a 3D digital twin to bring this information into a single, structured environment, improving data accessibility and operational visibility.
Both projects began by connecting and structuring technical data, then built analytics on top of that foundation.
Starting where the data is
Most data and AI programmes are designed from the IT side: the architecture is drawn, the platform selected, the models scoped… Until someone asks where the data will come from.
Equans Digital works from the operational reality. The starting point is always the source: the equipment, the systems, the constraints on the ground. Equans Digital teams are often already on site, maintaining the BMS, managing the infrastructure, running the facility, before any data programme begins. That proximity is what makes it possible to connect, structure and orchestrate operational data in a way that a pure technology vendor cannot: not just building the analytics layer, but making the layer underneath it reliable.
“ At Equans, we have a major advantage: the fact that we have developed, often over many years, a detailed knowledge of our sectors of activity, our clients and their facilities. Often, we are already on site, alongside our client, as part of recurring service contracts. ”
Ready to close the gap between what you collect and what you use?