Industrial AI has passed the proof-of-concept stage. Across energy, manufacturing, water, and process industries, operators are running AI-driven predictive maintenance, energy optimisation, and quality control programmes that are delivering measurable, auditable results. The question for most industrial operators is no longer "does AI work in OT environments?" It is "where should we start, and what should we realistically expect?"
Most AI deployments in OT environments that fail do not fail because the technology is wrong. They fail because the objective was too vague ("use AI to improve maintenance"), the data quality was insufficient, or — most critically — the wrong type of AI model was deployed for the risk level of the application. Getting these three things right is the difference between a successful pilot that scales and an expensive experiment that sits on a shelf.
The Three Maintenance Eras — and Why Predictive Is Different
Reactive and preventive maintenance share a fundamental weakness: they are not informed by the actual condition of the equipment. Reactive maintenance responds to failure; preventive maintenance responds to elapsed time. Predictive maintenance responds to the actual state of the asset — vibration signature, temperature trend, current draw, acoustic signature — and intervenes before failure occurs. The enabling capability is access to high-resolution OT historian data and an ML model trained to recognise the early signatures of specific failure modes.
Predictive Maintenance: From Data to Action
The most common first AI deployment in industrial operations is predictive maintenance, and for good reason: the ROI is clear, the data is usually available (most modern historians capture the relevant signals), and the risk profile is well-understood. An incorrect AI prediction does not cause a safety incident — it causes a maintenance event to happen at a non-optimal time.
What data is required: Vibration data from accelerometers (bearing wear, imbalance, misalignment), motor current draw (overload, phase loss, winding degradation), temperature (bearing, winding, process fluid), and pressure differential (impeller wear, filter blockage) are the most diagnostic signals for rotating equipment. Three to five years of historian data, with documented fault records from the CMMS, is typically sufficient to train a useful model.
A practical example: A cooling water pump at a process plant has been failing approximately every 14 months — always requiring emergency replacement during production. An ML model trained on 4 years of vibration and temperature historian data identifies a characteristic pattern: bearing vibration increases by 15% RMS in a specific frequency band approximately 3 weeks before failure. The model triggers a maintenance alert when this pattern is detected. The next failure is scheduled rather than emergency, bearing replacement happens at a planned shutdown, and zero production is lost.
AI for Quality Control
Quality defect detection is a strong second candidate for early AI deployment, particularly in manufacturing and food processing environments where inspection at line speed is either impossible for the human eye or economically impractical at scale.
The approach combines SCADA or DCS process data (temperature, pressure, residence time) with computer vision or sensor arrays to detect product deviations in real time. An AI model trained on historical production data — including examples of good product and defective product — learns to recognise the process signatures that precede quality failures.
Training data is the constraint: Quality AI models require labelled examples of defects. For computer vision applications, this means labelled images of each defect class — typically 1,000+ examples per class for reliable performance. For process-parameter-based models, SCADA historian data correlated with quality measurements from the lab is usually sufficient. Building this labelled dataset is often the longest phase of a quality AI project.
Energy Optimisation
Energy cost is frequently the largest controllable cost in process industries — oil and gas, chemicals, water treatment, and large manufacturing facilities can spend €5–50M annually on energy. AI-driven energy optimisation targets this cost by analysing production schedules, equipment load curves, energy tariff structures, and environmental conditions to shift loads, reduce peak demand, and improve equipment efficiency.
For compressor scheduling in a gas processing plant: an AI system analyses throughput demand, equipment efficiency curves, and energy tariff periods to schedule compressor loads during off-peak tariff windows while maintaining throughput targets. Achieved savings of 8–15% of energy cost are documented across multiple deployments.
The Critical Rule: Right AI for the Right Level
The most important principle in industrial AI deployment is matching the type of AI model to the risk level of the application. This is not a conservative suggestion — it is the boundary between safe and unsafe AI in OT environments.
The rule: Large language models (LLMs) and other stochastic AI systems are not appropriate for direct control applications at Purdue Levels 0–2. They produce probabilistic outputs that can vary unpredictably. Physical processes at Level 0–2 require deterministic, validated, bounded models where the output range is known and certified.
| Purdue Level | Application | Appropriate AI Type | Example |
|---|---|---|---|
| Level 0–1 Field Devices | Real-time anomaly detection, local protection | Deterministic ML — validated, bounded, certified | Vibration anomaly → safe-state signal to controller |
| Level 2 Control | Process optimisation, setpoint adjustment | Model Predictive Control (MPC), certified ML models | MPC adjusting column temperatures within validated limits |
| Level 3 Site Operations | Historian analytics, predictive maintenance | Statistical ML, time-series models (LSTM, Random Forest) | Bearing life prediction → planned work order |
| Level 4–5 Enterprise | Work order generation, procurement, reporting | LLM agents, NLP, generative AI | AI auto-generates purchase order from low-inventory alert |
Data Quality: The Real Bottleneck
In almost every industrial AI project I've seen, data quality is the binding constraint — not the model selection, not the compute infrastructure. Most industrial companies have years of historian data, but "having data" and "having usable data" are quite different things. It's scattered across multiple historians, CMMS systems, manual log books, and ERP exports that have never been integrated into a single view. Timestamps are inconsistent — different systems record at different rates with different clock references. Fault events exist in the historian as process deviations, but correlating them with the maintenance records that explain what actually happened requires slow, manual work. And the gaps tend to appear precisely during equipment failures, because the failure itself disrupts data recording. So the moments you most need captured are often the least reliably captured.
A successful AI project budget should allocate at least 40–50% of total effort to data engineering: collection, cleaning, alignment, and labelling. Projects that skip this step produce models that look good in testing and fail in production.
Measuring ROI: What to Track Before and After
Every industrial AI project should define its success metrics before deployment and measure against a documented baseline. Without a pre-deployment baseline, ROI claims are not credible to finance or to the board.
| Metric | How to Establish Baseline | Realistic 12-Month Target |
|---|---|---|
| Unplanned downtime hours per quarter | Average of last 8 quarters from CMMS | 20–40% reduction |
| Mean time between failures (MTBF) for target equipment | Maintenance records for 3–5 years | 15–30% improvement |
| Emergency maintenance cost (parts + labour + lost production) | Finance records, last 2 years | 25–50% reduction for targeted assets |
| Energy cost per unit of output | Energy bills vs. production records | 5–15% reduction |
| Product defect rate at final inspection | Quality system records, last 12 months | 20–50% reduction (if quality AI deployed) |
Getting Started: A Practical Sequence
I'd start by identifying one asset — not a category of assets, one specific piece of equipment — that has clear maintenance cost data, a documented failure history, and good historian coverage over the past three to five years. Pull that historian data, go through it manually, and correlate it with the CMMS maintenance records for that asset. That correlation work takes longer than most people budget for, but it's also where you find out whether you actually have a labelled dataset or just a pile of time-series numbers. Once you have the labels, start with the simplest statistical approach that could possibly work — rolling window anomaly detection, or even just percentile thresholds — before entertaining neural networks. You need to prove the pattern exists in the data before you invest in model complexity. Define your success metric before deployment and write down the baseline. Then run in advisory mode — alerts only, no automated work orders — for 60 to 90 days. That window catches the false positives before they become a trust problem. If the first use case holds up, the next budget conversation is much easier.
The organisations that scale industrial AI successfully are those that start with the smallest credible use case, measure it rigorously, and use a proven result to build the internal confidence and capability to move to more ambitious applications. The technology is ready. The question is whether the data and the process discipline are in place to use it well.