What is AI (Artificial Intelligence)?
AI (Artificial Intelligence) is technology that enables a computer system to infer from its input how to generate predictions, recommendations, decisions or content, without every scenario being written out in rules in advance. Modern AI mainly learns this ability from example data: a model sees thousands of vibration measurements, camera images or process trends and then recognises patterns in new data by itself. In industrial environments AI is used for maintenance, quality control, energy management and process optimisation, and it is a key driver behind Smart Industry and OT convergence.
AI does not make systems ‘smart’ in a human sense, but it is very good at recognising patterns that are too complex or too large for people to see.
🧠 How does AI work?
The OECD (revised definition of November 2023) and the European AI Act use almost the same definition: an AI system is a machine-based system that operates with some degree of autonomy, may adapt after deployment, and infers from its input how to generate outputs that can influence a physical or virtual environment. The key word is infer (inference). A simple if-then rule written out entirely by a person, such as an alarm when a temperature exceeds 80 °C, does not fall under it according to the European Commission’s guidelines. An expert system that uses an inference engine to draw its own conclusions from captured knowledge does.
In practice an AI application goes through two phases:
- Training - an algorithm learns the relationships in historical data, for example from a historian or time series database. This needs a lot of computing power and usually happens in a data centre or the cloud.
- Inference - the trained model applies that knowledge to new data and produces a result, such as “bearing likely to fail within three weeks” or “reject product”. Inference needs far less computing power and can run locally on an edge device or IPC.
The output of an AI model is a statistical estimate with an error margin, not a certainty: even a good model is wrong in a certain percentage of cases. That is the fundamental difference from a PLC program, whose logic is explicitly defined and can be tested in full.
🔄 What is the difference between AI, machine learning and deep learning?
AI is the umbrella term. Machine Learning and Deep Learning are subfields that dominate the current wave of AI.
| Term | What it is | Typical industrial example |
|---|---|---|
| AI | Any system that infers from data how to generate outputs | Collective name for all the forms below |
| Rule-based AI (expert systems) | Expert knowledge captured in rules and an inference engine | Fault diagnosis based on a decision tree |
| Machine learning | Models that learn patterns from example data | Predicting the remaining useful life of a pump |
| Deep learning | Machine learning with multi-layer neural networks | Detecting scratches and dents in camera images |
| Generative AI and LLMs | Large language models that generate text, code or images | An assistant that searches fault manuals |
| Agentic AI | A language model that plans steps and calls tools on its own | An agent that retrieves data and proposes a work order |
Almost all AI running in factories today is narrow AI: a model that does one task well, such as detecting defects, but knows nothing about other tasks. General, human-level intelligence does not exist (yet).
🕰️ How has AI developed?
| Year | Milestone |
|---|---|
| 1955–1956 | John McCarthy and colleagues introduce the term artificial intelligence in their proposal for the 1956 Dartmouth workshop, the starting point of AI as a field |
| 1970s–1980s | Expert systems capture human knowledge in rules, including for fault diagnosis |
| 1997 | IBM’s Deep Blue defeats world chess champion Garry Kasparov |
| 2012 | The neural network AlexNet wins the ImageNet image recognition challenge by a wide margin; start of the deep learning wave |
| 2017 | Google researchers introduce the transformer architecture, the basis of today’s language models |
| November 2022 | ChatGPT makes generative AI widely accessible |
| 2024 | The AI Act, Regulation (EU) 2024/1689, enters into force on 1 August |
For industry, 2012 was the real turning point: since then image recognition has been reliable enough for visual inspection on the production line.
🎯 Where is AI used in industry?
| Application | What AI does | Data used |
|---|---|---|
| Predictive Maintenance | Predicts failing components such as bearings, pumps and motors | Vibration, temperature, current draw |
| Visual quality control | Detects defects in camera images of the production process | Images from Vision systems |
| Process optimisation | Recommends optimal setpoints, often alongside or within APC | Real-time process data and recipes |
| Energy management | Optimises consumption based on production planning and grid load | Energy meters, planning, tariffs |
| Anomaly detection | Detects abnormal behaviour in OT processes or network traffic | Process data, network traffic |
| Digital Twin | Combines physical models with AI for simulation and prediction | Design data and live process data |
The added value lies in speed and scale: a model monitors thousands of signals at once, learns from both historical and real-time data and spots weak correlations an operator would miss. It reduces human error in repetitive judgements such as inspection. More depth on these applications can be found in Industrial AI.
🔁 How does AI differ from traditional automation?
| Traditional automation | AI-based systems |
|---|---|
| Pre-programmed rules, for example in IEC 61131 | Learn relationships from data and apply them to new situations |
| Deterministic: explicit logic, the same input gives the same output | Statistical: the outcome is an estimate with an error margin |
| Behaviour fully verifiable and validatable | Behaviour harder to explain and test |
| Reacts to a measured deviation | Can predict before a deviation occurs |
| Changes only when an engineer modifies the program | Performance can degrade when the process changes (model drift) |
AI therefore does not replace a PLC or PID controller but works on top of it: it analyses, advises and optimises, while the proven control system drives the machine.
🏭 How do you deploy AI in an OT environment?
In industrial environments AI usually runs in a layered architecture:
- Data comes from SCADA, MES and the historian, increasingly via OPC UA or a Unified Namespace
- Training happens centrally; inference runs on edge computing or IPCs close to the machine, so the model keeps working if the cloud connection drops
- Feedback to the process preferably goes through the operator (advice), or through a bounded write action with fixed minimum and maximum values
- Retraining is controlled: a new model is tested first and only then released, never silently during operation
A proven approach in five steps:
- Choose one measurable problem, for example unplanned downtime of a compressor or rejects on a packaging line
- Check the data - are there enough historical failures, are the timestamps correct and are the sensors calibrated?
- Start with advice - the model gives an alert or a recommendation; the operator decides
- Measure the effect against a baseline, such as less downtime or fewer rejects
- Manage the model - monitor accuracy, plan retraining and define who is responsible
AI agents that take actions on their own need strict boundaries; see Agentic AI in OT.
🔐 Which risks and rules apply to AI in OT?
AI brings specific risks: polluted or manipulated training data (data poisoning), input crafted to deliberately mislead a model, fabricated answers from language models, and model drift. In addition, every connection between OT and an AI platform enlarges the attack surface. You secure those connections according to IEC 62443 as separate zones and conduits, preferably with data pushed out of OT, possibly through a data diode.
| Framework | What it covers | Status |
|---|---|---|
| AI Act | Legal requirements by risk; strict for AI as a safety component of a machine or in the management of critical infrastructure | In force since 1 August 2024; high-risk requirements from 2 December 2027 and 2 August 2028 |
| ISO/IEC 22989 | Concepts and terminology for AI | Published July 2022 |
| ISO/IEC 42001 | Certifiable management system for AI | Published December 2023 |
| NIST AI RMF 1.0 | Voluntary risk approach with Govern, Map, Measure and Manage | Since January 2023 |
| Principles for the Secure Integration of AI in OT | Guidance from CISA, ASD’s ACSC and partners, including the Dutch NCSC | Published 3 December 2025 |
The common thread: AI does not belong in safety decisions. The SIS and the functional safety of a plant remain deterministic and separate from AI, and people remain ultimately responsible. In addition, since 2 February 2025 the AI Act has required every organisation that uses AI to ensure sufficient AI literacy among its staff.
❓ Frequently asked questions
What is the difference between AI and machine learning?
AI is the umbrella term for systems that infer from data how to generate predictions, recommendations or decisions. Machine learning is a subfield of AI in which a model learns that ability from example data instead of hand-written rules. Almost all industrial AI in use today is machine learning.
Can AI replace a PLC?
AI cannot replace a PLC, because a PLC must control deterministically and verifiably while an AI model works with probabilities. AI therefore runs on top of the control system: it analyses data and recommends setpoints, while the PLC continues to drive the machine. Safety functions always stay outside AI.
Does AI need a lot of data?
AI needs enough representative data, and for failure prediction above all examples of real failures. A few months of clean, correctly timestamped process data is often enough for anomaly detection, whereas predicting a rare failure can require years of history. With AI, data quality matters more than data volume.
Does AI in a factory fall under the AI Act?
AI in a factory falls under the AI Act, but most industrial applications such as predictive maintenance fall into the minimal-risk class. AI only becomes high-risk when it is a safety component of a machine or of the management of critical infrastructure, such as water, gas or electricity. The AI literacy obligation does apply to every organisation that uses AI.
Where does AI run in an OT environment?
AI in an OT environment is usually trained centrally or in the cloud and runs inference locally on edge devices or industrial PCs close to the machine. This keeps the application working when the cloud connection is lost and keeps latency low. The connection to the OT network is secured as a separate zone according to IEC 62443.
What is model drift in AI?
Model drift is the phenomenon in which the accuracy of an AI model declines because the process, raw materials or sensors change compared with the training data. A predictive maintenance model can, for example, become unreliable after a pump is replaced. That is why you monitor AI performance continuously and plan controlled retraining.
📌 In summary
AI is technology that infers from data how to generate predictions, recommendations or decisions, making it a key driver behind Smart Industry and IT/OT convergence.
In industry AI works on top of proven control: it predicts failures, detects defects and recommends setpoints, while PLCs and safety systems remain deterministic. Success requires good data, a measurable problem, model management and security according to IEC 62443, within the framework of the AI Act.