By Dr Maurizio Pilu, PhD (AI), eMBA, Fellow IET
August 2025
Over the past decade, I have had the privilege of working directly with dozens of industrial clients on AI-based innovations and working “under the bonnet” with some of the best industrial AI tech on the planet.
The GenAI disruption of the past few years has blurred the difference between the world of industrial AI and the one we use every day on our computers or in the office.
Industrial AI differs fundamentally from commercial AI, which powers chatbots, personalised recommendations, and other applications. Industrial systems are cyber-physical, safety-critical, and in some cases highly regulated.
In an industrial plant, a mistake can lead to downtime, environmental damage, or even threaten human life. Industrial AI must seamlessly integrate with legacy systems, need to be assured and, in some cases, certified, endure harsh environments, and deliver reliability and safety.
That is why industrial AI demands a different mindset: one that prioritises resilience, explainability, cybersecurity, and trust at every layer of the stack.
I believe not enough attention is being given to the fundamental difference between the two types of AI, but I believe the coming years will redefine how industrial software is designed, commissioned, operated, and sustained.
But what exactly is industrial AI? How different is it from what has been going on to date? How will the latest disruptive development seep into the slower, high-stakes world of industrial technologies?
Below are ten forward-looking industrial AI scenarios that I have directly experienced in conversations, strategic advisory or actual projects with clients.
1. Industrial Software Becomes No-Code or Self-Writing
For decades, industrial software, from PLC logic to MES workflows, has been written by human engineers. (Disclaimer: My first job as a new graduate aeons ago was to develop PLC software for a manufacturing operation).
But we are now entering an era where AI can generate, test, and validate code automatically.
Companies are experimenting with AI-powered code synthesis using historical data, logs and configuration templates. There is an emergent class of digital twins that can be developed using WYSIWYG interfaces, reinforcement learning and, lately, using chain-of-reasoning from LLM tools.
This shift does not eliminate engineers, but it redefines their role: from writing code line-by-line to validating, refining, and certifying AI-generated logic. The productivity gains could be dramatic, reducing integration costs and accelerating time-to-market.
It is still early days, but system integrators are already exploring AI-assisted programming, and a new class of well-funded startups are targeting this market. And we are already seeing new markets developing: standardisation, independent assurance and certification services.
2. Industrial “GPTs” as Digital Operators
Just as GPT-like models have transformed how people interact with knowledge, domain-specific foundation models are emerging in industry. These “Industrial GPTs” combine sensor data, equipment manuals, and process knowledge to act as digital operators or co-pilots.
For example, generative AI has already been introduced to give plant operators natural-language insights into asset data, and systems already exist that integrate multimodal data for smarter decision-making. A GPT-based assistant has been recently deployed to help engineers draft approval documents and swiftly search for regulations, precedents, and other regulatory information buried in large datasets, thereby streamlining and accelerating nuclear-technology regulatory processes.
The trajectory is clear: operators will increasingly converse with their plants, with AI not just reporting on status but suggesting and eventually executing actions.
There is also another important angle. As experienced staff retire long before the asset’s natural life, AI-based knowledge capture has become a strategic priority for HR and operations to retain expertise and ensure business continuity.
3. Sensorless and Ubiquitous Sensing
Instrumentation costs have previously limited industrial process monitoring, but AI-driven virtual sensors and affordable IoT devices now enable comprehensive sensing.
Some companies are releasing AI-based “inferential sensors” to estimate unmeasured variables from historical data, and with ever cheaper battery-free IoT sensors, factories, plants, infrastructure and mobile assets can upgrade legacy systems efficiently, achieving real-time situational awareness without major investments.
Advanced AI, digital twins, and increasingly low or no-code are ready to receive all this data and create business and sustainability value.
4. Real-Time ESG-Aware Optimisation
Sustainability is evolving from a focus on annual reporting to serving as an operational control parameter.
Artificial intelligence can now be employed to optimise production not only as traditionally for yield or cost, but also for real-time metrics such as carbon, water usage, and social risks.
Platforms are being released to enhance energy efficiency for industrial clients. In the steel sector, AI is starting to be used to lower the carbon intensity of blast furnace operations. Several startups are now racing to develop ESG-oriented digital twins of operations, allowing managers to, e.g. scenario plan changes to the supply chain, or materials, thereby going into the more challenging Scope 2 and 3 domains of emission reporting.
Looking ahead, we can easily see a future where auditors may be granted read-only access to digital twins, enabling ESG compliance to be integrated directly into operational decisions rather than being addressed retroactively.
5. The Rise of the Industrial App Store
Imagine deploying a predictive maintenance module as easily as installing an app on your phone. Why has it not happened to date? Many reasons, but if I have to list the top 3: fragmentation, safety, data quality and integration issues.
That vision is emerging with large vertically integrated operators which offer marketplaces for certified AI and analytics apps on their stack.
But outside of it, AI is poised to lower barriers to the creation of flexible applications, enabling plants to buy capabilities instead of building from scratch or replacing the full stack.
AI could be used to act as a translation layer between data, protocols, automating compliance checks, or even using digital twins to simulate the deployment of a new app.
While this is still to be played out, there may be opportunities for new ecosystems of industrial software developers, mirroring what happened with smartphones.
6. Fully Virtual Commissioning
Virtual commissioning, the automation logic and plant behaviour in simulation before physical deployment, is rapidly becoming a possibility.
Digital twins allow companies to de-risk projects, cut commissioning time, and reduce expensive on-site surprises.
Companies are now investing in full digital twins of factories and simulating assembly lines before they exist, or using virtual commissioning to verify robotics workflows.
Looking forward, we will see AI playing an ever bigger role in simulation. With an LLM-based chain of reasoning coupled with more traditional twins, AI can help planners explore edge cases and failure scenarios, reducing the cost and risk of commissioning.
7. Digital Calibration, Assurance and Certification
Calibration, assurance and certification are a costly bottleneck in regulated industries, not only in direct costs (as high as 1-2% of turnover) but also in terms of opportunity and downtime costs.
AI could allow remote verification and automated assurance. For instance, companies are now experimenting with AI to predict when equipment needs calibration, reducing unnecessary downtime. Regulators are beginning to consider how digitally logged calibration and AI-verified compliance could replace some on-site inspections.
The long-term, but very realistic, outcome may be continuous assurance and compliance, where industrial assets are assessed in real time, for instance via read-only APIs to approved digital twins of operations.
8. Predictive Autonomy at the Edge
The rise of the latest LLM models created an unprecedented demand for cloud computing, in particular, AI computing for training and inference. But not all intelligence belongs in the cloud.
Increasingly, AI models run directly on edge devices such as drives, valves, and PLCs, and efforts are underway to create “small” LLMs that can run on embedded devices but exhibit, in limited domains, the same power. The imperative of data privacy and cyber-resilience as just two of the drivers on this counter trend, enabling autonomy in sensitive, disconnected or remote environments.
For example, numerous players are now developing real-time AI at the edge for industrial IoT, and energy companies are deploying edge AI on remote energy infrastructure to monitor safety-critical equipment even when connectivity is poor.
The overall trend is for Industrial AI to become narrower, more on-premises and local, but with powerful central oversight to get the best of both worlds.
9. Causal AI
Traditional AI and machine learning often struggle with causality, learning models from past data, but not really understanding why and how. And we have all experienced how even the most advanced LLMs are prone to hallucinating, something that the hyperscalers are trying to address.
In industrial environments, the physics of things does matter, hugely. Knowing what causes what is essential.
Causal AI promises to bridge the latest development of AI with an understanding of physics. This will allow to e.g. simulate interventions before execution, providing safer and more reliable recommendations, or making sure predictive models reflect the physics of the underlying assets.
Causal machine learning methods are being developed for industrial processes, and startups are developing causal AI platforms for decision-making, such as in energy and transport.
Causal AI is also bound to simplify certification of both the AI (e.g. digital twins) as well as plants and systems, by helping predict how process changes affect compliance, making audits faster, clearer, and more evidence-based.
10. Autonomous Drones and Robotics
In industries like oil, gas, mining, and construction, hazardous environment inspections are one of the core activities carried out to make sure assets are safe and efficient. Yet over 80% of them are still carried out by humans, and in some sectors or for some assets, human inspection is a regulatory requirement.
While drones and robots are increasingly used, they still require hard-to-find pilots, and it’s not uncommon to have gigabytes of visual data sent by hard disk to headquarters.
Advances in small, efficient AI models and unprecedented breakthroughs in visual analytics capabilities allow these systems to be increasingly autonomous, process data locally and respond in real time, even in remote or harsh conditions.
AI-powered drones and robots can now detect anomalies, map inspection routes, and feed real-time data into digital twins for predictive maintenance and simulation.
Looking ahead, autonomous inspection platforms will become increasingly capable of self-planning, collaborating, and making operational decisions, further extending workforce reach and transforming industrial operations.
The Cybersecurity Imperative
These forward-looking scenarios will open up tremendous possibilities and markets.
Yet, with AI embedded more and more in industrial software and even in control systems, the attack surface expands. Attacks such as adversarial inputs, model poisoning, and hijacking of autonomous drones are no longer theoretical risks.
Cybersecurity is and will remain a first-class requirement of industrial AI. Organisations will need cyber-physical security frameworks that protect not just IT and OT, but also the AI models themselves, with state-of-the-art ModelOps and governance frameworks.
Final Thought
AI has been used in one form or another in industrial sectors for decades. But the latest developments are game changers, and the industrial AI revolution is rebooting. AI can become a force for resilience, safety, and sustainable growth.
To be sure, AI innovation can be hard and expensive, and many industrial organisations do not have the depth of AI skills in-house to make the most of it, nor the urgency.
At Enexem, leveraging our experience of supporting some of the largest industrial players out there to innovate with AI and emerging tech, our perspective is simple: the winners will not be those who adopt AI fastest, but those who adopt it strategically, securely, and sustainably.
That means aligning AI with performance, sustainability, safety and resilience goals, embedding cybersecurity, and ensuring that human expertise and ethics remain at the core.
Are you working or planning to work with industrial-grade AI?