From control loop to learning system: The autonomous process plant is becoming the new benchmark

What is happening in Fadhili is therefore far more than just a successful AI application. A learning system is continuously intervening in a production process – within defined limits and with measurable benefits. This also makes it clear that autonomy is not synonymous with artificial intelligence. 

AI expands the process automation toolbox to include learning and adaptive methods.

However, the autonomous plant only comes into being through its interaction with conventional control technology, Advanced Process Control, reliable sensor technology, consistent data, secure automation functions and clearly defined intervention limits.

This is now possible because data availability, computing power and system integration have reached a level at which autonomous functions can be reliably implemented even on an industrial scale – albeit, so far, with a high level of project-specific effort. For widespread adoption, therefore, technologies and standards are needed that reduce this effort. And these are now available: Ethernet-APL brings digital communication right down to the field device; NAMUR Open Architecture (NOA) provides access to additional operational data without interfering with process control; management shells and digital twins describe plants in a standardised form; Module Type Package (MTP) simplifies the integration of modular processes, and open architectures such as the Open Process Automation Standard (O-PAS) break down proprietary barriers between automation systems.

The most visible evidence of this transformation is that autonomy can now be measured. The ARC Advisory Group has established a practical scale in the form of the Autonomous Operations Maturity Model. It ranges from Level 0, where humans make all decisions, to Level 5, where no human intervention is required. In between lie the crucial steps: Level 3 corrects defined deviations within programmed limits. Level 4 makes certain decisions independently and reports them to humans. NAMUR also uses a similar five-level classification. However, in a survey conducted at the user conference in November 2025, most participants still placed themselves at Level 2.

Such models serve more than just an academic purpose: they create a common language for operators, automation providers and investors. Rather than speaking in general terms about digitalisation, they allow specific capabilities to be described: which decisions are currently made by humans, which can be taken over by automation, and where it is sufficient for humans to monitor the process and intervene only when necessary.

The highest level of autonomy is not automatically the most economically sensible goal. The starting point should be a target vision agreed between the business and operations, from which a roadmap can be derived: Which functions are to become autonomous? What benefits do they promise? What prerequisites are required? For a batch-operated plant handling varying products, this objective looks different from that of a continuously operated refinery. Autonomy thus evolves from a vision of the future into a plannable transformation programme.

The Autonomous Maturity Report published by Schneider Electric in March 2026 shows how far the industry has already come along this path. In the report, 400 executives from energy and chemical companies across twelve countries rate their organisations’ maturity level at an average of 3.52 out of 5. By 2030, they aim to reach an average of 4.02. Almost a third of those surveyed cite autonomy as a ‘critical’ priority for the next five years, whilst fewer than five per cent consider it a low priority. ‘Autonomy is rapidly becoming the new operating model for industry,’ sums up Gwenaelle Avice Huet, Executive Vice President at Schneider Electric.

A look at process control systems shows that there is also a substantial market behind this development: Grand View Research estimates the global DCS market at around 23.5 billion US dollars for 2026 and expects annual growth of 5.4 per cent by 2033. The analysis explicitly cites the integration of IIoT, artificial intelligence and advanced analytics – alongside the modernisation of ageing infrastructure – as drivers of growth; the capabilities of traditional control systems are thus increasingly shifting towards autonomous functions.

The regional distribution is revealing: the Gulf region and Asia are leading the way in implementation, whilst North America is planning the most significant acceleration. Particularly far-reaching applications are emerging where operators manage large, highly integrated plants and combine digitalisation with substantial investment programmes.

ADNOC, for example, operates ‘Neuron 5’, an AI platform developed in collaboration with AIQ and AVEVA that autonomously monitors critical equipment. According to the operator, it can reduce unplanned outages by up to 50 per cent. At the same time, ADNOC is rolling out an agent-based AI solution across its upstream value chain under the umbrella of the ‘ENERGYai’ platform. The scale of the project is considerable: the three-year contract signed with AIQ for this purpose is worth 340 million US dollars. The aim is no longer merely to optimise individual units, but to integrate real-time data, analysis and decision support across the entire value chain.

Europe is taking a different approach. Whilst new production sites in the Middle East can design their facilities around modern digital architectures from the outset, the European process industry is focusing primarily on the gradual transformation of existing brownfield sites.

As impressive as the case studies from Fadhili or Hawiyah may be, they can easily obscure the actual engineering challenge. After all, learning systems can only make decisions as reliably as the data on which those decisions are based. And in many process plants, paper, walkie-talkies and filing cabinets are still the norm: OT and IT levels are disconnected. This divide between operational technology and information technology is the biggest hurdle on the path to an autonomous plant – not the AI model.

However, simply making data available is not enough. An autonomous system must also understand what the data means: a pressure reading only becomes a reliable basis for decision-making once the measurement point, the associated unit, the operating state, and the limit and setpoint values are clearly defined. For autonomous applications, data must therefore not only be machine-readable, but also contextualised, semantically unambiguous and consistent across system boundaries.

In today’s flagship projects, this foundation is still largely created on a project-by-project basis – which is time-consuming and difficult to replicate. For autonomous functions to scale widely, standardised data models are needed that can achieve this. This is precisely what the industry is working on: DEXPI, for example, no longer describes R&I flow diagrams as drawings, but as machine-readable data models that remain consistent throughout the plant’s lifecycle.

What is remarkable here is not so much the individual standard as a new willingness to cooperate: under the umbrella of the Asset Lifecycle Information Management initiative, eleven international industry associations and standardisation organisations – from DEXPI and NAMUR to CFIHOS and ISO – are coordinating their information models for the first time. If an engineering object has the same meaning regardless of the software used to create it, the information no longer needs to be re-translated at every system boundary – the database thus evolves from project-specific, painstaking work into reusable infrastructure.

This also shifts the focus of innovation. For decades, automation centred on more powerful controllers, faster bus systems and new field devices. Today, competition is increasingly shifting towards software architecture. AI models, optimisation algorithms and assistance systems are evolving much more rapidly than traditional automation hardware. It must be possible to update, expand or replace them without having to modernise the entire control system each time. Otherwise, more frequent system changes would tie up additional investment funds and slow down rather than accelerate the transformation to an autonomous plant.

Against this backdrop, software-defined automation is gaining in importance. Applications are decoupled from the hardware, functions are provided as software services and developed independently of one another. The innovation cycle is thus no longer based on the life cycle of control technology, but on the speed at which new software functions become available.

Emerson describes ‘Boundless Automation’ as a software-defined architecture in which functions are decoupled from the hardware and delivered across intelligent field devices, edge devices and the cloud. Honeywell emphasises the transition from automation to autonomy; ABB is pursuing a software- and data-driven platform strategy with Genix. The Chinese supplier Supcon is also developing in this direction: with UCS/Nyx as a software-defined control system, TPT2 as an AI platform, and a shared data space and Ethernet-APL for field connectivity.

Whilst the product strategies differ, the objectives are largely the same: open communication standards, software-defined control systems, standardised data models and AI-supported optimisation are now found in almost every roadmap. This points less to a short-term technological trend than to a fundamental shift in architecture.

The Open Process Automation Standard demonstrates that standardisation is not an end in itself. Described as a future project as recently as 2024, it now forms the basis of a new control system at ExxonMobil in Baton Rouge. The system, comprising over 100 controllers and 1,000 I/O points, went live at the end of 2024 and, according to the operator, ran throughout 2025 without any major disruptions.

In Europe, the path to autonomy involves many successive modernisation steps: decades-old process control systems, field devices from different generations and engineering data accumulated over time must be gradually integrated into a shared information environment – whilst the plants continue to produce.

Covestro in Dormagen demonstrates that autonomy is possible under these conditions. There, a production line has been operating autonomously in the existing polyester plant since June 2024. The AI not only controls individual process parameters but also manages the entire workflow, from order acceptance through production planning and manufacturing to logistics. To achieve this, sensor and process data are consolidated with information from the process control system, MES and ERP. A pilot project has thus evolved into an application deployed in industrial production.

This example is also noteworthy because it addresses the complexity of batch processes. Changing products, formulations and process sequences make end-to-end automation challenging. In Dormagen, the AI analyses the data collected in real time, recognises patterns and independently adjusts control parameters. The plant thus demonstrates that autonomous process control is not limited to large-scale continuous processes or newly built facilities.

ENEOS Materials had previously demonstrated a different technological approach in Yokkaichi, Japan. There, the reinforcement learning algorithm FKDPP controlled a butadiene distillation column autonomously for 35 days. Steam consumption fell by 40 per cent without any off-spec material being produced. Unlike traditional advanced process control systems, which are based on previously developed mathematical process models, FKDPP learns suitable control strategies directly from operational data.

SOCAR Carbamide demonstrates that autonomous control can also be applied to large-scale chemical processes. The fertiliser plant was recognised as a WEF Lighthouse by the World Economic Forum in early 2026, partly for its closed-loop, autonomous process control – achieving a 21 per cent higher throughput and 24 per cent greater natural gas efficiency.

Other European operators are also developing their existing plants along these lines. According to its own statements, BASF is transforming a “high double-digit number” of plants towards autonomy. Evonik and Siemens have announced a partnership for autonomous chemical plants – explicitly described as a “gradual, structured journey rather than a big-bang transformation”.

This is cementing regional differences: whilst the Middle East and Asia are demonstrating autonomous control primarily in large-scale continuous processes, Europe is developing methods to transform existing batch and continuous plants whilst they remain in operation. Given the brownfield capacity installed worldwide, this wealth of experience could, in the long term, prove at least as valuable as the spectacular greenfield projects.

The more decisions automation takes over, the more important the question becomes of what it is actually permitted to decide. An AI agent that independently alters setpoints or control variables must not undermine the established safety levels of a process plant. Functional safety in accordance with IEC 61511 therefore remains a separate layer even in an autonomous plant: autonomous optimisation can operate within defined operating limits, whilst safety-critical systems must, independently of this, prevent dangerous conditions from arising. Autonomy thus does not mean unlimited freedom of decision-making but rather acting within clearly defined technical limits.

At the same time, the necessary networking increases the attack surface. As OT and IT converge, data flows through edge and cloud systems, and new software services are continuously integrated, additional points of vulnerability arise. Cybersecurity thus evolves from a supporting IT issue into a prerequisite for autonomous production.

‘An experienced plant operator would normally intervene at this point’ – this was the image with which the discussion on Fadhili began. As autonomy increases, the ‘human in the loop’ is increasingly becoming a ‘human on the loop’: the system takes over routine decisions, whilst humans monitor and intervene when the system reaches its limits. They also remain responsible for targets, operating limits, exceptional cases and safe escalation. This can help alleviate the skills shortage, but it also poses a new risk: those who intervene less and less frequently may lose precisely the experience needed in exceptional circumstances.

Autonomy therefore changes skill sets rather than making people redundant. An understanding of processes remains key, alongside data literacy and the ability to assess the decisions made by autonomous systems. Trust can only be built if the operator is aware of the system’s limitations – and if there are clear rules governing who takes over when the system makes a mistake.