What problem does it solve?
Process plants run on control loops that hold levels, temperatures and flows near a setpoint. Most loops are handled well by the plant's distributed control system and by PID control and advanced process control. A few are not. At the ENEOS Materials Yokkaichi Plant, the liquid level in a butadiene distillation column and the supply of waste heat and steam had to be held tightly to keep the product pure, and changes in outside temperature made it hard for the control system and advanced process control to adjust two key valves. Operators set those valves by hand and stepped in often when conditions shifted. Using large amounts of steam wastes energy and is costly.
Gas treating has the same pattern. Aramco's Fadhili Gas Plant removes acid gases from sour gas in an acid gas removal unit that uses amine, steam and power, and Aramco reports that its AI agents cut all three and reduced the manual interventions operators had to make there. In both cases the target is the way an existing unit is operated, hour by hour, rather than building a new unit.
How does it work?
- Pick a loop that conventional control holds badly. Typical candidates are valves that operators adjust by hand and units where quality and energy use pull in opposite directions.
- Train on a simulator. A model of the plant is built first and the AI agents learn to control it there, so no learning happens on the live process. Yokogawa describes this step at both ENEOS Materials and Aramco.
- Test against the plant's own knowledge and data. The control model is evaluated and refined with data from the unit. At ENEOS Materials, operators first entered the values the AI proposed into the control system by hand.
- Connect it to the control system. The agents are integrated with the plant's production control system so the existing safety functions stay in place. At Fadhili this was Yokogawa's CENTUM VP.
- Widen in phases. Aramco introduced the agents in three phases, optimizing one section at a time until the core of the acid gas removal unit ran under autonomous control.
- Keep the way back. At ENEOS Materials the plant has safety instrumentation, and staff say they can simply stop the AI and go back to normal control system operation.
- Audience
- Back office
- Autonomy
- Autonomous
- Adoption
- Emerging
- Channels
- API and system to system
What is it worth?
Benchmarks are computed from the public deployments below: one data point per organization per KPI, with who made each claim.
| KPI | Median | Reported range | Data points | Claimed by |
|---|---|---|---|---|
| Energy savings | Too few to pool | 5% to 40% | 2 | 1 organization, 1 vendor |
Value drivers: Lower cost to serve, Employee productivity, Risk and loss reduction.
Indicative value
A gas processing or chemical site with one to three process units that use steam, chemicals and power
USD 100,000 to USD 1.8 million
Annual steam, chemical and power cost avoided per year
How this is calculated
Formula: units * unitEnergyCost * savingShare. The low scenario uses every low input, the high scenario every high input.
| Input | Low | High | Basis |
|---|---|---|---|
| Process units put under autonomous control units, units | 1 | 3 | Editorial assumption, replace with the number of units that have a loop operators still adjust by hand. |
| Annual steam, chemical and power cost of one unit unitEnergyCost, USD per unit per year | 2,000,000 | 6,000,000 | Editorial assumption, replace with the unit's own utility and consumables cost from your plant accounts. |
| Reduction in that cost savingShare, fraction | 0.05 | 0.1 | Aramco reports around 5% less power and 10% to 15% less amine and steam at the Fadhili Gas Plant. The range runs from the power figure to the bottom of the amine and steam range, because a unit's whole utility bill falls by less than its steam use alone. It stays well under the 40% steam reduction Yokogawa reports for ENEOS Materials, which has no stated baseline. |
What it leaves out: Counts only the reduction in steam, chemical and power use. It leaves out the cost of building and validating the plant simulator, control system and instrument work, software licenses, the effort to maintain the model, and the value of steadier quality and fewer manual interventions.
Who already uses it?
2 public deployments, strongest evidence first. Grades: A regulator or audit, B the organization itself, C vendor case study, D anonymous or estimate.
Aramco
Saudi Arabia · Energy and utilities · 2025
Between October 2024 and April 2025 Aramco and Yokogawa deployed multiple coordinated reinforcement learning agents (Yokogawa's FKDPP algorithm) that directly and autonomously control the acid gas removal unit at Aramco's Fadhili Gas Plant, the process that turns sour gas into sweet gas. The agents were trained on a simulator of the plant first, tested for reliability and validity, and then connected to Yokogawa's CENTUM VP control system so the plant's existing safety functions stay in place. Aramco reports a 10% to 15% reduction in amine and steam use and around 5% less power, and says the technology is now permanently adopted at Fadhili, with other Aramco facilities assessed for further deployments. Aramco says detailed evaluations are still ongoing.
- Energy savings: about 5%, power usage at the acid gas removal unit, first results of the rollout from October 2024 to April 2025
"a 10% to 15% reduction in amine and steam usage, around a 5% reduction in power usage, improved process stability, and a decrease in required operator manual interventions"
Claimed by: organization
ENEOS Materials Corporation
Japan · Manufacturing · 2023
Yokogawa and JSR (whose elastomers business is now ENEOS Materials) used Yokogawa's FKDPP reinforcement learning AI to control two valves on the butadiene distillation column at the Yokkaichi Plant in Japan, work that the plant's distributed control system and advanced process control could not do and that operators had done by hand.
A field test ran for 35 consecutive days in January and February 2022; the AI first learned on a plant simulator, then was refined with plant data, and its recommended values were entered into the control system by operators before it controlled the plant. A revised model has run since April 2022 and the case study says it remains in routine use, with Yokogawa reporting 40% less steam use.
- Energy savings: 40%
"Reduction of steam usage by 40%, saving energy and reducing costs"
Claimed by: vendor
How do you implement it?
A model agnostic playbook: what to prepare, the order to build in, and what goes wrong.
Data you need
- Historian data on the loop's process variables, valve positions and disturbances such as ambient temperature
- Engineering data to build and validate a plant simulator of the unit
- Control loop configuration, operating limits and product quality specifications
- A record of the manual valve moves operators make today and why
Systems to integrate
- The plant's distributed control system, for example Yokogawa's CENTUM VP at Fadhili
- Plant simulator used to train and test the agents before live use
- Safety instrumented systems and alarm management, which stay independent of the AI
- Process historian for monitoring and later retraining
Complexity: High
The build needs a plant simulator that reproduces the unit, a connection to the distributed control system with the existing safety functions kept in the loop, and agreement from operations, process safety and maintenance. Aramco's rollout at Fadhili ran in three phases over section after section of the unit before the core process was under autonomous control.
- 1
Choose one loop that operators still run by hand
ENEOS Materials started with two valves on one distillation column that its control system and advanced process control had not managed to automate. A single loop with a measurable energy cost is easier to judge than a whole plant.
- 2
Train and test on a simulator before the live unit
Build a plant model and let the agents learn on it, then check the control model against plant data and site knowledge. Reliability and validity checks belong here, before any live connection.
- 3
Start with operators in the loop
At ENEOS Materials the AI proposed valve positions and operators entered them into the control system. Running that way for a period lets operators compare the AI with their own judgement and gives the first evidence for safe autonomous operation.
- 4
Keep the plant's safety functions in force
Integrate the agents with the production control system so existing alarms, interlocks and safety instrumented functions act on whatever the AI does, as Aramco did at Fadhili.
- 5
Widen by section, then by site
Move from one section to the core process in phases, run a detailed evaluation, and only then assess other units. Aramco says other facilities are being assessed after Fadhili.
Guardrails
- Operating limits for every controlled variable set in the control system, outside the AI's authority to change
- Existing safety instrumented functions and alarms independent of the AI and unchanged by it
- A tested route back to normal control system operation that an operator can trigger at once
- Model changes released through the plant's management of change process
KPIs to instrument
- Steam, chemical and power use per tonne of product or per unit of throughput, against a matched baseline
- Time the loop runs under AI control, and the number of times an operator takes it back
- Variation of the controlled level, temperature or flow around its target
- Off specification product and the fuel and labor lost making it
- Manual valve interventions per shift
Human in the loop
Operators supervise the unit and can stop the AI and return to normal control. In the ENEOS Materials trial phase they entered the AI's proposed values themselves. As a matter of practice, process safety and operations engineers should own the operating limits, approve every new model version and review the AI's valve moves after upsets; no source on this page describes that split.
Common failure modes
- The simulator does not match the plant
- Agents trained on a model that misses a disturbance, such as weather driven changes in ambient temperature, act badly on the live unit. Refine the control model with plant data before going live, as ENEOS Materials did.
- Operators do not trust moves they would not make
- ENEOS Materials staff report that operators called to ask whether it was OK to stop the AI during the 35 day test, and that the AI sometimes makes valve moves people would not think of. Show operators what the AI is doing and keep stopping it easy.
- A result that rests on one unit
- Aramco says detailed evaluations are still ongoing, and Yokogawa's release calls the Fadhili figures initial results. Treat one unit's savings as a range, and measure the next unit before counting on the same.
What are the risks and rules?
EU AI Act
Depends on design
Annex III point 2 lists AI systems used as safety components in the management and operation of critical digital infrastructure, road traffic, or in the supply of water, gas, heating or electricity as high risk. An agent that only optimizes a process unit inside operating limits and safety systems that act independently is not obviously a safety component, while one whose failure would endanger the supply of gas, heating or electricity could be. Which side a given plant falls on depends on its design and use. The deployments recorded here are in Saudi Arabia and Japan, outside the EU market.
Guidance
- Annex III, high risk AI systems referred to in Article 6(2) (European Union, Europe). Point 2 covers AI used as a safety component in the supply of water, gas, heating or electricity and in critical digital infrastructure.
- AI Risk Management Framework (NIST, North America). Voluntary framework for mapping, measuring and managing AI risk, including reliability and safety.
- ISA/IEC 62443 series of standards (ISA, Global). Security standards for industrial automation and control systems, relevant when AI agents are connected to a plant's control network.
Controls to put in place
- Plant simulator validation record for every model version before live use
- Operating limits and safety functions documented as independent of the AI
- Logging of every setpoint or valve change the AI makes, with the plant state at the time
- Network separation and access control between the AI and the plant control network
Frequently asked questions
- Who runs AI that controls a plant directly?
Aramco has autonomous control AI agents running the acid gas removal unit at its Fadhili Gas Plant in Saudi Arabia, built with Yokogawa, and ENEOS Materials has used Yokogawa's AI to control two valves on a distillation column at its Yokkaichi Plant in Japan since 2022. Both are reported by the vendor and the operator, not by an outside auditor.
- How much energy does it save?
Aramco reports a 10% to 15% reduction in amine and steam use and around 5% less power at Fadhili, and says detailed evaluations are still ongoing. Yokogawa reports a 40% reduction in steam use at ENEOS Materials. The two figures come from different processes and baselines, so measure your own unit against a matched baseline before using either.
- Is it safe to let AI set valves?
Both deployments built safeguards around the AI. Yokogawa trained the agents on a plant simulator first and connected them to the control system so its existing safety functions apply, and ENEOS Materials staff say the plant has safety instrumentation and that the AI can be stopped to return to normal control. Those are the operators' and vendor's descriptions of their own setups.
- How is this different from advanced process control?
The ENEOS Materials case describes two valves that PID control and advanced process control had not managed to automate, so operators adjusted them by hand. Yokogawa says its AI can handle conflicting targets such as high quality and energy savings, and in that case it learned its control model on a simulator, was refined with plant data, and then ran the two valves.
How to cite this page
Blits.ai AI Use Case Library, "AI for autonomous control of industrial plant processes", last verified 14 October 2026, https://www.blits.ai/ai-use-cases/autonomous-plant-process-control. Licensed under CC BY 4.0. Method: how we verify use cases.
Changelog
- 14 October 2026: First published