Machine-learning-based advanced process control that continuously retrains against live plant data, rather than the fixed linear models of first-generation APC.
Agentic AI for closed-loop process control
Control Systems — DCS · PLC · ESD · HMI
Emerging Idea
Mid-term · 5–10 yrs
Evidence E1 · Concept
AI / ML
AI agents that don't just recommend a setpoint change but are given standing authority to execute a bounded set of control actions autonomously.
Short entry. This page summarises our current placement. Full, sourced analyses are published first for the technologies closest to practical use; see the list of full analyses.
How it works
An advisory system suggests; an agent acts. In an agentic architecture, software is given a goal, such as minimising energy use while holding product quality, together with a set of tools it may use: reading process data, running a simulation, and writing a limited set of setpoints. It plans a sequence of steps, executes them, observes the result and adjusts.
The engineering effort goes into the boundaries. Every proposed action is checked against hard limits before it reaches the control system, the agent normally writes to supervisory setpoints rather than directly to valves, and the basic control and safety layers beneath it remain unchanged and independent. Operators can see what the agent intends to do and can take control back at any moment.
Why it matters
Represents the frontier beyond advisory APC — genuinely autonomous control decisions inside defined guardrails — with the largest potential efficiency upside and the largest governance question mark.
Where it fits
The plausible early applications are supervisory: coordinating several existing controllers across a unit, managing grade transitions, or optimising utilities, where a mistake is slow to develop and easy to reverse. Published demonstrations of autonomous control over extended periods exist, but they are few. Broad adoption depends on accumulated operating evidence and on agreed engineering practice as much as on the technology.
Why we place it at the Emerging Idea stage
Agentic AI is heavily promoted in enterprise IT, but in closed-loop process control there are only demonstrations and tightly bounded pilots. Industrial deployment lags the general agentic-AI narrative by years, which is why this entry sits at the Emerging Idea stage rather than at Over-Promise. Estimated time to the Proven in Use stage: 5–10 yrs. What the stages mean.
Key risk
Functional-safety certification, liability and operator trust for an autonomous agent with write-access to a control loop are unresolved — expect this to move slower in ICS than in IT.
Questions to ask before committing
- Exactly which variables can the agent write to, within what ranges and rates of change, and what enforces those limits?
- Can operators see the reasoning behind an action, and is every action logged?
- How was the agent tested against abnormal situations, and what does it do when its inputs are bad?
- Who is accountable when an autonomous action contributes to an off-specification batch or a trip?
Standards and terms to know
- Guardrail
- A hard constraint enforced outside the AI that an action must satisfy before it is applied.
- Human on the loop
- A person who supervises an autonomous system and can intervene, without approving every action.
- Supervisory layer
- The level above basic regulatory control where setpoints, rather than valve outputs, are adjusted.
Related technologies
See this entry on the full maturity map · All Control Systems entries · How placements are made. Stage and time-horizon placements are editorial judgements based on public information as of October 2026, not engineering, safety or investment advice; see the Terms of Use.