Skip to main content
Captia Technology

What is Closed loop

Definition

What is Closed loop?

A closed loop is a system in which the output is measured and used to correct the input without waiting for a person. In a digitalised plant it means the software does not just display the data: it decides and executes under supervision through rules, workflows, energy setpoints or maintenance work orders. It is level 3, Acting, of the Physical AI Ladder, and it requires connected, reliable data.

In control engineering, a loop is closed when the effect of the action is measured and that result feeds back into the next decision. This entry explains that definition, the three actuation patterns, the guardrails that make it acceptable and what the loop demands in data and in governance.

What it is in control engineering

In open loop an action is issued and the result is taken on trust. In closed loop the output is measured, compared with the target, and the difference corrects the next action. The operating criterion is blunt: the loop is closed only if the effect of the action is measured. If the effect is not measured there is no loop, there is a recommender. That is exactly the boundary between level 2 and level 3 of the Physical AI Ladder. A PID controller is the simplest case of the loop: one measured variable and a fixed law. The decision layer described here does not replace it, it sits above it and sets its setpoints.

The three actuation patterns

  1. Setpoint on an existing loop. The system writes a reference and the local controller follows it: target temperature, flow, contracted power, line speed. It is the lowest-risk pattern: the fast dynamics remain governed by the cabinet controller.
  2. Discrete action on the sequence. Starting, stopping, switching a machine, changing recipe or diverting product. Here the interlock rules: there are plant states in which the action never executes, and that condition is a deterministic rule.
  3. Order to a person or to another system. A maintenance work order, an adjustment notice to the operator, a task created in the ERP. The actuator is human or is management software, and the effect is measured just the same.

The guardrails

  • Confirmation. Writing a setpoint is not acting. Acting is writing and verifying in the measured signal that the equipment accepted and executed it.
  • Maximum data age. Every action declares how old the data justifying it may be. Past that limit it does not execute: it would be deciding on a state that no longer exists.
  • Safe default behaviour. With missing data the answer is not to act, unless the opposite is justified in writing and tested.
  • Independent envelope. The safety layer and the interlocks remain deterministic rules and do not depend on the model. With robots involved, the applicable framework is ISO 10218-1:2025 and ISO 10218-2:2025, published in February 2025, which also absorb the content of the former ISO/TS 15066, as reported by The Robot Report on 18 February 2025.
  • Cost of a false positive. Asking for human confirmation on everything produces alarm fatigue and the operator stops looking. The cost of a false positive is that attention.

What it demands in data

Bounded latency, not just low latency: a loop with variable delay cannot be tuned, either it becomes slow in order to be stable, or it oscillates. The order of magnitude is set by periodic industrial communication, with common cycle times of 10 ms, 1 ms and 0.5 ms according to the 5G-ACIA and ZVEI paper on industrial Ethernet and 5G networks, published in November 2019. It also demands a common clock, which in measurement and control networks is solved with the precision time protocol IEEE 1588-2019, published on 16 June 2020. Without a common clock the history lets you count, not explain. And it demands a baseline: without a declared before and after reference nobody can show that the loop improves anything.

What it demands in governance

Traceability of every decision: input data with its timestamp, version of the rule or model applied, the person who supervised, the command written, the equipment confirmation and the effect measured afterwards. Declared authority: who approves the action, who can revert it and under which conditions it does not execute. Separation of read and write permissions, named access control and a rollback path that has been tested, not merely described. The industrial cybersecurity framework is the ISA/IEC 62443 series, whose part 2-1, in its August 2024 edition, sets the security programme requirements of the asset owner. Writing into the process crosses the responsibilities of production, maintenance and safety: if nobody owns the action, the project stays in pilot.

Related terms

The closed loop is level 3 of the ladder and the prerequisite for level 4 of physical AI, because it is the only one that produces pairs of action and effect. It rests on data acquisition and on the historian, it often runs from edge computing, and its organisational guardrail is human-in-the-loop. When a model decides, industrial AI agents and the monitoring of drift come into play. The architecture that supports it is described in the industrial data platform.

Related terms

Related solutions

How we apply this concept in practice:

Frequently asked questions

What is the difference between a closed loop and a dashboard?
The loop measures the effect of its own action. A dashboard shows state and leaves the decision outside the system; a recommender proposes and does not check what happened either. The closed loop writes a setpoint or issues an order, verifies that the equipment executed it and measures the result against a declared reference.
Does a closed loop replace the PLC or the PID?
No. Fast dynamics stay in the cabinet controller, which is where determinism can be guaranteed. The decision layer sits above: it sets setpoints, changes recipes, issues orders and leaves the regulatory loop where it is. Interlocks and the safety layer remain deterministic rules independent of any model.
What happens if the data justifying an action is missing?
The default behaviour is not to act. Every action also declares a maximum data age: past that limit the action does not execute, because it would be deciding on a state that has already changed. Both rules are written before the loop goes live and are tested with a missing-data and stale-data trial.
What is needed to close the loop without losing control?
Bounded and measured latency on the action path, a common clock across the sources being crossed, an immutable record of every decision with its effect, an action owner with authority to revert it, and a safety envelope that does not depend on the model. Without those five conditions, the loop is automation with no one accountable.

Keep reading

This term belongs to the scope of Captia AI. You can find every other definition in the full glossary.