THE REFERENCE SCALE FOR PLANT AUTONOMY

The Six Levels of Autonomous Plant Operations

Self-driving cars had the same problem once: 'autonomous' meant everything and nothing — until the industry agreed on six levels of driving automation. Plant operations needed the same clarity. So we wrote it: six levels, from the instrumented plant to the goal-directed plant. Find yours.

Find your plant's level

The Six Levels of Autonomous Plant Operations · Arundo 2026 · v1.0 · September 2026 · Authored and maintained by Arundo Analytics

  1. Levels 0–3The operator runs the plant. The system's job is to see more, reason better, and recommend — the human decides.
  2. Levels 4–5The system runs the plant. Humans set objectives and boundaries — the system decides and acts within them.
  3. Level 2This is where Arundo Foundation runs plants today.
  4. Open standardDefinitions, entry criteria, level boundaries — free to download, cite, and reproduce.
  1. The operator runs the plant
  2. 0InstrumentedMost plants are here
  3. 1Point analytics
  4. 2Integrated assistantArundo Foundation today · engineered for Level 3
  5. 3Supervised autonomy
  6. The system runs the plant
  7. 4Bounded closed loop
  8. 5Goal-directed plant

TWO ERAS, ONE HANDOFF

Levels 0–3: The operator runs the plant.

The system's job is to see more, reason better, and recommend — the human decides.

Levels 4–5: The system runs the plant.

Humans set objectives and boundaries — the system decides and acts within them.

0LEVELInstrumented

A Level 0 plant is measured but not understood by its systems. Historians, dashboards, threshold alarms, trips, regulatory PID control. Automations are fixed-rule and context-free. The human provides all situational awareness.

In practice: The control room sees thousands of numbers; people supply all the meaning. Alarms say that something crossed a threshold — never why.

So the plant's understanding lives in your most experienced heads, and walks out the gate with every shift change. Data-rich, understanding-poor.

1LEVELPoint analytics

At Level 1, analytics exist, but each model sees only its own asset. Single-purpose models on single assets: vibration-based condition monitoring on a pump, fouling detection on a heat exchanger. Useful — but isolated, and someone still has to go and look.

In practice: Islands of intelligence in an ocean of dashboards. Every model has its own pipeline, its own dashboard, its own maintainer — and none can see beyond its own asset.

The value is real but capped: a pump model can't tell you the problem started upstream. This is the decade-old plateau most of the industry is stuck on.

2LEVELIntegrated assistant

At Level 2, one model of the whole plant can answer any question — but only when asked. A unified asset and domain model with orchestration to run analysis on demand. Plant-wide, cross-domain reasoning in human language. Still reactive — the human needs to know what to ask, but for the first time the plant can answer anything.

In practice: Every sensor, asset, model, and manual in one Domain. Engineers converse with the whole plant in plain language.

Cross-team questions that took a week get answered in minutes — with the evidence attached. This is where Arundo Foundation runs plants today.

3LEVELSupervised autonomy

At Level 3, the plant watches itself and proposes actions unprompted; people approve and execute every one. The plant monitors itself 24×7, unprompted. It detects deviations, diagnoses causes, simulates options, and emits recommended actions with confidence scores. Humans approve and execute.

In practice: The most attentive engineer on shift — one that never sleeps, never rotates out, and watches every unit at once. Deviations get caught at the drift stage, not the alarm stage.

Every recommendation arrives with its diagnosis, its alternatives, and a confidence score — trust is built into the mechanism. This is what Foundation is engineered for — and where we are taking our customers next.

Companion read: One Year Closer to the Autonomous Factory

Level 3 in practice

The difference between Level 2 and Level 3 is who starts the conversation. At Level 2, your engineers ask the plant; at Level 3, the plant speaks first:

An ethylene operator's furnaces coke at different rates as feedstock shifts. Unprompted, the plant simulates the alternatives and proposes the economically optimal decoking sequence for the next 30 days — with the margin impact of each option.

A chemicals producer runs sites in several European energy markets. As tomorrow's power prices diverge, the plant recommends shifting production between sites — with the cost delta attached.

An LNG carrier's reliquefaction efficiency drifts mid-voyage. The system flags it before anyone asks, diagnoses the likely cause, and proposes the burn-vs-reliquefy adjustment that protects the boil-off warranty.

detected unprompted diagnosed simulated recommended with a confidence score approved and executed by people

Different industries, same shape every time. That last step is what "supervised" means — and why Level 3 is an operational upgrade your operators trust, not a leap of faith.

These scenarios are drawn from real Foundation deployments at Level 2 today; Level 3 is the frontier we are building toward with these operators.

The operator runs the plant The system runs the plant

4LEVELBounded closed loop

At Level 4, the system acts on the control system itself — strictly inside an envelope you define. The system writes back to the control system within a defined envelope: named conditions, confidence thresholds, hard limits. Outside the envelope, it degrades to a safe state. Humans set objectives and handle exceptions.

In practice: The delay between insight and action disappears. Setpoints are trimmed toward the optimum around the clock; people supervise exceptions instead of executing adjustments.

The envelope keeps it safe: every permitted action named in advance, every action logged and reversible, safe state the moment conditions leave the envelope. Not the plant doing what it wants — the plant doing exactly what you allowed, perfectly, all the time.

5LEVELGoal-directed plant

At Level 5, people set business objectives and the plant runs itself to meet them. Humans supply business objectives — margin, throughput, yield, an emissions cap. The plant self-orchestrates across all units and conditions. No operating-envelope restriction.

In practice: You run the business; the plant runs itself. Prices move, markets diverge, a furnace nears its coking limit — and the plant re-plans across every unit at machine speed, holding the economic optimum continuously.

No plant operates here today. But the architecture that makes it reachable is decided at Levels 2 and 3 — which is why Foundation was designed backwards from here.

Companion read: Towards the Sentient and Autonomous Factory

Take the standard with you.

The scale is published as an open standard: definitions, entry criteria, level boundaries, and the correspondence to SAE J3016's levels of driving automation — free to download, cite, and reproduce.

Reproduce the scale freely in your own decks and documents, unmodified, with the citation. Derivative scales should identify themselves as such.

The vocabulary → glossary

Download the standard (PDF)

The Six Levels of Autonomous Plant Operations · Arundo 2026 · Version 1.0 · September 2026

Cite as: Arundo Analytics AS (2026). The Six Levels of Autonomous Plant Operations, Version 1.0. https://doi.org/10.5281/zenodo.23476035

Revision history: v1.0 — first public release, September 2026. Reviewed annually.

Pick a level to see what the next step from there looks like.

What it takes to move up

Moving up a level is an architecture problem before it is a modeling problem:

0 → 1

Takes models. Most vendors stop here.

1 → 2

Takes a Domain: one connected model of every asset, sensor, model, and document — plus the orchestration to reason across it. This is the step most plants never make, because point tools can't make it.

2 → 3

Takes always-on orchestration at scale, simulation, and confidence-scored recommendations your operators can trust.

3 → 4

Takes hard envelopes, degradation-to-safe behavior, and write-back integration with the control system.

4 → 5

Takes objective-driven orchestration across every unit and condition.

Arundo Foundation was designed backwards from Level 5. That is what "end-to-end" means — not a feature list, but an architecture with a destination.

Questions about the scale

What are the six levels of autonomous plant operations?+
  • Level 0 — Instrumented: historians, dashboards, and alarms; humans provide all situational awareness.
  • Level 1 — Point analytics: single-purpose models on single assets; useful, but isolated, and someone has to go and look.
  • Level 2 — Integrated assistant: one unified model of the whole plant; engineers ask anything in plain language and get plant-wide answers.
  • Level 3 — Supervised autonomy: the plant monitors itself 24×7, diagnoses, simulates, and recommends; humans approve and execute.
  • Level 4 — Bounded closed loop: the system acts on the control system within a strictly defined envelope; humans handle exceptions.
  • Level 5 — Goal-directed plant: humans set business objectives; the plant self-orchestrates across all units and conditions.
What level are most plants at today?+

Level 0–1. Most operations have historians, dashboards, and a handful of point analytics — condition monitoring on critical pumps, fouling detection on an exchanger. What almost nobody has is the step that changes the slope: one connected model of the whole plant that analytics, people, and AI can reason over together. That is why the industry has spent a decade adding models without getting fundamentally closer to autonomy — point tools accumulate, but they don't integrate.

What is Level 3 (supervised autonomy)?+

The level where the plant starts the conversation. At Level 2 your engineers ask and the plant answers; at Level 3 the plant watches itself around the clock, detects deviations as they develop, diagnoses the cause, simulates the options, and proposes an action with a confidence score — unprompted. People stay in command: every action is reviewed, approved, and executed by your team. It is the last level before the system touches the control layer, which makes it the natural proving ground for trust.

What does it take to get from Level 1 to Level 2?+

An architecture change, not more models. Level 2 requires a Domain — one connected representation of every asset, sensor, model, document, and their relationships — plus the orchestration to run analysis across it on demand. No number of additional point solutions adds up to this, because each one arrives with its own isolated view of the plant. This is the step where most digitalization programs stall, and the step Arundo Foundation is built to make fast: connect the data, model the Domain, and Level 2 capability follows.

Do the levels require replacing my control system?+

No. Levels 0–3 leave the control layer exactly as it is — the system reads from the plant and advises people, who act through existing procedures. Only at Level 4 does anything write back to the control system, and then strictly inside an envelope you define: named conditions, confidence thresholds, hard limits, and automatic degradation to a safe state outside it. The scale is designed so trust is earned level by level before autonomy touches control.

Can a plant skip levels?+

You can skip projects — a decade of dashboard initiatives is not a prerequisite. You cannot skip architecture. The Domain built at Level 2 is the same one that Level 3 monitoring reasons over and Level 4 acts through; without it there is nothing for the higher levels to stand on. The practical path is to enter at Level 2 directly with a first high-value use case, then climb — which is exactly how Foundation deployments are structured.

Where is your plant on the scale?

Thirty minutes with our team, and you'll know your plant's level, the constraint holding it there, and what Level +1 is worth to you.

Find your plant's level Book a level assessment