The Physical AI Ladder places a plant by what it can do with its data, not by the technology it has bought. This entry describes the five levels, the observable criteria that place a plant on one of them, how it is used in a diagnosis and what it is not.
What the ladder is
It is a Captia framework for ordering a conversation that usually starts at the end. Physical AI does not start with the robot: it starts with the data from the machine you already have. The ladder turns that idea into five rungs with fixed names, each defined by a specific capability over the data rather than by a product. The rule that governs it is cumulative: a level counts as reached when its own criteria and those of the lower levels are met. No rung is skipped, because each one produces something the next one consumes.
The five levels
- Level 0, Isolated. Data lives in each machine, on a screen or on paper, with no comparable history. To know what happened yesterday you have to walk to the equipment.
- Level 1, Connected. Equipment speaks a common language and data reaches a single place with time and context. This is the work of protocols, edge, normalisation and source timestamping.
- Level 2, Aware. The plant knows what is happening: OEE, energy, stoppages and quality in real time per line, over a history that allows periods to be compared.
- Level 3, Acting. The system decides and executes under supervision through rules, workflows, energy setpoints or maintenance work orders, and records the measured effect of every action.
- Level 4, Physical. Perceiving, reasoning and acting autonomously on the real world. It is the state of the art of the sector and is described in the third person: Captia works on levels 1 to 3.
Observable criteria
Each level is checked with evidence that can be requested during a plant visit, not with a statement of intent.
- Level 1. Plotting data from two different machines on the same chart and the same time axis, and a signal dictionary with asset, unit and frequency maintained by a named person.
- Level 2. A single written definition of OEE and of planned time, and a breakdown of stoppage reasons in which the unclassified share is known and low.
- Level 3. At least one recorded action with its effect measured against a reference, an identified supervisor and a rollback path that has already been used.
- Level 4. Quantitative evaluation of the policy on cases that were not in the training set, and a safety envelope independent of the learned model.
How it is used
It serves to place, to order and to stop. Place: one question per level is enough for a first cut. Can each machine’s data be seen outside that machine? Does it reach a single place with time and context? Are OEE and energy visible per line in real time? Are there actions that execute under supervision and leave a record of the effect? Order: it fixes the investment sequence, because each rung produces value on its own and prepares the next one. Stop: it prevents the direct jump from level 2 to level 4, which is the most expensive failure. The context justifies the caution. According to the third barometer of industrial digitalisation and automation, presented at Advanced Factories 2026 and reported by Metalindustria on 22 May 2026, only 3.3% of companies say they have fully digitalised smart factories.
What it is not
It is not a generic maturity model: it does not score strategy, organisation or culture, but a single dimension, what can be done with equipment data. Nor is it a certification: there is no seal, no accredited auditor and no number to display. It does not measure purchased technology, because a plant can have a platform and still sit on level 0 if data does not arrive with time and context. And it does not describe a product catalogue: level 4 is not installed on top of level 2.
Related terms
The term the ladder orders is physical AI. Level 1 rests on OT/IT convergence, edge computing and data acquisition. Level 2 is measured with indicators such as OEE. Level 3 is the closed loop, with human-in-the-loop as the guardrail. At level 4 inference moves next to the machine and Edge AI appears. The architecture that supports levels 1 to 3 is described in the industrial data platform, and the guide industrial physical AI develops each rung with examples.