Should each automated function have its own independent level?

Parasuraman, Sheridan and Wickens (2000) argue that "automation can be applied to four broad classes of functions: 1) information acquisition; 2) information analysis; 3) decision and action selection; and 4) action implementation," and that within each of these a system can sit anywhere "across a continuum of levels from low to high, i.e., from fully manual to fully automatic." A single system, they write, "can involve automation of all four types at different levels" — a conflict-detection tool that only flags and suggests a resolution for an air traffic controller is "level 4" automation of decision selection, while a system that executes its own resolution unless the controller intervenes is "level 6 or higher," and either could pair with a fully manual or fully automated information-acquisition stage. The paper's own ten-point scale (Table I) runs from the computer offering no assistance to acting fully autonomously and ignoring the human altogether, with intermediate points such as level 2 (the computer offers several options the human simply picks from) and level 4 (the computer suggests one option but the human still decides whether to execute it or pick another).

The reasoning for keeping type and level separate, rather than running everything up or down one autonomy dial, is that "the human performance consequences of particular types and levels of automation constitute primary evaluative criteria for automation design," with "automation reliability and the costs of decision/action consequences" as secondary criteria. The authors ground this in a four-stage model of human information processing — acquisition, analysis (working-memory manipulation and inference), decision, and action implementation — and argue each stage automated differently changes the human's job differently: information automation (organizing, highlighting, integrating data) tends to reduce mental workload and improve "situation awareness," while automating decision and action selection raises separate risks of "complacency" and "skill degradation" that automating information does not. Their method is explicitly a design procedure: predict or test the performance consequences across a candidate range of levels for a function, use those consequences to bound the acceptable range (their example: performance is enhanced at level 4 and degraded above level 6, so the bounds become 4–6), then narrow further using the secondary criteria.

Set against the single-axis leveling schemes already in the library, this model's distinguishing move is that it is two-dimensional before it is a scale at all. Does machine agency exist on a spectrum rather than binary? and Can we design AI autonomy separately from capability? each place a system at one point on one axis (degree of autonomy/coordination, or user role); Parasuraman, Sheridan and Wickens instead ask which of four functions is automated and at what level within that function, so a single agent can be "level 6" on decision selection while staying "level 2" on information acquisition. Their primary evaluative criterion — human performance consequences first, reliability and consequence-cost second — also predates and reframes the capability/autonomy separation that the later agent-levels paper treats as its core move: here the ordering is not capability versus autonomy but human effects versus system properties, with human effects given priority by design.

The excerpt is a 2000 paper about automated systems in domains like air traffic control, anesthesiology, and finance, built from laboratory and field evidence of that era; it does not describe large language models or AI agents, and nothing in it should be read as a direct finding about current agentic AI. Its scale and four-function breakdown are also explicitly offered as "a simple guide," not a validated measurement instrument — the authors state the model "does not therefore prescribe what should and should not be automated in a particular system," only that it gives a more objective basis for the choice than technological capability or economics alone.