The four-stage allocation table: which stages may be fully automated and which may not
A one-person studio brought several dozen automated surfaces under a single four-stage allocation table. Each surface is split into acquisition, analysis, decision and execution, and each stage gets a setting plus a stated reason. Settings rest on two criteria of the studio's own: whether an error at that stage is reversible, and whether it propagates to the next stage. This report records the table itself, two incidents where a setting was crossed, and the transferable steps.
1. Background and setting
This studio is run by one person, and daily operation depends on several dozen automated surfaces. Some scan topic sources every morning. Some pre-assemble data for the weekly review. Some propose archiving at the end of a working day.
Every new automated surface raises the same question: how far can this be let go. Arguing it case by case costs too much, so the four-stage model from Levels of automation was written down as a standing allocation table: four information-processing stages, one setting each, one line of reasoning each, with the setting at each stage decided by two criteria of the studio's own.
2. Intervention and method
| Stage | Setting | Instances | Reason |
|---|---|---|---|
| Information acquisition | Fully automated, highest setting | The topic radar scans and prepares material every morning; file hygiene is counted and reported automatically | Missing something is cheap, the error is reversible, and a re-run fixes it |
| Information analysis | High setting, output always marked as machine-derived | The telemetry part of the weekly review is pre-assembled by machine; the person only does the part that needs judgment | An analysis error propagates into the decision; derived output has to stay traceable and reviewable |
| Decision selection | Locked at suggest, do not execute | Expired items are listed as candidates by machine, and each deletion is decided by the person; archiving likewise only produces suggestions | A decision carries responsibility, and the cost of deleting wrongly is not symmetric with the cost of keeping wrongly |
| Action execution | Lowest setting, with a procedural lock | A restore point is created before any batch change; automatic restore is permanently forbidden; merging into the main line needs human authorisation | For an irreversible error there is no such thing as a re-run |
The point of the table is not the four rows. It is the rightmost column. The four stages sit at different places on the two criteria, reversibility and error propagation, so their settings differ. A single on-off switch has no place on this table.
3. Observations and results
Writing the table down did not make it hold. Both records below are cases where a setting on the table was actually crossed. The corrected wording is now the reasoning line for the matching row.
The monthly review once ran on an unattended schedule. The problem was not accuracy of execution. A review that needs a person to supply judgment, hung on an unattended schedule, degrades into a formality nobody reviews. The correction: the review became manually triggered, and the schedule now only carries the reminder. A reminder can be automated. A review cannot.
The whole-tree restore point taken before batch operations came close to being wired into automatic rollback on failure. The risk is that a rollback returns the entire working environment to the snapshot, which can overwrite progress made elsewhere after the snapshot was taken. The ruling was written in one line: automatic restore is permanently forbidden, and any restore must be run by hand or explicitly authorised.
4. Interpretation
This case matches the blind-spot line in Levels of automation: automation is not one switch, it is a setting per stage. Setting automation stage by stage is the skeleton of the Parasuraman model. The two criteria, reversibility and error propagation, are this studio's own allocation rules and do not come from the original paper. The four stages sit differently on both, so there is no single correct degree of automation.
Record one also bears out the attention mechanism in Automation complacency and bias: the more reliable an automated surface, the faster the person's sampling rate falls. An unattended schedule with no separate spot-check sets that sampling rate to zero, and the piece missing in record one was exactly that human judgment step. Record two is the direct reason for keeping the execution layer at its lowest setting: the cost of an error in an irreversible operation is asymmetric, so a procedural lock beats repair after the fact.
5. Limits and boundaries
This report is one studio's record of testing itself (n=1). It makes no claim to statistical validity.
The setting is a one-person, multi-line studio. Whether the allocation table transfers to a multi-person organisation has not been tested.
Both incidents are internal observations with no control group, and the assessment of the corrections relies on self-reported compliance. Cite this report as single-case evidence.
6. Transferable steps
- Split one automated surface into four stages: acquisition, analysis, decision, execution.
- Answer two questions for each stage. If this stage goes wrong, is it reversible? Will the error propagate to the next stage?
- Stages that are reversible and do not propagate can go to a high setting. Stages that carry responsibility or are irreversible stay locked at suggest, do not execute, or below.
- Write the reason on every row. A setting with no reason drifts sooner or later.
- Update the table when something goes wrong, and do not blame the individual. Missing things has a structural cause; see Automation complacency and bias.
- [1] Parasuraman, R., Sheridan, T. B., & Wickens, C. D. (2000). A Model for Types and Levels of Human Interaction with Automation. IEEE Transactions on Systems, Man, and Cybernetics, Part A, 30(3), 286–297. (Source of the four-stage-by-level model. The table's two criteria are the studio's own and do not come from this paper.)
- [2] Parasuraman & Manzey (2010). (Basis for the reading of record one; full reference on the Automation complacency and bias node page.)