TASKSCAN · rubric v1
How TASKSCAN scores a task
TASKSCAN answers one question: which parts of a job could an agent already take over, and what does that leave for a person? It scores tasks, never people, and it never says a job will disappear.
The four inputs
An AI model reads each task and gives four numbers from 0 to 5:
| Input | 0 means | 5 means |
|---|---|---|
| D — digital end-to-end | needs hands, tools or a place | the whole task happens on a computer |
| C — output checkable | pure judgement, nothing to check | a test, reconciliation or checklist proves the result |
| E — error cost | a mistake costs nothing | a mistake is severe (safety, legal, money, harm to a person) |
| L — liability or licensure burden | nobody needs a licence or carries liability | personal legal liability |
The model also marks whether the task needs a person to be physically present.
The arithmetic
exposure = 100 * (0.35*D + 0.3*C + 0.2*(5-E) + 0.15*(5-L)) / 5 if the task needs physical presence: exposure = min(exposure, 15) occupation exposure = importance-weighted mean of its task exposures
Weights (a v1 opinion, golden set proposed, pending owner hand-scoring):
- d — weight 0.35
- c — weight 0.3
- e — weight 0.2
- l — weight 0.15
Lanes, with the thresholds shown to every visitor:
| Exposure | Lane | What it means |
|---|---|---|
| 65 and above | Automate | an agent could do the task with a person checking the result |
| 35 to 64 | Delegate | an agent can draft or prepare it; a person still decides |
| 34 and below | Keep | the value is in the judgement, the presence or the relationship |
What the model does and does not do
The model supplies only the four 0–5 inputs, the physical flag and a rationale of at most 200 characters. Every published number is computed from those inputs by the formula above, in code, so the arithmetic can be checked by anyone. Task text is treated as data, never as instructions.
Scores are re-run when the rubric version changes. A task that a person has reviewed is marked reviewed and is never overwritten by a later run.
Limits, in plain words
- Rubric-based estimate, not a prediction. A high score means the task looks automatable on these four axes, nothing more.
- The weights are an opinion. The golden set is proposed, pending owner hand-scoring.
- Tasks come from the O*NET database, which describes an occupation in general, not one employer's version of it.
- Free AI tiers can change how a model behaves; scores are snapshots, not measurements.
Data and licence
Task statements and occupation data come from the O*NET database (CC BY 4.0). They were modified for this site: tasks were scored and summarised here. The U.S. Department of Labor did not approve these modifications. Wages and employment are public domain from the BLS Occupational Employment and Wage Statistics program.
The scoring batch runs on the owner's machine, in small batches, and writes its results after every batch. There is no queue to wait in.