The figures marked "author's estimate" in the text are not quoted from published research; the author estimated them from public data. They are: the reachable share and the ceiling on the amplification factor for each industry and occupation under three tiers of conditions; the value and conditions of 332 work activities; the ceilings for finance professionals and for investors themselves; and the consequences of default and the verifiability of results for each work activity. The value of these figures lies in their relative size and structure, not in precision after the decimal point. All the data, rules and programs used in the estimates are published below, so that readers can check them, or substitute their own judgments and recalculate.
The estimates use two public data sets. The first is the Occupational Information Network (O*NET) database, version 31.0. For 899 occupations it lists work tasks one by one, gives each task a rating for importance and for frequency, and maps tasks to work activities common across occupations. The second is Eurostat's Labour Force Survey table lfsa_eisn2: employment in the 27 EU countries in 2024, by section of economic activity and by major occupational group.
Going from occupational tasks to industry ceilings takes five steps.
| Step | Method | Data |
|---|---|---|
| 1 Weight tasks | Each task's weight is (importance − 1) × the weighted average of the frequency categories, used as an approximation of its share of the occupation's working time | O*NET: task_ratings |
| 2 Assign to activities | A task's weight is split equally among its detailed work activities, which are then grouped into 332 intermediate work activities | O*NET: tasks_to_dwas, gwas_to_iwas_to_dwas |
| 3 Double coding | Two coders independently judge each intermediate work activity's domain, its reachable proportion under three tiers, and its two trust dimensions | Rules in Section 2 |
| 4 Aggregate by occupation | An occupation's reachable share p is the sum over activities of time share × reachable proportion; occupations are approximately matched to the 9 ISCO-08 major groups by Standard Occupational Classification major group, job zone and occupation title, and averaged within each major group | O*NET: job_zones |
| 5 Weight by industry | Weighted by each industry section's employment composition across ISCO major groups; an activity's share of all working time is weighted by the employment composition of all industries | Eurostat: lfsa_eisn2 |
Three things are judged for each intermediate work activity.
Domain: which kind of state the direct result of the activity changes. One value is chosen: one's own judgments and decisions; other people's bodies, minds and relationships; things and the environment; information whose inputs are digitized or obtainable through systems; information that must be gathered on site by the senses; rights and obligations. The domains elaborate the five categories of human affairs, with the information category split in two by whether its inputs can be obtained digitally. Each activity is also assigned to one of 17 subcategories under the five categories, called classes, which are used in the explanations in the figures.
Reachable proportion: the share of the activity's working time that could be handed to an agent, assuming the agent can carry it out completely and correctly. It is judged separately for each of three tiers of conditions, taking only one of the values 0, 0.1, 0.25, 0.5, 0.75 and 0.9. A Chat works only from the material provided in the conversation and from public information. An Agent can, in addition, connect to the principal's and the organization's data and systems, run code, send messages externally with authorization, and conduct business through institutional interfaces. An Embodied Agent can, in addition, perceive and act on things on site. The parts the principal must still initiate, check and confirm are not counted in the reachable proportion; they belong to the checking effort in the denominator of the amplification factor. Where the principal must take part or decide in person, where the value comes from who does the work, or where the law requires the principal to act personally, the reachable proportion should be low. Reliability is not discounted; the judgment is made on theoretical premises.
The two trust dimensions: imagine the activity handed to an agent, and judge the most typical case. Consequences of default have four grades: minor and reversible; a loss, but remediable; a sizeable loss, partly hard to reverse; major and irreversible. Verifiability of results has four grades: can be judged only in the long run; requires professional judgment; can be checked item by item; can be judged on the spot before it takes effect. Activities whose two coders' average falls at 2.5 are placed in the boundary band and counted in no zone.
The full coding rules are in rubric-reach.md and rubric-trust.md.
The two coders are two independently run instances of a large language model (Anthropic's Claude). Each instance read only the coding rules and the list of 332 activities, not the other's judgments. The author set the coding rules and reviewed the aggregated results, but did not manually review the 332 activities one by one.
Using a model as coder has the advantage that the rules are applied consistently and the process can be fully repeated. Its limitation is that the two instances come from the same model and may share the same biases. The agreement figures below therefore show only that the rules are clear and the judgments stable; they cannot prove the judgments correct. Readers who disagree with the judgment on an activity can change it in the data file and re-aggregate.
Agreement on domain was 96%, with a κ coefficient of 0.95; agreement on class was 92%, with a κ coefficient of 0.90. For reachable proportion under the three tiers, the correlation coefficients were 0.88, 0.93 and 0.84, and the mean absolute differences 0.05, 0.06 and 0.08. The linearly weighted κ coefficients for consequences of default and verifiability of results were 0.72 and 0.83. Disagreements were handled as follows: where domains differed, each domain counted for half; for reachable proportion and trust grades, the average of the two coders was taken.
All files are UTF-8 encoded; the tabular files open directly in spreadsheet software.