Skip to content
man bot

Method · version provisional-0.1

How the score is made

In plain language first, then the arithmetic, then everything it gets wrong. If you only read one section, read the limitations.

This build is not measuring anything Every score on this site is illustrative placeholder data. No figure here has been derived from a real source yet. The scoring method is Phase 0 work and is not finished. The method below is the shape the real one will take.

Jobs are bundles of tasks

Almost no job is a single thing. It is thirty or forty tasks tied together by a job title, and the tie gets re-knotted every few years. When people say a job is "at risk", what is almost always happening is that some of the tasks in the bundle are moving — to a machine, to a colleague, to the customer — and the rest stay put.

That is why this site scores tasks and not titles, and why it refuses to publish a single exposure number on its own. A job where a machine can help with almost everything and replace almost nothing is in a completely different position from one where the reverse is true, even though a single number would put them in the same place.

The four inputs

Every occupation carries four declared numbers, each from nought to a hundred. They are never added into a single "exposure" figure, and they are always shown next to the score.

Tasks a machine can help with Assisted
Higher is generally good news. It means more of the work gets faster without the work going away.
Tasks a machine can do without you Done outright
This is the number most people mean when they say exposure. It is kept separate from assistance on purpose.
Tasks needing hands, presence or a body Physical
Protective only for as long as robotics cannot reach it. Read this one next to the robotics number, never alone.
How close robotics is to that physical work Robotics reach
Low reach turns physical work into shelter. High reach turns it into the main exposure.

Worked example — Plumber

Assisted
22
Done outright
8
Physical
88
Robotics reach
12

cognitive = 0.65 × 8 = 5.2

embodied  = 0.35 × (88 × 12/100) = 3.7

reach    = (5.2 + 3.7) × 1.6 = 14.2

credit   = 0.15 × (22 − 50) = -4.2

score    = 100 − 14.2 + (-4.2) = 82

Physical work counts against a score only in proportion to how far robotics has actually reached. That single multiplication is what puts plumbing at the top and truck driving a long way below it, even though both are highly physical.

The scale

Nought to a hundred, where a high number means the shape of the work holds. The labels describe states of change, not grades — there is no "safe" band and no "doomed" one, because neither is a thing the data can support.

Scale, labels and thresholds are provisional. Open question Q-02.

0–24
Rebuilding Most of what this job was is being rebuilt. What replaces it is a different job with the same name.
25–44
Shifting A large share of the work is moving. The remaining core is narrower and more senior than it used to be.
45–64
Blending The job is becoming a partnership. Output goes up, and the judgement calls become most of the value.
65–79
Holding The shape of the work holds. The admin around it thins out considerably.
80–100
Anchored Anchored by physical presence or by responsibility that cannot be handed over. Tools change; the job does not.

The three scenarios

Chapter five lets you move the map. Those are not forecasts. Each one is a stated arithmetic assumption, written out here in full so you can disagree with the assumption rather than the number.

Today The sources as published.
The inputs exactly as the sources report them, with nothing projected forward.
Slow Automation +10%, robotics reach +10%.
Capability keeps improving but deployment lags: what a machine can do rises about a tenth, and robotics reaches into physical work about a tenth further.
Fast Automation +45%, robotics reach +90%.
Capability and deployment both move quickly: what a machine can do rises by around half, and robotics reach nearly doubles.
Uneven Fast where assistance is already high, slow everywhere else.
Adoption concentrates where it has already started. Occupations already heavily assisted (60 or more on the assistance input) move at the fast rate; everything else moves at the slow rate. This is the scenario that most resembles how the last four shifts actually went.

How these multipliers should be derived is open question Q-03.

Source register

Every source with its publisher, the version this site is pinned to, how often it changes, what it covers, and its licence. When a source changes its method between releases, this site stays on the pinned release until the change has been reviewed.

No access dates recorded yet — nothing has been ingested. Licences marked "Confirm" block launch (Q-09).

ISCO-08

International Labour Organization

ISCO-08 Confirm

Role  Master occupation list

Coverage  Global

Updates  Stable

Every occupation on this site is anchored to an ISCO-08 unit group. Where a unit group is too broad for the story, the page says so.

Working Paper 140, Generative AI and Jobs

International Labour Organization

2025 index Confirm

Role  Global AI exposure scores

Coverage  427 four-digit ISCO-08 occupations

Updates  Irregular (2023, 2025)

The only genuinely global exposure index in the register. Everything else is national and has to be mapped.

O*NET

US Department of Labor

31.0 CC BY 4.0

Role  Task-level detail, and task history through the release archive

Coverage  Around 900 US occupations and 19,000 task statements

Updates  Quarterly, with the main update in Q3

US-centric. This is the single largest known limitation on a site with a worldwide audience.

SOC to ISCO crosswalk

US Bureau of Labor Statistics

Current Confirm

Role  Mapping O*NET detail onto ISCO units

Coverage  US to international

Updates  As published

Many-to-one mappings are flagged wherever they affect a score.

ESCO

European Commission

v1.2 Confirm

Role  Skills, and adjacent moves in a later phase

Coverage  EU, multilingual, linked to ISCO-08

Updates  Periodic major versions

Not used in v1 scoring. Registered because adjacent moves depend on it.

Gen AI Capacity Study

Jobs and Skills Australia

September 2025 detailed data release Confirm

Role  Australian layer, and the augmentation versus automation model

Coverage  ANZSCO 1.3 occupations

Updates  One-off study

The clearest published treatment of the augmentation and automation split, which is why the score copies that shape.

Global Automation Atlas

Global Automation Atlas

2026 release Confirm pending Q-01

Role  Candidate robotics and physical-work dimension

Coverage  Multi-country, task to occupation

Updates  Unknown

Labelled by a language model. Adopting it, or building a custom measure from O*NET physical-work data instead, is open question Q-01 and blocks SCO-03.

Anthropic Economic Index

Anthropic

Pin one release Confirm AI company pending Q-10

Role  Optional real-world usage layer

Coverage  Claude usage only, mapped to O*NET tasks

Updates  Periodic. The method changed between the March and June 2026 releases.

Published by an AI company about its own product. Disclosed here for that reason. Whether it appears in v1 at all is open question Q-10.

What this cannot tell you

The task data is American
O*NET describes US occupations. This site has a worldwide audience and a global master list, so every task-level claim travels further than the data that produced it. ISCO-08 and the ILO index carry the global layer, but the detail underneath is US-shaped.
Some of it was labelled by a language model
Parts of the candidate robotics data, and the usage layer if it is used at all, were produced by models rather than by observation. That is disclosed here rather than buried, and it is the reason Q-01 is still open.
Occupation groups hide real differences
ISCO-08 appears to place product designers and garment designers in one unit. That single merge would have hidden the move at the centre of the essay. Where a group is too coarse, this site drops to O*NET detail and says so on the page.
A score is not a forecast
Nothing here predicts job losses, timing, or wages. It describes what a machine can currently do with the tasks in an occupation, as declared by the sources, on the day those sources were published.
One source is published by an AI company
The Anthropic Economic Index measures usage of one company’s own product. Whether it appears in v1 at all is open (Q-10). If it does, it is labelled everywhere it is used.

Data changelog

Every refresh is logged here with what changed and why. Data is replaced as versioned files, so a refresh never requires a code change.

  1. 0.1-placeholder 2026-09-11

    First build of the site with illustrative placeholder values across all ten occupations.

    To review the shape of the product before the scoring method (Phase 0) is settled. Nothing here is a real measurement.

What this site does not do

  • It does not scrape or mine job ads, social media, or anything you do here, to produce a score.
  • It does not collect personal data beyond a newsletter address, if you give one.
  • It does not give personal career advice. It explains; it does not tell you what to do.
  • It does not predict when anything will happen.