Every occupation figure on this site comes from one of four independent studies, or is our own estimate derived from them — and each is labelled so you can tell which. Three studies measure AI exposure directly: Microsoft Research (from 200,000 Copilot conversations), the Anthropic Economic Index (from observed Claude usage), and the Felten-Raj-Seamans AIOE index (the index Statistics Canada uses for its own Canadian estimates). The Anthropic Economic Index also supplies the measured automation-versus-augmentation split, for 305 of 342 occupations; where it has no coverage we place the occupation on that axis not at all rather than substituting a judgement for a measurement. O*NET supplies the task statements. Statistics Canada and ESDC supply the Canadian occupation code and labour-market outlook.
Coverage is not universal and we do not pretend otherwise: 336 of 342 occupations carry a measured exposure composite. Where a study has no data for an occupation, that occupation's page says “not available” and does not cite the study.
Coverage counts are computed from the dataset each time this page is built.
AI applicability score, 0–1, per occupation.
200,000 anonymised Copilot conversations classified against the O*NET catalogue of work activities, weighted by how often each activity came up and how well AI completed it. Tomlinson et al., arXiv:2507.07935.
Observed AI usage, 0–1, per occupation.
The share of an occupation's tasks observed being performed with Claude, from Anthropic's openly published per-occupation dataset (CC-BY 4.0). This is usage that happened, not usage that might.
The occupation's actual task statements and work activities.
Task statements collected from people doing the job, reproduced verbatim and grouped by the work activity each belongs to. This is the only per-occupation join in the dataset that cannot be wrong: O*NET-SOC is built on SOC 2018, so no crosswalk or title matching is involved. It says what the work IS — it makes no claim about which tasks AI touches, and no published dataset supports one.
Academic exposure percentile, 0–100.
The AIOE index, which scores occupations by how much their required abilities overlap with what AI systems can do, across 774 occupations. This is the index underlying Statistics Canada's own Canadian AI-exposure estimates (Mehdi & Morissette, 2024).
Relative AI exposure tier: Low, Moderate, High or Very high.
Published with the 2025–35 Employment Projections in August 2026. BLS combined five public measures, including the Microsoft, Anthropic and Felten scores above, into a four-tier ranking of 831 occupations, so it is a cross-check on those scores rather than a separate reading. BLS is explicit that exposure does not imply job loss, automation, productivity gains or wage effects. Where one of our occupations groups several BLS occupations that sit in different tiers, we show the range.
Robot exposure index, 0–3, per occupation.
Published September 2026 (Legate-Yang & Massenkoff, CC BY 4.0). Claude, with web search, rated every physical O*NET task for the least structured setting where an existing robot can already do it: 1 a purpose-built environment, 2 a structured workplace, 3 an unstructured place such as a home or a building site. Non-physical tasks count 0, and the index is the time-weighted average. It is a model's estimate of what robots can technically do, not robots in use: Anthropic finds them cost-competitive for about 0.3% of work. A low score on a desk job mostly means its tasks are not physical. Where one of our occupations groups several that score differently, we show the range.
Automation versus augmentation, measured, per occupation.
The share of observed Claude usage in an occupation that follows an automation pattern — the task handed over — against an augmentation pattern, where the person iterates with the model. The two sum to 100%. This is the second axis of the grouping below, and it replaced a third-party classification in September 2026. It describes how people use AI for this work, not how much of the work AI can do.
Canadian occupation code and TEER level.
Each US SOC 2018 code is mapped through StatCan's published correspondence tables — SOC 2018 to NOC 2016 V1.3 to NOC 2021 V1.0. TEER is read off the resolved NOC code rather than stored separately, so the two cannot disagree.
Projected Canadian labour-market outlook.
The Canadian Occupational Projection System's assessment for each NOC 2021 unit group: balance, or a moderate or strong risk of shortage or surplus over the projection period.
Comes straight from one of the studies above, for that specific occupation:
Derived by us from the measured values, calibrated against published aggregate findings:
Calibrated against aggregates published by Anthropic and by BCG. Neither published per-occupation figures; these estimates are ours, and are not presented as theirs.
A high exposure score says AI reaches deep into the work. It does not say whether that replaces the person or amplifies them — and those are opposite outcomes. Statistics Canada, Goldman Sachs and the IMF have each converged on pairing an exposure measure with a second axis describing whether AI complements the work or substitutes for it.
That second axis is measured: the share of observed Claude usage in an occupation that follows an automation pattern rather than an augmentation one, for 237 of the 342. The remaining 12exposed occupations are not placed on it at all — their pages say so. Until September 2026 the axis was a third-party classification for every occupation; it was replaced by the measurement, then removed rather than kept as a fallback, because a fallback is still an unsourced number deciding what a reader is shown.
The line sits at half: more automation-pattern usage than augmentation. 80 of the measured occupations sit close enough to it that the side should be read loosely, and their pages say so. Combining the two axes puts every occupation in one of three groups:
The counts above are computed from the current dataset. See how they break down by domain.
The occupation taxonomy and the measured telemetry are reviewed quarterly; the Canadian concordance and COPS outlook annually, tracking the publication cycle of the underlying government tables. Each dataset records when it was last updated and how long it may go before it is considered stale, and a health check reports any that have run past their window.
A dataset whose date cannot be read is reported as stale rather than skipped — the one nobody can date is the one most likely to have been forgotten. Occupation pages carry the date of the figures they show.
An occupation average tells you where the ground is moving, not where you stand on it. Map the skills you actually own against it, and see which pathways stay open.