Canada is short of the workers
AI reaches furthest into.
Two official pictures of the same labour market disagree. ESDC projects a shortage of pharmacists, psychologists, dietitians, teachers and engineers over the next decade. Measured AI-usage data puts those same occupations in the highest exposure band we track. Both are true, and almost nobody is reading them together.
- Short-staffed and heavily exposed
- 15 occupations
- Of those, a strong shortage risk
- 4 occupations
- Exposed but amplified, not replaced
- 54% of exposed occupations
The overlap
Of 342 occupations, 100 carry an ESDC projection that Canada will struggle to fill them. Separately, measured telemetry from Microsoft Research and the Anthropic Economic Index places a set of occupations in the top exposure band — the work AI already reaches deepest into. 15 occupations appear on both lists.
That combination is the awkward one. A shortage says employers will want more of these people. High exposure says more of the task is already being done with AI. Neither cancels the other out, and the policy response to one is not the response to the other.
Exposure: Microsoft Research (arXiv:2507.07935) and Anthropic Economic Index, measured. Outlook: ESDC COPS 2024–2033, verbatim. Percentile is a composite of the two telemetry measures.
Exposure is not the same as replacement
It is worth being precise about what a high exposure score means, because the number is routinely read as a countdown. It measures how much of an occupation's work AI already touches. It says nothing on its own about whether the person doing that work is amplified or removed.
Pairing exposure with the measured automation-versus-augmentation split separates the two. Across 342 occupations, 249 (73%) are measurably exposed — but in 54% of those, most of the AI usage actually observed is augmentation rather than automation. Only 102 occupations (30%) are both measurably exposed and automation-majority.
One thing to hold alongside that number: it counts occupations, not people. Every occupation here is weighted equally, whether it employs a few thousand people or a million. An analysis weighted by employment can land somewhere quite different — the Future Skills Centre, working from 13 million job postings and weighting by workers, found a far more even split between work AI complements and work it can automate. We hold no per-occupation employment counts, so we cannot weight ours, and the honest reading is that these two answers are measuring related but different things.
Exposure measured; the automation/augmentation split measured for most occupations, not ours.
The displacement that exists is not where the headlines put it
Grouped by domain, substitution is not spread evenly and it is not concentrated in the fields that dominate coverage. Technology accounts for the sharpest concentration: 21 of its 43 occupations. Technology, by contrast, has none — despite a mean exposure percentile in the same range.
| Domain | Roles | Automation-majority | Mean exposure |
|---|---|---|---|
| Marketing & Sales | 10 | 440% | 85th |
| Business, Finance & Legal | 29 | 1138% | 69th |
| Technology | 43 | 2149% | 67th |
| Science & Education | 43 | 921% | 67th |
| Creative & Media | 27 | 519% | 65th |
| Management & Ops | 50 | 1836% | 54th |
| Trades & Healthcare | 140 | 3424% | 29th |
Exposure does not track how much training a job requires
Canada classifies occupations by TEER — the training, education, experience and responsibility a role typically demands. If AI exposure were a story about low-skilled work, displacement would fall as TEER rises. It does not. It sits between 20% and 29% across every level but one, and spikes at TEER 4 — 63% — before falling again.
| TEER | Typically requires | Roles | Automation-majority | Mean exposure |
|---|---|---|---|---|
| 0 | Management | 17 | 24%4 roles | 54th |
| 1 | University degree | 76 | 24%18 roles | 65th |
| 2 | College / apprenticeship, 2+ yrs | 75 | 29%22 roles | 41st |
| 3 | College / apprenticeship, under 2 yrs | 28 | 21%6 roles | 37th |
| 4 | Secondary school | 16 | 63%10 roles | 68th |
| 5 | Short-term demonstration | 5 | 20%1 roles | 36th |
TEER 4 covers 16 occupations here — a small group, and the figure should be read with that in mind. The table covers the 217 occupations whose NOC code is certain; TEER read off a code we picked from several candidates is not reported.
Method, and what this does not show
- Exposure is measured, not forecast.It comes from Microsoft Research's analysis of 200,000 Copilot conversations against O*NET work activities, and from the Anthropic Economic Index's open per-occupation usage data. 336 of 342 occupations carry a measured composite; the rest say so rather than showing a number we invented.
- The Canadian outlook is ESDC's, verbatim. COPS 2024–2033, per NOC 2021 unit group, for 327 of 342 occupations. National only — the source publishes no provincial split, so neither do we.
- The occupations are US SOC codes mapped to NOC 2021through Statistics Canada's published correspondence tables. 177 resolved to a single unit group; 158 had several candidates and took the closest by title.
- The mapping is many-to-one, so these are not distinct Canadian occupations. The 335 mapped occupations resolve to 258 distinct NOC unit groups, and 128 of them share a code with at least one sibling. Rows sharing a unit group inherit one COPS assessment and one TEER level between them, so a Canadian count here counts US occupations, not Canadian ones.
- TEER is reported only where the NOC code is certain.TEER is the second digit of that code, so it inherits the code's uncertainty. 217 occupations either resolved to one candidate or matched our title on every word; the other 118 were the closest of several and are excluded from the table below rather than counted under a level we chose.
- Automation versus augmentation is measured for 237 of the 342 occupations, from observed Claude usage (Anthropic Economic Index). The remaining 12 exposed occupations carry no measured split and are left off that axis rather than assigned a side. Note the split describes how people use AI for the work, not how much of the work AI can do.
- Every figure counts occupations, not workers. All 342 are weighted equally regardless of how many people they employ, so no percentage here describes a share of the workforce. Weighting by employment would need per-occupation headcounts we do not hold — the Statistics Canada Labour Force Survey is the source that would supply them — and it could move the balance substantially, since the automation-majority occupations are not necessarily the small ones.
- No salary data. We hold none we can stand behind across all 342 occupations, so we publish none.
Full method and sources · The Canadian picture · All 342 occupations
Citing this
Free to quote and link with attribution. Underlying figures are computed from the occupation dataset at build time, so the numbers on this page and on each occupation page cannot disagree.
Fractional Manager™ (2026). Canada is short of the workers AI reaches furthest into. https://fractionalmanager.org/research/canada-labour-shortage-ai-exposure