AI Exposure of the Canadian Job Market

Mapping how susceptible every occupation in the Canadian economy is to AI and automation — using open government data from Statistics Canada and ESDC, scored by a large language model.

Based on karpathy/jobs by Andrej Karpathy. The original project built this visualization for the US job market using Bureau of Labor Statistics data (342 occupations). This repository adapts the entire pipeline for Canada: replacing BLS with Statistics Canada and ESDC sources, swapping SOC codes for NOC 2021, and expanding coverage from 342 to 516 occupations. The scoring rubric, visualization engine, and site layout are inherited directly from the original; everything data-facing is Canadian.

---

What it shows

An interactive treemap of 516 Canadian occupations (NOC 2021). Each rectangle's area is proportional to the number of Canadians employed in that occupation. The colour runs green → red based on AI Exposure — a 0–10 score of how much cognitive AI is expected to reshape the work.

A second view, Exposure vs Outlook, stacks occupations into columns by AI exposure score with colour showing labour market outlook (surplus to shortage). An exposure score range slider lets you narrow columns to any subset of scores (e.g. 4–8 only).

The site supports light and dark mode (persisted via localStorage), eleven category filter chips (including a dedicated Robotics Risk filter for manufacturing, trades, and natural resource sectors), and a Methodology & Stats page with verified job-weighted statistics. Hovering any occupation shows its robotics/physical automation risk alongside the AI exposure score — these are two independent dimensions that together give a complete picture of technological displacement. Occupation names are sourced directly from Job Bank's canonical titles; tooltips show employment requirements scraped from Job Bank's requirements pages.

---

Data sources

All data is publicly available under the Open Government Licence – Canada.

Source	Publisher	What it provides
Canadian Occupational Projection System (COPS) 2024–2033	ESDC	Employment 2023, projected job openings 2024–2033, future & recent labour market conditions
3-Year Employment Outlooks 2025–2027	ESDC / Job Bank	Job outlook ratings (Shortage / Balance / Surplus) by occupation and province
Table 98-10-0586-01 (2021 Census)	Statistics Canada	Median annual employment income by NOC 2021 unit group, income reference year 2020
National Occupational Classification (NOC) 2021	Statistics Canada	Occupation titles, 5-digit codes, TEER levels, major group structure
Job Bank	ESDC	Canonical occupation titles, employment requirements, and profile URLs (scraped via scrape_jobbank.py)

---

The occupations

The 516 occupations come from the COPS 2024–2033 summary CSV, which lists every unit group in the NOC 2021 classification with labour market projections. The "All Occupations" aggregate row is excluded; so are TEER and major-group subtotals. What remains is 516 individual unit groups.

By NOC major category:

Category	Occupations
Trades, transport and equipment operators	94
Manufacturing and utilities	68
Natural and applied sciences	66
Sales and service	61
Education, law and social services	52
Business, finance and administration	51
Health occupations	43
Art, culture, recreation and sport	37
Natural resources and agriculture	24
Management occupations	15
Other	5

Employment coverage: 485 of 516 occupations have employment data (31 have N/A due to small sample sizes or suppression). The 485 with data represent 20.1 million jobs — the full Canadian employed labour force.

---

AI exposure scoring

Each occupation is scored 0–10 on a single axis: how much will cognitive AI reshape this occupation?

The score measures cognitive/digital automation only — language models, AI agents, and RPA tools acting on information-processing tasks. It explicitly excludes industrial robots, autonomous vehicles, and physical automation machinery. A welder scores low even though welding robots exist; a long-haul truck driver scores low even though autonomous vehicles are coming. Physical automation risk is surfaced separately (see Robotics below).

The score captures both direct automation (AI doing the cognitive tasks currently done by humans) and indirect effects (AI making each worker so productive that fewer workers are needed). A key signal is whether the work product is fundamentally digital — if a job can be done entirely from a computer, AI exposure is inherently high.

Scoring is done by an LLM (GPT-4o-mini via OpenAI) reading a structured Markdown description of the occupation. The description is generated from COPS data and includes: NOC code, TEER level, major category, employment figures, labour market conditions, and projected job openings.

Calibration anchors:

Score	Tier	Canadian examples
0–1	Minimal	Underground miners, roofers, landscapers, oil field workers
2–3	Low	Electricians, plumbers, firefighters, welders, heavy equipment operators
4–5	Moderate	Registered nurses, police officers, veterinarians, social workers
6–7	High	Teachers, accountants, engineers, HR managers, journalists
8–9	Very high	Software developers, graphic designers, translators, paralegals
10	Maximum	Data entry clerks — routine digital processing with no physical component

Current weighted average (job-weighted, all 516 occupations): 3.8 / 10.

---

Robotics and physical automation

Physical automation affects a different set of occupations than cognitive AI — primarily those scoring low on the AI Exposure axis. The tooltip for every occupation shows a Robotics Risk level (Low / Moderate / High / Very High) based on its NOC sector.

NOC Sector	Robotics risk	Drivers
Manufacturing and utilities	Very High	Auto assembly robots (ON), food processing and packaging lines
Trades, transport and equipment operators	High	Autonomous truck platforms, automated equipment
Natural resources and agriculture	High	Mining haul trucks (Caterpillar AutoMine), robotic harvesters
Sales and service	Moderate	Self-checkout (40%+ penetration), robotic food prep
All other sectors	Low	Physical automation has minimal direct impact

Research basis: Brookfield Institute (2016) — 42% of Canadian jobs at high automation risk; Bank of Canada (2018) — ~2M jobs at high risk; OECD (2023) — ~27% at high risk; Acemoglu & Restrepo (2020) — each robot per 1,000 workers reduces employment-to-population ratio 0.18–0.34%.

A set of dual-threat occupations face both cognitive AI and physical automation: transport truck drivers, cashiers, agricultural workers, warehouse pickers, postal workers, food processing labourers.

---

Pipeline

Eight scripts take you from raw government data to the website:

build_occupations.py   →  occupations.json   (516 NOC occupations + fallback URLs)
build_jobbank_urls.py  →  occupations.json   (upgrades to direct Job Bank profile URLs)
scrape_jobbank.py      →  occupations.json   (adds canonical titles + requirements)
generate_pages.py      →  pages/*.md         (one Markdown file per occupation)
make_csv_ca.py         →  occupations.csv    (pay, jobs, outlook, education)
score.py               →  scores.json        (AI exposure scores — cognitive AI only)
build_site_data_ca.py  →  site/data.json     (merged, ready for frontend)
make_prompt.py         →  prompt.md          (LLM-ready research document)

See process.md for a detailed walkthrough of every calculation, data join, and design decision.

---

Key files

File / Directory	Description
occupations.json	516 occupations: NOC title, canonical Job Bank title, NOC code, category, slug, URL, employment requirements
occupations.csv	One row per occupation: median pay (CAD, 2020), employment 2023, outlook, education
scores.json	AI exposure scores 0–10 with LLM rationale, keyed by slug
pages/	516 Markdown files — one per occupation — used as LLM input for scoring
prompt.md	Full research document (all 516 occupations + robotics analysis) for pasting into an LLM
data/cops_summary.csv	COPS 2024–2033 summary (downloaded, 516 + aggregate rows)
data/census_wages/98100586.csv	Statistics Canada 2021 Census median annual employment income by NOC 2021 unit group
data/outlook_ca.xlsx	3-year employment outlooks by occupation × province/region
site/index.html	Self-contained frontend: treemap, scatter view, filter chips (incl. Robotics Risk), sliders, light/dark mode
site/about.html	Methodology & Stats page with data sources, robotics research, and verified statistics
site/data.json	Compiled dataset loaded by the frontend
build_occupations.py	Step 1 — build occupation list from COPS
build_jobbank_urls.py	Step 1b — fetch Job Bank concordance IDs, update profile URLs
scrape_jobbank.py	Step 1c — scrape canonical titles and employment requirements from Job Bank
generate_pages.py	Step 2 — generate Markdown descriptions
make_csv_ca.py	Step 3 — compile structured CSV (wages from 2021 Census)
score.py	Step 4 — LLM scoring (OpenAI GPT-4o-mini); cognitive AI exposure only
build_site_data_ca.py	Step 5 — merge into site/data.json
make_prompt.py	Step 6 — generate prompt.md (all data + robotics research in one LLM-ready document)

---

Differences from the original (karpathy/jobs)

Dimension	US — karpathy/jobs	Canada — this repo
Occupation standard	SOC (Bureau of Labor Statistics)	NOC 2021 (Statistics Canada / ESDC)
Occupation count	342	516
Data acquisition	Playwright scraping (BLS website)	CSV/XLSX downloads (Open Government Portal)
Employment year	2024	2023
Currency	USD	CAD
Wage source	BLS OOH pages (scraped directly)	Stats Canada 2021 Census median annual employment income (2020 reference year, CAD)
Outlook format	% employment growth over 5 years	Labour market conditions: Shortage / Balance / Surplus
Education classification	BLS entry-level education labels	NOC TEER (0–5)
Occupation descriptions	Rich BLS HTML pages (duties, pay charts, work environment)	Generated Markdown from COPS data
Education labels	BLS entry-level education labels	Human-readable (High school / Vocational / College / University+)
Occupation titles	BLS OOH canonical titles	Job Bank canonical titles scraped via scrape_jobbank.py
Employment requirements	Scraped from BLS OOH pages	Scraped from Job Bank requirements pages (462/516 occupations)
Methodology page	—	site/about.html with verified job-weighted statistics
Light/dark mode	—	Toggle in sidebar, persisted to localStorage
Scatter view controls	—	Exposure score range slider (min–max 1–10)

The US version scrapes the BLS website with Playwright because BLS blocks bots; the Canadian version downloads open CSV/XLSX files directly — no browser automation required after the initial setup.

---

Setup

uv sync

Requires an OpenAI API key for the scoring step:

# .env
OPENAI_API_KEY=your_key_here

Running the full pipeline

# 1. Build the occupation list from COPS data
uv run python build_occupations.py

# 1b. Fetch Job Bank concordance IDs for direct profile URLs (~30s)
uv run python build_jobbank_urls.py

# 1c. Scrape Job Bank for canonical titles + employment requirements (~3 min)
uv run python scrape_jobbank.py

# 2. Generate Markdown pages (used as LLM input)
uv run python generate_pages.py

# 3. Compile structured CSV with pay, jobs, outlook
uv run python make_csv_ca.py

# 4. Score AI exposure for each occupation  (needs OpenAI key)
uv run python score.py

# 5. Merge everything into the site dataset
uv run python build_site_data_ca.py

# 6. Serve locally
cd site && python -m http.server 8000

Steps 1–3 and 5 are deterministic. Step 1b queries the Job Bank Solr API for concordance IDs (462/516 get direct profile links; 54 fall back to a search URL). Step 1c scrapes each profile's summary and requirements pages (~924 HTTP requests at 0.1 s each, 3 min total); it saves progress every 50 occupations and is safe to interrupt and resume. Step 4 is the only paid API call ($0.50 for all 516 occupations with GPT-4o-mini) and is also incrementally resumable.

---

Credits

• Original concept and code: Andrej Karpathy — karpathy/jobs
• Canadian data: Statistics Canada · Employment and Social Development Canada · Job Bank
• AI scoring: GPT-4o-mini via OpenAI API
• Occupation classification: National Occupational Classification (NOC) 2021

---

How to cite this project

If you use this repository in research, policy work, reporting, or derivative tools, please cite it as:

Krunal. (2026). AI Exposure of the Canadian Job Market [Computer software]. GitHub. https://github.com/krunal16-c/jobs

You can also use this BibTeX entry:

@software{krunal2026_ai_exposure_canada,
	author = {Krunal},
	title = {AI Exposure of the Canadian Job Market},
	year = {2026},
	url = {https://github.com/krunal16-c/jobs},
	note = {Adapted from karpathy/jobs using Canadian data sources (Statistics Canada, ESDC, Job Bank)}
}