Which entry-level jobs are most and least exposed to AI/automation — and what does that mean for young people entering the UK labour market right now?
An interactive treemap and data pipeline scoring UK occupations (SOC 2020) for AI and automation exposure, scoped to entry-level roles accessible to under-25s and the ~900,000-strong NEET (Not in Education, Employment or Training) population. A UK adaptation of karpathy/jobs.
⚠️ The scoring is directional, not empirical. Read a score as “likelihood of significant role change or headcount reduction by 2030” — not “this job will disappear”. It is an LLM’s structured judgement, useful for comparison and conversation, not a forecast.
A treemap of UK occupations grouped by SOC 2020 major group, tile area proportional to UK employment, coloured green (lower exposure) → red (higher exposure). Toggles:
- Colour by: AI exposure, or employment growth (green = growing, red = declining)
- Filters: entry-level only · high-risk entry-level · public sector · apprenticeship available
- Hover: score, risk band, median pay (£), employment, recent growth, the model’s rationale, and — for high-risk entry-level roles — a suggested safer pivot.
All 412 SOC 2020 unit groups are scored across all 9 major groups. The entry-level slice (SOC major groups 6–9 — caring/leisure, sales/customer service, process/plant operatives, elementary occupations) is where the NEET population disproportionately enters work and is the primary focus; the remaining groups provide context. Use the Entry-level only filter to isolate the NEET-relevant subset.
LMI for All (the spec’s original source) was decommissioned at the end of October 2025, so the pipeline sources data from:
| Field | Source |
|---|---|
| SOC 2020 codes, titles, major groups | NOMIS NM_218_1 codelist (APS occupation SOC2020) |
| UK employment + recent growth | NOMIS NM_218_1 (Annual Population Survey), UK, 2021→latest |
| Median annual / hourly pay | ONS ASHE Table 14 (4-digit SOC 2020, latest SOC2020 release) |
| Role content for scoring | National Careers Service job profiles (httpx scrape) |
Growth is a recent year-over-year APS change, not a 5-year projection, and is noisy for small occupations — labelled honestly in the tooltip and capped in the colour scale.
uv sync
uv run playwright install chromium # optional; the scraper uses plain HTTP
cp .env.example .env # add ANTHROPIC_API_KEY for the scoring step
uv run python scripts/fetch_soc.py # → data/occupations.csv (all ~370 unit groups)
uv run python scripts/scrape_ncs.py --all # → data/raw/*.html (all matched NCS pages)
uv run python scripts/parse_pages.py --all # → data/pages/*.md (NCS content or SOC stub)
uv run python scripts/score.py # → data/scores.json (Anthropic Batch API)
uv run python scripts/merge.py # → data/site/data.json + site/data.json
uv run python scripts/validate.py # coverage / sanity report
cd site && python -m http.server 8000 # open http://localhost:8000- Scoring uses the Anthropic Message Batches API (Sonnet 4.6,
max_tokens=600), one request per occupation with a shared, prompt-cached system prompt and a strict JSON-schema output. ~412 occupations costs roughly $8–10. RequiresANTHROPIC_API_KEYin.env. - Where no NCS profile matches a SOC unit group,
parse_pages.pywrites a compact title + major-group stub so every occupation in the slice is still scored.
- Which entry-level roles have the lowest AI exposure right now — where should a school leaver or NEET young person focus?
- Which entry-level roles are already high-risk and likely to see headcount decline?
- What are the best pivot paths — low-barrier moves from a high-risk entry role to an adjacent lower-risk one?
- Does the public sector provide a meaningful buffer for entry-level workers, or a false sense of security?
- Which SOC major groups have the healthiest mix of low exposure + growing employment + accessible entry?
Use the filters and the two colour modes to read these off the treemap directly.
- Directional, not empirical — see the note at the top.
- “AI exposure” ≠ “job loss”. The framing is role change / headcount reduction, not disappearance.
- Pay is from the latest SOC2020 ASHE release and may lag the current year; some 4-digit cells are suppressed (shown as n/a).
- The focus is realistic entry points for under-25s — law, finance, medicine and management consulting are deliberately not the centre of attention.
scripts/ fetch_soc · scrape_ncs · parse_pages · score · merge · validate (+ common.py)
data/ occupations.csv · scores.json · site/data.json · raw/ · pages/
site/ index.html (D3 treemap)