Work · Stable Future · 2026
AI exposure for every UK job
The best research on AI and jobs scores tasks in a US database. UK jobs are coded differently, and there was no official link between the two at the detail needed. I built one, and a score for each of 412 UK job groups.
- Role
- Sole builder
- When
- 2026
- Scale
- 1,016 occupations · 17,951 tasks · 412 UK job groups
- Stack
- Python, pandas, parquet, LLM agents, pytest
The problem
Almost all task-level research on AI and jobs uses O*NET, the US government’s database of occupations and their tasks. UK jobs are classified by the ONS (SOC 2020), and there is no official link between the two at the level of detail this work needs.
To give UK families evidence rather than headlines, I needed a score for every UK job group that I could defend line by line.
What I built
- Snapshot the sources. O*NET 30.0, ONS SOC 2020 and published research scores, stored unchanged with checksums.
- Crosswalk. Give each of 1,016 O*NET occupations a primary UK job group. LLM agents propose links from the official job descriptions.
- Review. Automated checks and reviewed corrections, then a comparison with a published UK mapping as an internal test.
- Score. Two measures per job: exposure (can AI learn the tasks?) and substitution (would AI replace the worker or help them?).
- Release. Versioned parquet tables with a manifest of row counts and checksums, feeding the website.
Exposure uses Tomei and Klein Teeselink’s published scores for 17,951 tasks, weighted by each task’s importance to the job. Substitution follows Pizzinelli and colleagues’ work-context measure as adapted by Liu, Wang and Yu.
The key decision: not re-scoring 17,951 tasks myself
I built my own LLM scoring run and calibrated it on a 50-task sample. It agreed with the published scores at r = 0.854 and would have cost about $38 a run. A run would have bought provenance and fresher data, not a different ranking. So I shipped the published scores with full credit and parked my run until it pays for itself.
That run has since grown into its own project: GPTs are GPTs, re-rated for 2026.
Engineering
- Raw snapshots are never edited. A new release is a new file with a new checksum.
- Paid model calls are cached on a hash of content and prompt, so a rebuild never pays twice.
- 27 test files cover the crosswalk, scoring and report builders.
- Every published number keeps its source and its caveat.
Where it went
The scores power the free career check on stablefuture.uk, and they fed the next study: are AI-exposed jobs losing young UK workers?
The same method answers a business question: which of a team’s tasks can AI speed up? That became the opportunity scanner.