What the tool is, and isn't
It's a map of how much of each UK occupation's work a language model can do today. The unit is the SOC 2020 4-digit occupation group, employment comes from ONS ASHE, and the exposure scores come from Eloundou et al's 2023 paper.
It is not a prediction that any specific role disappears. It is a snapshot of where AI has already arrived. Which tasks get automated, which get augmented, and which stay human is the more useful question; the treemap is a starting point for that conversation, not the end of one.
Three exposure variants, and which one leads the headline
The per-task scores come with three aggregation recipes. Alpha counts tasks the model can do directly, on its own. Beta credits partial exposure when the model is paired with tools. Gamma counts any exposure, direct or with tools. The three numbers are genuinely different readings of the same data.
- Alpha: £107.2bn UK wage exposure. The strict "can an LLM replace this task unaided" reading.
- Beta: £332.75bn. Mid-case: partial credit for tasks where LLM+tools is the real story.
- Gamma: £558.32bn. The most generous; closest to how the tool is actually being used in 2026.
The landing page headline uses gamma because that's the shape of the Eloundou paper's own US $711bn figure, and because it's a more honest read of what "exposed to AI" means in 2026 (tools are normal now). The tiles and the per-occupation scores use alpha, which keeps the individual readings conservative. Both are in the data; the UI just picks one for each surface.
How occupations are joined to task scores
UK SOC 2020 unit groups are bridged to O*NET tasks through a chain of public crosswalks:
- ONS SOC 2020 Volume 2 coding index (~32,000 job titles) gives a modal ISCO-08 code per UK SOC unit group.
- US BLS ISCO-08 ↔ SOC 2010 crosswalk maps each ISCO to one or more US SOC 2010 codes.
- O*NET-SOC codes truncated to 6 digits match US SOC 2010.
- Per-task Eloundou scores are aggregated up that chain: importance-weighted mean within an O*NET-SOC, then unweighted means as we aggregate up through SOC 2010 and ISCO-08 back to UK SOC 2020.
That chain loses coverage at every step. The final number: 379 of 412 UK SOC unit groups have a score (91.99%). 326 of those also have ASHE employment data, so 299 are fully scorable for wage exposure.
What's honest
- Eloundou 2023 was labelled against early GPT-4. The capability frontier has moved; a fresh scoring pass with Claude (April 2026 capability) is planned for a Phase 2 update and will likely push the numbers up.
- ASHE is employees only. Self-employed workers aren't counted; roles with heavy freelance share (creative, consulting, trades) are under-weighted in the wage bill.
- The ISCO-08 bridge is lossy. The coding index gives multiple ISCO codes per UK SOC; we take the modal one. Other choices (weighted, multi-valued) give slightly different numbers.
- "Exposure" ≠ "job goes away." It's a ceiling on how much of the task inventory a model can touch. How much of that gets automated, augmented, or stays human is a design choice, not a technical one.
Sources
- ONS SOC 2020 Volume 1 — structure and descriptions of unit groups
Office for National Statistics (2025). Standard Occupational Classification 2020 (SOC 2020), Volume 1: Structure and descriptions of unit groups, revised 3 December 2025. Licensed under the Open Government Licence v3.0.
- ONS SOC 2020 Volume 2 — coding index (contains SOC 2020 ↔ ISCO-08)
Office for National Statistics (2025). Standard Occupational Classification 2020 (SOC 2020), Volume 2: The coding index, revised 3 December 2025. Licensed under the Open Government Licence v3.0.
- O*NET 30.2 database (tab-delimited, ZIP)
National Center for O*NET Development. O*NET 30.2 Database. Employment and Training Administration, U.S. Department of Labor. Used under CC BY 4.0.
- Eloundou et al — per-task AI exposure scores (full_labelset.tsv, alpha = direct exposure)
Eloundou, T., Manning, S., Mishkin, P., & Rock, D. (2024). GPTs are GPTs: Labor market impact potential of LLMs. Data: github.com/openai/GPTs-are-GPTs (commit 0471612f).
- US BLS — ISCO-08 ↔ US SOC 2010 crosswalk
U.S. Bureau of Labor Statistics (2012). ISCO-08 to 2010 SOC Crosswalk. Public domain.
- ONS ASHE Table 14 — occupation (4-digit SOC 2020), 2025 provisional (includes employment count and gross annual pay)
Office for National Statistics (2025). Annual Survey of Hours and Earnings (ASHE), Table 14: Occupation (4-digit SOC 2020), 2025 provisional. Licensed under the Open Government Licence v3.0.
Full writeup (PDF)
The full technical methodology - source-by-source notes, aggregation formulas, coverage drop-off at each join, code fragments - is in the PDF below. It's generated from the same manifest as this page.
Get in touch
Notice a SOC that's badly mapped, a task score that doesn't match your day, or a methodology choice you'd make differently? Email hello@alexlockey.com. I'll fold the better calls into the next build.