Method and sources
How every number is built
Every figure in this tool comes from a named, public dataset. Nothing from the book is presented as data. Each chart has a CSV below, and the three scripts that build everything from the raw files are published too (downloads), so a data journalist can rebuild any number.
What the tool does not claim
- Exposure is not loss. Exposure measures how far the abilities a job relies on overlap with what common AI applications can do. It does not say whether AI will help with the job or replace it.
- No forecasts. There are no predictions of jobs lost, by year or by number.
- Correlation stays correlation. The fall in job adverts since November 2022 is shown, not blamed on AI. Adverts also move with the economy, interest rates and the cost of hiring.
- The DfE scores date from 2023. They measure which abilities AI applications overlap with, not what employers are doing. In DfE’s words, the analysis “measures the exposure of jobs to AI, rather than distinguishing whether a job will be augmented (aided) or replaced (substituted) by AI.” The ILO’s 2025 measure is shown alongside as a check.
- Advert data is one signal, not a measure of hiring. See the cross-checks below.
- No named employers appear anywhere in the tool.
The backbone: SOC 2020 occupation codes
Everything joins on the ONS Standard Occupational Classification 2020 at four-digit level: 412 unit groups, from “Chief executives and senior officials” (1111) to “Other elementary services occupations n.e.c.” (9269). The job search also uses the ONS SOC 2020 coding index, which maps about 29,000 everyday job titles to these codes.
Sources
| Measure | Source | Edition | Detail |
|---|---|---|---|
| AI exposure | DfE Unit for Future Skills, The impact of AI on UK jobs and training, Annex 1 | November 2023 | AI occupational exposure score (AIOE, all AI applications), standardised. SOC 2010, 365 unit groups. |
| SOC 2010 to 2020 | ONS relationship tables | July 2021 | Share of each SOC 2010 unit group coded to each SOC 2020 unit group in dual-coded Census and LFS samples. |
| Pay | ONS ASHE Table 14.7a | 2026 provisional (October 2026) | Median annual gross pay, full-time employees; mean annual gross pay, all employee jobs. |
| Employment | ONS Annual Population Survey, Nomis NM_218_1 | July 2025 to June 2026 | People in employment, and employees, by SOC 2020 four-digit, UK and 12 regions. |
| Job adverts | ONS labour demand volumes by SOC 2020 (Textkernel), Table 3 | January 2017 to September 2026 (October 2026) | New online job adverts each month by SOC 2020 four-digit and region. Official statistics in development. |
| Age profile | ONS Census 2021, occupations by age and sex | March 2021 | Share of each job held by people aged 16 to 24. England and Wales. |
| Degree share | ONS APS ad hoc 3501 | April 2025 to March 2026 | Share of each job whose highest qualification is a degree or equivalent. |
| Advert cross-checks | ONS X06 single month vacancies and ONS online job advert estimates (Adzuna) | November 2026; Adzuna final edition October 2024 | Total vacancies from ONS’s survey of employers, by industry; a second advert source by job category. |
| Second exposure measure | ILO Working Paper 140, Generative AI and jobs, with the ONS SOC 2020 coding index | 2025; index version 14, August 2026 | ILO mean exposure score by ISCO-08 four-digit, mapped to SOC 2020 through the ISCO code the ONS index gives each job title. |
How the measures are built
What the exposure score is
DfE’s AI Occupational Exposure score follows the method of Felten, Raj and Seamans (2021). It links ten common AI applications, such as image recognition, language modelling and translation, to the human abilities each job relies on, using the US O*NET database of abilities, and scores each job by how much those abilities overlap with what the applications can do. The applications were chosen from those where the Electronic Frontier Foundation recorded progress from 2010 onwards, so the measure predates ChatGPT. DfE also publishes a separate score for large language models; we use the all-AI score. The two agree closely (correlation 0.98), and 69 of the 82 most exposed jobs are in the top fifth on both.
Exposure on SOC 2020
The DfE scores use the older SOC 2010 codes. Each SOC 2020 job takes a weighted average of the SOC 2010 scores of the jobs its workers were coded to before, weighted by the number of people in the ONS dual-coded Census sample (men and women combined; the LFS sample where the Census one is empty). 363 of the 412 SOC 2020 jobs take 90% or more of their workers from a single SOC 2010 job, so most scores carry over directly. Four jobs have no score, because DfE published none for their SOC 2010 equivalents: elected officers and representatives, officers in the armed forces, other ranks in the armed forces, and waste disposal and environmental services managers.
The ILO’s 2025 generative AI scores, mapped separately, rank jobs in much the same order: the rank correlation between the two measures is .
Highly exposed
The most exposed fifth of the 408 scored jobs (82 jobs). It is a relative measure: these are the jobs whose abilities overlap most with what AI applications can do, compared with other UK jobs.
Pay comparisons
“The most exposed fifth pay £53,093” means each job’s median full-time annual pay, averaged across the jobs in that fifth and weighted by how many people do each. ASHE annual pay covers “employees on adult rates who have been in the same job for more than a year”.
The ladder: junior, mid-level and senior
Each field is shown as rungs that follow the SOC 2020 major groups, which ONS links to skill level: senior (1, directors and managers), mid-level (2, professionals), junior (3, associate professionals and technicians) and entry and support (4 administrative, or 6 and 7 for support roles). Which jobs belong to which field is our judgement, set out in full in ladders_mapping.csv. A rung’s exposure is the average of its jobs’ positions among all UK jobs, weighted by people. Some fields have no job at a given rung; we leave it out rather than borrow one. Rungs show skill levels within a field, not career routes, and where a rung includes jobs from outside its usual group the page says so.
Wages riding on it
Employees in each job (APS) multiplied by mean annual gross pay for all employee jobs (ASHE; employees in the same job for more than a year). Where ONS does not publish a four-digit mean, the three-digit (or, once, two-digit) figure fills the gap; the CSV marks every filled value. Added up across all jobs this comes to about £1.2 trillion a year, in line with total UK wages and salaries. Regional wage bills use UK pay with each region’s employment mix, so they show where exposed jobs are concentrated, not regional pay differences.
Adverts, and why to treat them with care
The door chart shows new online job adverts collected by Textkernel and published by ONS as official statistics in development. Adverts are not vacancies or hires.
Withheld months and ONS warnings. ONS withholds some monthly cells as unreliable, including March 2026 for every occupation. We treat withheld cells as missing, never as zero. Each comparison uses the average month over the months ONS published, and each point on the chart compares only the jobs published that month with the same jobs’ 2019 average, so a withheld cell is left out rather than counted as zero. Where a whole month is withheld (March 2026), or a job’s months are, the chart shows an estimate, drawn dotted and flagged in the CSVs: the straight line between the published months either side, scaled by how that month compared with the same line in normal past years (2017 to 2025, leaving out the 2020 and 2021 lockdowns; the median of those years, needing at least three). For all adverts, March has typically run about 10% above the average of February and April. Estimates are for the chart only; every figure we quote uses published months alone. ONS also says the series from October 2025 to September 2026 is partly imputed, after a major source failed, and that the latest months carry more uncertainty. Our headline window is therefore October 2024 to September 2025, the latest 12 months before that period; later months are shaded on the chart. A job’s comparison is left blank where its base year had fewer than 100 adverts, and a quarter is left blank where ONS withheld any of its months.
We tested this series against two others before publishing it:
- The fall is much steeper than in ONS’s survey of employers. against 2019: Textkernel adverts ; ONS X06 vacancies . January to September 2024 against 2019: Textkernel , Adzuna adverts , ONS vacancies .
- A second advert source ranks types of job differently. Across ten Adzuna categories matched to SOC 2020 groups, January to September 2024 against 2019, the rank correlation between the two sources is . The matching is approximate, because Adzuna uses job-board categories. The full comparison is in advert_crosscheck.csv.
- ONS’s survey by industry (latest 12 months against 2019) shows the largest falls in retail, other services, arts and hospitality, which are less exposed sectors. Industry is not occupation, so this does not contradict a finding about exposed jobs, but it is a reason not to read the advert lines as hiring.
We therefore show adverts as one signal, and point to published studies for stronger evidence on entry-level work.
Who does the job
The share aged 16 to 24 comes from Census 2021 (England and Wales), so it describes the workforce before ChatGPT. The degree share comes from the APS for April 2025 to March 2026.
Missing and suppressed data
- ASHE does not publish median full-time pay for 50 of 412 jobs, where ONS judges the estimate unreliable (coefficient of variation above 20%) or disclosive. Those jobs are left off the pay chart but stay in every other view.
- APS publishes UK employment for all 412 jobs. By region, ONS suppresses unreliable four-digit cells: for employees, which the map uses, from 25 (South East) to 143 (Northern Ireland), with 61 in London. Regional shares cover the published cells only. The suppressed cells are mostly small, less exposed jobs, so regional shares in exposed work are, if anything, slightly high: filling the gaps roughly moves London from 52.2% to about 51.7%.
- Census 2021 suppresses 118 small age cells; they are counted as zero, which slightly lowers the under-25 share for small jobs.
- Degree share is left blank for jobs with fewer than 5,000 people with a known qualification; cells ONS marks as too small are counted as zero.
- The ILO score for each SOC 2020 job is a simple average over the job titles the ONS coding index lists for it.
- The 39% of wages in highly exposed jobs is measured against all 412 jobs; against the 408 with a score it is 39.6%.
Downloads
- occupations_master.csv: one row per SOC 2020 job, with every measure
- employment_by_region_and_occupation.csv: APS employment by region and job
- regions_exposed_wage_share.csv: the map
- door_adverts_by_group.csv: the door chart, monthly counts and index
- adverts_quarterly_by_occupation.csv: quarterly adverts for every job
- ladders_mapping.csv: which jobs sit on which rung of each field
- advert_crosscheck.csv: the two advert sources and the ONS vacancy survey compared
- fetch.sh, build.py and export.py: the three scripts that download every source and build every number (Python with pandas)
- join_log.md: every join, with counts of matched, missing and suppressed rows
Use and credit
Free to use and embed. Please credit: “Data: ONS, DfE. Analysis: Chris Simmance, author of The Next Rung.” Questions about the data: chris@simmance.ai.
Contains public sector information licensed under the Open Government Licence v3.0. ILO data © International Labour Organization.