by Claude Opus 5.5

What is the most defensible way to estimate “AI exposure” of occupations in the UK, what are the methodological pitfalls, and why do recent UK exposure estimates differ so widely?

The most defensible approach starts with tasks, not job titles. Score each task against what AI can actually do, weight the tasks by how much time UK workers spend on them, combine several independent measures instead of trusting one, and test the result against real outcomes such as vacancies and actual AI usage. Headline UK figures differ so widely mainly because they measure different things: a share of tasks, a worst-case job count, a displacement path over time, or growth in AI-related jobs.

What “exposure” means, and what it doesn’t

An exposure score says how far an occupation’s tasks overlap with what AI can do. It is a map of where AI could have an effect, not a forecast of job losses. Whether exposure turns into displacement, augmentation or new demand depends on things the score leaves out: the cost of adoption, regulation and liability, how much demand grows when output gets cheaper, and whether firms redesign work. The IPPR’s 2024 study shows the gap neatly. Starting from the same exposure estimates, its scenarios range from zero job losses (full augmentation, with GDP up 13.4%) to 7.9m (full displacement), with a central case of 4.4m.

The main methods

Ability-based: Felten, Raj and Seamans’ AI Occupational Exposure (AIOE, 2021). This links progress in ten AI applications, such as image recognition and language modelling, to the 52 abilities that the US O*NET database assigns to each occupation. A 2023 version focused on language models put telemarketers and post-secondary teachers near the top. It is transparent and widely reused, including in the Department for Education’s 2023 UK report and in the IMF’s work. Its weakness is that abilities are a coarse proxy for what people actually do, and the 2021 version predates generative AI.

Patent-based: Webb (2020). This matches the text of AI patents to O*NET task descriptions. Its notable finding was that AI, unlike robots or earlier software, is most exposed to high-skill, high-wage work. Patents show what inventors are aiming at, which is not the same as what firms deploy.

Rubric-based task scoring: Eloundou et al., “GPTs are GPTs” (2023; Science, 2024). Humans and GPT-4 rated each O*NET task by whether an LLM, alone or with extra software, could cut the time it takes by at least half. The study found about 80% of US workers have at least 10% of their tasks exposed, and about 19% have at least half. Separating “LLM alone” from “LLM plus tools” was an important step, because the second category depends on integration work that may never happen.

International organisations. The ILO’s refined index (May 2025) scored tasks in the ISCO-08 occupational classification using an AI model, a survey of 1,640 workers in Poland and expert panels. It found that one in four workers globally, and 34% in high-income countries, are in occupations with some GenAI exposure, but only 3.3% are in the highest-exposure group. The ILO’s own conclusion is that transformation is more likely than replacement. The IMF (January 2024) adjusted AIOE for “complementarity”, meaning how far a job’s social and physical context shields it from substitution. It estimated that almost 40% of global employment is exposed, rising to about 60% in advanced economies, and that roughly half of exposed jobs may gain from AI rather than lose.

Usage-based measures. The Anthropic Economic Index maps conversations on Claude to O*NET tasks and separates automation-style use (delegating the task) from augmentation (working with the model). It measures what people actually do with AI, not what AI could do. Its limits are that it covers one platform, its users skew towards early adopters, and it captures consumer and API use rather than whole occupations.

Job-ad approaches: PwC’s AI Jobs Barometer. PwC compares growth in postings, wages and skill demand across occupations grouped by exposure. Its UK 2026 release found postings in low-exposure occupations grew 2.27 times since 2012, against 0.98 times in the highest-exposure quartile. This is an outcome analysis built on an exposure ranking, and the ranking’s method sits in a separate technical note rather than in the headline material.

UK-native work. Henseke and colleagues’ Generative AI Susceptibility Index links LLM task ratings to the British Skills and Employment Survey. That survey records what UK workers say they do, so the index avoids importing US job content.

Mapping O*NET to UK SOC

Most of these methods depend on O*NET because the UK has no equivalent task database. The DfE’s 2023 study said so explicitly and used a crosswalk to SOC 2010. Getting from O*NET to UK occupations usually means going through the US SOC and the international ISCO classification to UK SOC 2010 or 2020. Every step is many-to-many, so a UK code ends up as an average of several US occupations with different task mixes.

Three problems follow:

  • US job content is imported. A UK paralegal, practice nurse or claims handler may do a different mix of tasks from the nearest US equivalent.

  • Employment weights are uncertain. Weights come from the LFS, which ONS itself says has been volatile since 2023, and which also had coding problems during the move from SOC 2010 to SOC 2020.

  • Differences within an occupation disappear. A junior and a senior accountant get the same score even though AI bears on their work very differently.

The IPPR mapped O*NET to ONS LFS data, and the Tony Blair Institute (TBI) used O*NET’s roughly 20,000 tasks. Both therefore inherit these errors.

Task weighting and thresholds

The second big source of disagreement is arithmetic.

  • Weighting. Some methods treat every task as equally important. Others weight tasks by how important or how frequent workers say they are. An occupation where the exposed task takes an hour a week scores very differently from one where it takes 30 hours.

  • Thresholds. What counts as “exposed” is a choice. The IPPR asked GPT-4 whether a task could be done “at least 50 per cent faster”, while other studies use smaller time savings. Lower thresholds produce bigger headline numbers.

  • Which AI is assumed. The IPPR found 11% of tasks exposed to “here and now” AI but 59% to “integrated” AI connected to company systems. That one assumption moves the result more than any other.

  • LLMs as raters. Ratings shift with prompts and model versions, and models may have seen the scoring rubrics during training. Repeated runs, stored outputs and human checks reduce these problems, but they don’t remove them.

Why the UK estimates differ so widely

  • Coface/OEM (2 Apr 2026). What it actually measures: Share of tasks at risk across 923 occupations and 12 countries. UK headline: Nearly 20% of UK tasks, against about 17% in Germany and the US.

  • IPPR (27 Mar 2024). What it actually measures: Jobs lost in scenarios built from exposed tasks. UK headline: 0 to 7.9m; central 4.4m.

  • TBI (8 Nov 2024). What it actually measures: Time savings, adoption paths and re-employment over time. UK headline: 23.8% of private-sector time saved; 1m–3m jobs ultimately displaced; peak unemployment effect in the low hundreds of thousands.

  • Warwick IER for DSIT (28 Jan 2026). What it actually measures: Growth in jobs involving AI activities, projected from patents and vacancies. UK headline: About 158,000 (2024) to about 3.9m (2035).

These figures are not four answers to the same question:

  • Coface reports a share of tasks. Twenty per cent of tasks is not 20% of jobs.

  • The IPPR’s “8m” is the edge of a range: everything exposed under the integrated-AI assumption is lost and nothing is created. It is a stress test, not a forecast.

  • The TBI adds adoption timing and re-employment. That is why its displacement figure is large but its unemployment effect is small.

  • Warwick IER isn’t an exposure measure at all. It projects job creation using Working Futures-style projections adjusted for AI patent and vacancy trends, and its authors say the patent element is “exploratory”.

Read side by side, the four figures are fairly consistent. Roughly a fifth of current UK tasks are within reach of today’s tools, much more is reachable with integration, and how much of that becomes job loss depends on adoption speed and new demand.

The defensible recipe

  1. Score tasks, not job titles, and keep “the model alone” separate from “the model plus integration”.

  2. Use UK task data where possible, such as the Skills and Employment Survey or vacancy text, and publish the crosswalk where O*NET is unavoidable.

  3. Weight tasks by time or importance, and show how sensitive the results are to the threshold.

  4. Combine measures by rank. Bank of England staff did this in August 2026, taking percentile ranks across Felten, Henseke, the Anthropic Economic Index and Eloundou. A ranking that holds across all four is more reliable than any one score.

  5. Validate against outcomes. That composite correlated strongly with vacancy falls (Spearman ρ = −0.70). Adverts fell 15% in the high-exposure third of occupations, against 6% in the low-exposure third, with customer service down 23% and administrative roles down 22%.

  6. Say what the score can’t show. The same authors concluded that pandemic over-hiring, remote work, labour costs and AI overlapped, so “confident attribution remains premature”.

Bottom line

Use exposure scores to decide where to look, never to count job losses. When you see a UK headline number, first check whether it counts tasks, jobs or new roles, which AI it assumes, and whether it includes adoption and re-employment.

Sources

From AI and Jobs: UK, October 2026