The same data, cut six ways.
One measured signal, how AI is actually used, sliced by pay, gender, education, UK region, across 121 countries, and by who over-indexes on it. Each cut is one finding and one chart. The full deep-dive sits under each.
The better a job pays, the more AI assists rather than replaces.
Split every occupation into pay fifths and line up the measured augmentation share. It climbs monotonically, from 42.1% in the lowest-paid fifth to 50.7% in the highest, a 8.6 point tilt. The famous claim that AI automates the low-paid and augments the high-paid half holds: true in direction, gentle in size. Even the bottom fifth is 42% augmentation.
Measured: AEI augmentation share × ONS/ASHE pay, Apr+May 2026, CC-BY. Claude users only; US to UK SOC crosswalk drops ~9% of codes; usage is not a count of jobs lost. The full earnings cut →
The measured twist: male-dominated jobs lean more to automation than women's do.
The usual story says women's jobs automate first. The measured usage says the reverse. Male-dominated occupations sit at 43.8% augmentation (so more automation-leaning), female-dominated ones at 49.3%. Across 356 occupations the correlation between female share and augmentation is positive (+0.31). Sector sorting, not gender itself, is doing the work.
Measured: AEI augmentation share, Claude.ai usage, Apr+May 2026, CC-BY. Claude users only; automation-leaning usage is how AI is used, not a count of jobs lost. The full gender cut →
The more education a job asks for, the more AI assists it.
Line the measured augmentation share up against the qualification a role requires and a clean ladder appears, from 39% at entry and low-GCSE level to 53% at postgraduate. School-gated work sits near the automation end, degree work near the assisted end. It is the same tilt as pay, read through the credential axis.
Measured: AEI augmentation share, employment-weighted within each qualification band over AEI-covered roles. Claude users only; Apr+May 2026; usage is not a count of jobs lost. The full education cut →
Most-exposed is not most-at-risk.
London has by far the most workers in high-exposure roles, 39% of its workforce. But the largest pools of at-risk wages sit in the bigger mid-skill belts: South East tops the list at £7.6B. Same AI, different regional economies, different exposure. Concentration and damage are not the same map.
Employment-weighted regional aggregates. Exposure is the 0-10 scored axis, validated against the academic Felten index at rho 0.812; wages at risk = employment × pay in high-exposure roles. The full regional cut →
Where AI is barely used it automates. Where it's everywhere it augments.
The same measured split, drawn across 121 countries. The horizontal axis is how intensely each country uses Claude (a proxy for adoption, which tracks income); the vertical axis is the augmentation share. The two rise together. Low-adoption countries lean toward automation, high-adoption ones toward augmentation, and the UK sits high on both.
Measured: AEI, real Claude.ai usage, Apr+May 2026, CC-BY. Claude users only, so adoption reflects Claude's footprint (a proxy for AI adoption, itself a proxy for income); augment vs automate is how AI is used, not a count of jobs lost. The full international cut →
Claude usage doesn't mirror the workforce. Writers are a sliver of Britain's jobs and a giant slice of its AI use.
Compare each occupation's share of all Claude usage with its share of the UK workforce. Authors, writers and translators are 0.33% of workers and 5.74% of usage, 17.43× their size. Journalists, actuaries and economists run at 10× or more. At the other end, Britain's biggest physical occupations barely register: lorry drivers are 0.9% of the workforce and effectively 0% of usage.
“Sales related occupations n.e.c.” is greyed as a mapping artefact: the US→UK crosswalk routes five whole US sales categories into this one small residual code (50.8k workers), so its ratio measures the crosswalk, not the occupation. The rule: n.e.c. residual codes stay on the chart, flagged, and never headline it.
Some usage is people asking AI to teach them the domain, about 9% on average. It concentrates in rule-heavy professions:
Share of each occupation's Claude usage that is learning. Occupations work-filtered (≥45%), 10k+ jobs.
Measured: the AEI usage-share metric vs ONS employment, Apr+May 2026, CC-BY. Global usage fingerprint applied to UK weights; 75.8% of usage maps to a UK SOC; ranked rows filter to work-use ≥25% and 50k+ workers per the methodology display rule. Every role on the map →