A writing machine. Not a coding one.
Forget the forecasts. Anthropic publishes what millions of real Claude conversations actually do, job by job. Weighted across the whole UK workforce, the three things it most often produces are documents, explanations and advice: 43% of all measured output. Everything a developer would call building software, the apps, scripts, fixes and queries, adds up to about 6%.
Assist the person, or do the task? Almost dead even.
Cross the measured usage with ONS employment and the workforce splits down the middle. Of 23.4M mapped workers, 12.3M are in occupations where usage leans toward doing the task, and 11.1M where it leans toward helping the person. “Leans” is a majority-share cut of how AI is used in a job, not a count of jobs lost.
How that near-even split actually happens.
Every Claude conversation gets tagged with how the person and the model worked together. The single most common mode is directive: hand it a task and take the output, which counts as automation. But the three assisting modes, iterating, learning and checking, together nearly match it. That texture is why the split lands near even: 45.7% of collaboration assists the person, 50.3% does the task.
The lower the pay, the more the AI just does the task.
Split the occupations into pay fifths and the split stops being even. Automation-style usage runs 57.9% in the lowest-paid fifth and falls, step by step, to 49.3% at the top, while the assisting share climbs from 42.1% to 50.7%. Automate the poor, augment the rich. The tilt is real, and it is gentle.
The bottom rung isn't where the fear says it is.
The fear says AI eats the young's tasks first. Measured, automation-style usage is flat across every age band: between 51.7% and 53.8% from 16 to 65+. What is not flat is where the autonomy sits. Only 14.1% of 16–24 employment is in occupations in the top quarter for AI autonomy, against 25% for every other band: 44% under-represented.
AI isn't doing young people's jobs. It may be doing the jobs young people would have been hired to learn. The squeeze shows up at the hiring gate, in entry-level postings and first jobs, and the realised side of that story is on the gap tracker.
AEI usage per occupation crossed with age-band employment (ONS APS ad-hoc 3410, 2025) on the dual-coded SOC crosswalk; coverage 92–96% of each band. How this is built →
One of the heaviest AI adopters on earth, and one of the most collaborative in how it uses it.
Two caveats, carried openly. Rich, high-adoption countries skew toward augmentation-style usage, so part of this is who Britain's users are rather than a national virtue. And this strip is the one genuinely UK usage cut on the site: every per-occupation number elsewhere is a global usage fingerprint applied to UK employment weights. Said plainly in the method →
The same split, across the jobs most people actually do.
Kitchen, warehouse, care and admin work lean toward automation-style use. Teaching, finance and sales lean toward augmentation. Software developers sit close to the middle. The wave, when it comes, will not land evenly.
Measured: Anthropic Economic Index, real Claude.ai usage, Apr+May 2026, global, CC-BY. Confounds: Claude users only, not the whole labour market; the US O*NET-SOC to UK SOC-2020 crosswalk drops ~9% of codes; only ~45% of mapped usage is work, not personal; two-month cross-section. Employment: ONS. How this is built →