Barclays: Only 20% of Workplace Skills Are Highly Replicable by AI; Amplification Effect Covers a Much Broader Scope
nashnova research
Barclays' new "3A Framework" assesses AI's impact on 830-plus occupations at the skill level: only about 20% of workplace skills are highly replicable, but AI's amplification effect reaches nearly every corner of the labor market, and its long-term productivity gains may far exceed direct displacement.
How many jobs will AI actually replace?
Barclays' core finding: only about 20% of workplace skills can be highly replicated by AI. Fears of mass displacement are overstated.
This means → most jobs won't be "taken" by AI — they'll be reshaped. Replacement is narrow; amplification is broad.
The report's unit of analysis is skills, not occupations. In plain terms = instead of asking "will accountants disappear?", it breaks down which specific skills in accounting a machine can perform.
Which jobs face the highest exposure — and which are safest?
Highest cognitive-AI automation exposure: data scientists, statisticians, and actuaries — precisely the roles most deeply entwined with data.
Highest physical-AI exposure: agricultural workers, textile-machine operators, and structural steelworkers — roles built on repetitive manual labor.
Roughly 120 occupations show low exposure to both AI types, covering about 20% of the U.S. workforce — mainly teachers, childcare workers, and coaches. This reflects that face-to-face human interaction remains AI's hardest frontier.
Why does the "amplification effect" matter more than replacement?
Unlike automation's concentrated distribution, AI amplification spreads broadly and evenly — doctors, teachers, and engineers all stand to benefit from AI tools.
Highest cognitive amplification: air-traffic controllers, healthcare managers, and training-and-development managers. Highest physical amplification: bus and truck mechanics, industrial mechanics.
This means → AI's long-term productivity boost may exceed its direct displacement effect. In plain terms = AI is more like a universal magnifying glass than an axe that fells only a few trees.
Is the hiring market already confirming this?
Since 2022, job-vacancy share for the most automation-exposed occupations has fallen from nearly 30% to about 25%, with software engineering contributing the sharpest decline.
Vacancy share for the least-exposed occupations edged up from about 20% to 21%–23%.
Across the U.S., U.K., France, and Germany, job postings requiring AI skills have risen from roughly 2% in 2019 to 6%–10% by August 2026, with the U.K. showing the steepest climb. This means → AI skills are shifting from "programmer-only" to "everyone's baseline."
Where is AI-skill demand spreading?
In 2019, over 70% of AI-related postings sat in tech and data roles. By 2026, that share has dropped to about 50%.
Management roles' share of AI-skill demand has climbed from roughly 10% to about 25%.
This means → AI is migrating from a specialist tool inside tech departments to a day-to-day competency expected of managers. In plain terms = it's no longer just engineers who need to use AI — managers do too.
What new models will physical AI create?
Phase one: humans as "robot trainers." Physical AI lacks real-world training datasets comparable to LLM text corpora. The industry is using tele-operation and first-person video capture to generate demonstration data — Figure's Index and Scale AI are already building this crowd-sourced ecosystem.
Phase two: value shifts to the "operating procedures" themselves. Welding, surgery, and cooking expertise could be encoded into scalable, licensable "robot playbooks" — a new form of intellectual property. Unitree's UniStore and NEURA Robotics' NeuraGym / NeuraVerse are early examples.
Barclays' extreme-case projection: hardware gradually commoditizes, and the real high ground belongs to whoever can capture, encode, and control high-value skills. In plain terms = the company that builds the most robots may not win — the one that owns the "instruction layer" could dominate the new skills race.
市场有风险，内容仅供研究参考，不构成投资建议。