AI's effect on the labor market may be more subdued than market narratives suggest, yet its reach runs deeper. New Barclays research indicates that only about 20% of core workplace skills can be highly replicated by AI, casting doubt on fears of a massive "replacement wave." However, the "amplification effect" of AI on productivity paints a starkly different picture—its coverage is far broader than automation-driven substitution, touching nearly every corner of the labor market.

On September 22, Barclays released a report introducing the "Barclays AI Automation and Amplification Framework" (3A Framework), which uses skills rather than occupations as the core analytical unit to systematically assess the impact of cognitive AI and physical AI across more than 830 occupations. The study found that automation risk is highly concentrated in a small set of occupational groups. Cognitive AI's high-exposure roles are mainly clustered in computer, engineering, and scientific fields, while physical AI's high-exposure roles are centered in construction, production, installation, maintenance, and repair. Meanwhile, U.S. job posting data already reflects structural shifts consistent with AI—since 2022, the share of postings for occupations with the highest automation exposure has retreated from about 30% to roughly 25%.

The rapid expansion of AI skill demand reveals another layer of dynamics: in the United States, United Kingdom, France, and Germany, the share of job postings requiring AI skills has climbed from about 2% in 2019 to between 6% and 10% by August 2026. Notably, the center of gravity for AI skill demand is shifting—in 2019, over 70% of AI-related roles were concentrated in tech and data positions, but by 2026 that share has fallen to about 50%, while management roles have risen from about 10% to 25%, indicating that AI is spreading from specialized technical domains into the broader workplace ecosystem.

What AI Automates or Amplifies Are Specific Skills, Not Entire Occupations

Traditional analyses of AI's labor impact typically use "occupations" as the unit of analysis. Barclays argues that occupations are collections of tasks, tasks are collections of skills, and skills are the "atomic unit" for understanding AI's influence. AI is more likely to automate or amplify specific skills—such as information gathering, writing, or data analysis—rather than entire occupations themselves.

The 3A Framework is built on the U.S. Department of Labor database, which covers 104 cross-occupational skills mapped to over 800 occupations. The framework's innovation spans three dimensions: first, it incorporates both cognitive AI and physical AI, covering both digital and physical intelligence forms; second, it replaces tasks or occupations with skills as the fundamental unit of analysis; third, it captures both "automation" and "amplification" effects—the former measuring the degree to which AI substitutes for human labor, and the latter measuring AI's potential to enhance the productivity of human skills.

The results show that only about 20% of workplace skills are rated as highly replicable by AI. About 40% of skills show strong resistance to cognitive AI, while the proportion of skills with strong resistance to physical AI is as high as 70%, indicating that physical AI's current automation footprint remains far narrower than that of cognitive AI.

Automation Exposure Is Highly Concentrated, Leaving Most Workers Relatively Safe

On automation exposure distribution, Barclays' key finding is that extreme levels of exposure are confined to a very small number of occupations, and most occupations lean primarily toward exposure to either cognitive AI or physical AI, rather than both simultaneously.

The occupations with the highest cognitive AI automation exposure include data scientists, statisticians, and actuaries, whose core skills—information retrieval, pattern recognition, document processing, and procedural decision-making—overlap heavily with AI capabilities. The highest physical AI exposure is found in agricultural workers, textile machine operators, and structural steel workers, who rely on repetitive manual labor.

By contrast, about 120 occupations show relatively low combined exposure to both types of AI automation, covering roughly 20% of the U.S. workforce. These mainly include teachers, childcare providers, coaches, and some hospitality and entertainment workers—roles that depend heavily on interpersonal interaction, on-site supervision, or live performance, making them difficult for AI to replicate.

Amplification Effect: Broader Reach, Potentially Larger Gains

Unlike the concentrated distribution of automation, AI's amplification effect is far more widespread and uniform. In Barclays' model, the amplification effect is most pronounced in occupations with the most balanced mix of replicable, partially replicable, and non-replicable skills—where AI handles the replicable components while boosting the value and output efficiency of remaining human skills.

The occupations with the highest cognitive AI amplification include air traffic controllers, healthcare services managers, and training and development managers. In these roles, AI can take over administrative documentation and information processing tasks while amplifying core human capabilities such as leadership, coaching, and personnel management. Physical AI amplification is most prominent across various mechanical and technical roles, such as bus and truck mechanics and industrial machinery mechanics, where some routine physical tasks can be automated while diagnostic, decision-making, and adaptive capabilities remain human-led.

Barclays points out that the core message of the amplification distribution is this: AI's automation capabilities are concentrated in specific occupations, but its productivity-enhancing capabilities cover an extremely broad range of occupational groups. Doctors, teachers, and engineers—vastly different professions—can all benefit substantially from AI tools. This characteristic suggests that AI's overall productivity gains may exceed its direct substitution effects over the long term.

Job Posting Data Confirms Structural Shifts, with AI Skill Demand Spreading to Management

Labor market data has begun to corroborate the framework. According to Barclays' analysis of LinkUp job posting data, since 2022, the share of vacancies for occupations with the highest automation exposure has declined from nearly 30% to about 25%, with software engineering roles contributing the most to the drop. The share for occupations with the lowest automation exposure has risen slightly from about 20% to between 21% and 23%.

Meanwhile, demand for specialized AI skills is accelerating. Based on data from Indeed Hiring Lab and other sources using broader keyword searches, the share of AI-related job postings in the U.S., U.K., France, and Germany has risen from about 2% in 2019 to between 6% and 10% by 2026, with the U.K. showing the most significant increase.

More structurally significant is the shift in demand sources. In 2019, tech and data roles dominated over 70% of AI-related job postings; by 2026, that share has fallen to about 50%, while management roles have risen from about 10% to about 25%. This trend is observed across the U.S., U.K., France, and Germany, indicating that AI is permeating the broader workplace as a general-purpose skill rather than a specialized technology.

The Rise of Physical AI: Skills as Licensable Intellectual Property

The report further looks ahead to new economic models enabled by the proliferation of physical AI, dividing them into two phases.

In the first phase, humans act as "robot trainers." Since physical AI lacks real-world datasets equivalent to the training corpora used for large language models, the industry is relying on human demonstrations—through teleoperation, first-person video capture, and similar methods—to generate training data for robots. Companies like Figure AI and Scale AI are already building this crowdsourcing ecosystem.

In the second phase, the center of value shifts toward "operation protocols" themselves. As physical AI matures, the focus will move from hardware to the embedded skills and workflows—welding, surgery, cooking, and other human expertise could be encoded into scalable, licensable "robot operation manuals," creating a new form of intellectual property. Unitree's UniStore and NEURA Robotics' NeuraGym/NeuraVerse are already seen as early examples of such skill marketplaces, though they currently remain hosted on humanoid robot hardware vendor platforms.

Barclays argues that in the extreme scenario of this trend, hardware will gradually become commoditized, and the true competitive high ground will belong to those who can capture, encode, and control high-value skills. The company that builds the most robots may not be the ultimate winner—the one that controls the "instruction layer" could dominate this new battle for skills.