AI and the labour market

Our first conviction is that AI diffusion, rather than model capability, is the key variable when considering AI and unemployment. Technology diffusion takes time as it relies on broader system change, rather than just the capability of the technology itself. Companies must identify use cases, integrate systems, redesign workflows, train staff, validate outputs, manage cyber and data risks, and assign accountability when systems fail. In many cases, implementation costs still exceed the labour savings available, favouring selective augmentation over wholesale automation. History shows why these complements matter. US agricultural employment fell from 41% of the workforce in 1920 to around 5% by 1970, but tractors alone did not drive the shift: harvesting equipment, seed innovation, logistics and production methods transformed the sector over decades⁵. Digital technologies have often followed the same J-shaped productivity path^{6 7}. AI adoption is similarly ahead of organisational transformation. Gallup found in 2026 that half of employed adults used AI at least occasionally, yet only around one in ten workers in AI-adopting organisations strongly agreed that their work had been transformed⁸.
Our second observation is that higher productivity also need not translate directly into lower labour demand. This is because gains can appear as better service, new tasks, greater output and lower prices that stimulate demand — a labour-market version of the Jevons paradox. Historical examples including cotton, steel and motor vehicles show how automation can co-exist with expanding output and employment while demand continues to grow for long periods of time⁹.

Employment depends on the balance between displacement, productivity and new-task creation, not simply exposure¹⁰. Our third observation is that jobs requiring judgement, accountability and human interaction are often complementary to AI, while routine cognitive work is more substitutable. Radiology illustrates this dynamic. Despite predictions of AI-driven replacement, workforce shortages persist as imaging demand continues to outpace capacity^{11 12}. Similarly, while AI can improve coding productivity, but high error rates increase the importance of debugging, architecture, cybersecurity, and systems judgement¹³. This helps explain why current labour-market evidence remains mixed¹⁴. Amazon, Meta and HSBC have cited AI when discussing workforce changes, but reductions have been modest relative to total headcount, and Meta later acknowledged that some decisions were premature as adoption proved harder than expected¹⁵. As a result, we observe that it remains difficult to separate genuine AI displacement from cyclical weakness and conventional restructuring, particularly as technological innovation continually reshapes labour markets: roughly 60% of today’s jobs did not exist in 1940¹⁶.

AI may be software, but we observe that its bottlenecks are increasingly physical. Scaling requires compute, energy and critical materials: data-centre costs estimated to rise from $7.7m/MW in 2020 to an estimated $11.3m/MW in 2026, while copper supply is expected to meet only 70% of demand, and semiconductor supply remains concentrated in geopolitically sensitive Taiwan^{17 18}. Institutions can also add friction: Hyundai's wage negotiations became intertwined with concerns over AI-enabled robots, while Chinese courts have challenged dismissals justified solely by AI replacement. Governments could also respond to visible disruption through retraining programmes, income support or even measures such as AI-related profit taxes. Market structure matters, too. If frontier AI remains concentrated among a small number of providers with higher pricing power, adoption may be slower. A more competitive environment with cheaper models could accelerate both productivity and substitution. Our view is that widespread commoditisation of advanced AI models remains at least five years away. These constraints and interventions are likely to temper diffusion rates and help spread adjustment costs over time.

The labour debate often assumes too many workers, but many developed economies may face the opposite problem. The OECD projects the population aged between 20 to 64 to decline by 8% by 2060, while the old-age dependency ratio rises from 31% in 2023 to 52% by 2060¹⁹. At current productivity trends, annual per-capita growth could slow from 1.0% to 0.6%²⁰. Informal caregiving already costs between 0.8% to 4.9% of GDP through lower hours and productivity²¹. We would therefore observe that AI-enabled automation may therefore be necessary to sustain output and essential services as fewer workers support more dependants. Japan provides an early example: robots are increasingly used in services and elderly care, with at least 26% of long-term-care providers adopting care robots²². Alongside immigration, training and higher participation, AI productivity may alleviate labour shortages rather than create an aggregate labour surplus.

We are keen to highlight that our central case of gradual labour market disruption is not benign. Labour markets may weaken first through slower hiring, which appears to already be happening in pockets. Stanford and Census studies found a 12–16% relative employment decline among young workers in the most AI-exposed roles²³. A persistent “no-hire economy” could reduce income, mobility and demand, while fewer junior roles would limit opportunities to develop judgement, client handling, communication, and professional skills. All of this weakens the future leadership pipeline. There is also a cyclical risk. If the current AI investment boom and related equity wealth effects reverse, weaker GDP and corporate margins could encourage further automation, reinforcing recessionary unemployment. AI capability may also advance faster than organisations and policy can adapt or we expect. The Centre for AI Safety’s Remote Labour Index found that AI can currently match or exceed human performance on 16% of economically valuable freelance tasks²⁴. Separately, MIT Media Lab research links heavy AI assistance with weaker independent analysis and memory retention²⁵. AI can accelerate learning, but without appropriate governance and training it risks de-skilling workers and weakening human judgement. While unemployment considerations dominate news flow, we highlight that there is a myriad of ways AI can impact the capabilities of humans and the labour market that will need to be monitored.

AI is a core investment theme for Carmignac, sitting at the intersection of geopolitics, economics and ESG considerations. US-China competition gives it the characteristics of an arms race, sustaining investment in compute, semi-conductors and energy despite uncertain near-term returns. Combined with defence spending and supply constraints, this could keep inflation and capital costs higher and slow diffusion. We have carefully curated our AI exposure across the value chain, from Nvidia, TSMC and ASML to hyper-scalers and specialists including Nitto Boseki, Innodisk and Asia Vital²⁶. The likely winners are those that convert AI into durable productivity without weakening resilience, workforce capability or social licence. Additionally, our proprietary Responsible AI framework assesses companies across seven pillars, with particular emphasis on governance, data, and transparency, helping identify firms better placed to capture AI’s benefits while limiting operational, regulatory and societal risks.

At Carmignac we believe that AI will reshape labour markets, but adoption is likely to be gradual, uneven and contested. Some jobs will disappear, many will evolve and new roles will emerge. The widely discussed "jobs Armageddon" scenario remains unlikely in our central case because economic, institutional, physical, and demographic frictions slow diffusion. For investors, the decisive question is who can translate AI capability into commercially valuable, responsibly governed deployment at scale.