In March 2026, Bloomberg reported that HSBC was weighing job cuts that could affect as many as 20,000 positions, roughly one in ten of the banking group’s global workforce, as chief executive Georges Elhedery placed a bet on artificial intelligence to shrink the bank’s middle and back offices.
HSBC is not alone. DBS, Southeast Asia’s largest bank, said in February 2025 that it expects to cut roughly 4,000 positions over the next three years as AI takes over key tasks. The bank plans to lose the roles through attrition and create 1,000 new AI-enabled positions. But outgoing chief executive Piyush Gupta told a banking conference last year: “In my 15 years of being a CEO, for the first time I’m struggling to create jobs. I’ve always had a line of sight to what jobs I can create. This time, I’m struggling to say how I will repurpose people to create jobs.”
Bloomberg Intelligence has projected that global banks could eliminate as many as 200,000 positions in the next three to five years as AI encroaches on tasks currently performed by humans.
Similar job cuts are anticipated at tech companies. Cryptocurrency exchange Coinbase announced plans in April to lay off 700 of its 5,000 employees. CEO Brian Armstrong said, “AI is bringing a profound shift in how companies operate, and we’re reshaping Coinbase to lead in this new era.”
The largest AI companies, meanwhile, are incentivised to tout the disruptive effect of their models to attract funding and business customers. Anthropic CEO Dario Amodei spent much of last year arguing that AI could eliminate half of entry-level, white-collar knowledge work within years.
While announcements of AI-related restructuring are accelerating, the academic evidence on AI’s actual labour market effects remains, as of mid-2026, more ambiguous than the headlines suggest.
The academic evidence of AI disruption
The Budget Lab at Yale found that the pace of occupational change since the generative AI era began is not dramatically faster than during previous technological transitions, such as the PC era after 1984 or the internet era after 1996, and that no discernible economy-wide employment disruption attributable to AI is visible in the data as of late 2025.
A meta-analysis of empirical studies by Carbonara and Santarelli at the University of Bologna found no statistically significant average effects of AI adoption on employment or wages.
Goldman Sachs Research projects that AI could ultimately displace 6%–7% of the US workforce if widely adopted, but estimates that this impact will be temporary, with unemployment rising by only about half a percentage point during the transition. This would be consistent with frictional unemployment seen in previous periods of rapid technological change.
One reason the aggregate data appears stable is that AI does not entirely replace jobs. It replaces specific tasks within jobs. Automation-oriented AI most visibly substitutes for human labour in back-office functions such as reconciliations, document processing and routine compliance checks.
Complementary AI, in turn, augments human capabilities, allowing analysts and advisers to delegate data retrieval and first-draft work while focusing on advisory, judgment and client relationships.
For most professional roles, it appears we are still predominantly in an augmentation phase, in which AI serves as a productivity tool rather than a replacement. The labour market disruption seen in previous waves of automation tends to arrive later, when substitution rather than augmentation becomes the dominant mode of deployment.
The macro picture is thus stable overall, but the average does not allow us to draw conclusions about how any potential disruption is distributed across the labour market.
The observed versus the theoretical exposure gap
A study published by Anthropic economists in March 2026 constructs an “Observed Exposure” metric that combines occupational task data, theoretical feasibility scores and real usage data from Claude.
The results show that AI disruption is not arriving quickly and uniformly. Despite 94% of tasks in Computer and Mathematical occupations being theoretically feasible for AI, Claude is currently used for only 33% of those tasks.
In financial services, financial and investment analysts score an observed exposure of 57.2%, which is high by any measure, but substantially below the theoretical ceiling.
This gap reflects the friction of organisational change, including workflow redesign, training requirements, liability concerns, regulatory caution and the difficulty of codifying tacit knowledge. For now, observed exposure scores indicate a more measured pace of disruption than the dramatic corporate announcements suggest.
The productivity J-curve: Too early for measurable productivity gains
A related question is why the enormous investment in AI tools has not yet produced a visible productivity boom, if AI mainly augments human labour. The answer is likely the so-called productivity J-curve. When a powerful new general-purpose technology is introduced, measured productivity initially stalls because firms are in a costly phase of intangible investment. They have to reorganise workflows, redesign processes and re-skill staff.
Major banks are already logging millions of learning hours and creating internal AI academies. These investments, which are invisible to traditional productivity statistics, are the necessary groundwork for the eventual upswing. The J-curve suggests the productivity gains from AI may still lie ahead, arriving only after the current period of organisational adjustment has run its course.
The accounting profession is an indicator of this. The Big Four accounting firms have collectively committed over $6 billion to AI transformation. A Stanford-MIT study found that accountants who use generative AI finalised monthly statements 7.5 days faster, reallocating 8.5% of their time from data entry to business communication and quality assurance, and achieving a 12% increase in reporting granularity.
Yet, internal survey evidence from the Big Four tells a more nuanced story: 61% of employees already use generative AI tools, often without their managers’ knowledge, but only 12% are daily users, and AI is deployed primarily for communication tasks rather than core accounting functions.
Financial services: disproportionately exposed
Within that more measured picture, financial services remain among the most exposed sectors. The World Economic Forum estimates that 50%–70% of tasks across the sector will be automated or augmented, with insurance underwriters facing potential augmentation across 100% of their tasks. Bank tellers, loan interviewers, personal financial advisers and claims processors all feature prominently in sector-level exposure analyses.
Research into occupational automation risk provides a picture of where the pressure is most acute within financial services. Frey and Osborne’s foundational 2013 Oxford study assigned automation probability scores to 702 occupations, resulting in a wide spread within financial services: insurance underwriters scored a 99% probability of automation and loan officers and credit analysts 98%, while personal financial advisers scored 58% and financial managers just 7%. The risk gradient runs from near-certain to negligible depending on how codified and routine the core tasks are.
More recent work has refined this beyond a single risk score. IMF researchers Cazzaniga and colleagues propose a typology distinguishing between high-exposure, low-complementarity occupations, which are at genuine risk of replacement, and high-exposure, high-complementarity occupations that are more likely to be augmented. Applied to financial services, back-office processing roles and rule-based tasks such as account reconciliations and KYC document processing fall into the first category, whereas relationship-intensive and judgment-heavy roles fall into the second.
Analysis of actual AI usage reinforces this: current deployment is oriented more toward augmentation (57%) than automation (43%), and remains lower in both the lowest- and highest-paid roles, reflecting the practical barriers to deployment at both ends of the skill distribution.
Where the pain is falling: the entry-level problem
However, the placid macro picture masks a more troubling distributional pattern. Brynjolfsson, Chandar and Chen’s “Canaries in the Coal Mine” paper, published in November 2025, using ADP payroll data covering millions of workers, found that early-career workers aged 22 to 25 in the most AI-exposed occupations experienced a 16% relative decline in employment compared to their less-exposed peers. Workers aged 26 and above showed no such decline; experienced workers in AI-exposed roles saw employment grow by 6%–9%.
Complementary studies reach consistent conclusions from different data. Klein Teeselink’s King’s College London study found that AI-exposed firms reduced total employment by 4.5% on average, with the effect concentrated in junior positions, which fell by 5.8%. Highly exposed firms became 16.3 percentage points less likely to post new vacancies, with a 23.4% drop in job postings and a 9.6% fall in employment in high-paying firms, compared to almost no change in low-paying firms.
Hosseini and Lichtinger at Harvard, using resume data covering 62 million workers, find the same pattern. Following AI adoption, junior employment declines sharply while senior employment remains stable or grows, driven by slower hiring rather than increased layoffs.
The Economist published its own analysis in May 2026, which adds a fourth independent dataset to this picture. Using 10 years of graduate employment surveys from the National Association of Colleges and Employers, which track whether new alumni are working, unemployed or in graduate school six months after graduation, the paper compared outcomes across fields with varying levels of AI exposure before and after the arrival of large language models.
Graduates in the least-exposed quintile, studying subjects such as education, philosophy and civil engineering, saw their full-time employment rate fall by just 1.5 percentage points between 2022 and 2024. Those in the most exposed quintile, computer science, computer engineering and information science, suffered a 6.6 percentage-point drop over the same period.
Updated figures from 13 universities extended the trend to the class of 2025. Full-time employment in the most exposed fields fell from nearly 70% to 55% over the three years following ChatGPT’s release, after remaining stable in the years prior.
The vulnerability of junior staff lies in the nature of the labour they provide. Young workers primarily supply codified knowledge in the form of routine, standardised tasks that are easily replicated by AI.
Veterans are insulated by tacit knowledge that applies institutional context and high-level judgement accumulated through years of experience, which has never been digitised.
The potential erosion of training grounds creates a critical long-term risk. Junior roles have traditionally been the pathway through which workers master progressively complex tasks. As AI absorbs more of these routine functions, the foundational experiences that once developed senior talent are diminishing.
What history tells us
Feigenbaum and Gross’s study of AT&T’s mechanisation of telephone exchange operations between 1920 and 1940 offers an instructive parallel. About 80% of telephone operator jobs were eliminated. Those in post when the wave hit were significantly less likely to be employed ten years later, and those who found work tended to be in lower-paying roles. Their skills had been partially devalued.
The finding on subsequent cohorts is more reassuring but also more complex. Employment rates for women who would have become operators were not systematically lower. Instead, they moved into other fields as new demand emerged. But the process took years and was painful for those in its path.
The analogy has clear limits. Financial analysts are not early-twentieth-century switchboard operators, and large language models have a different adoption curve than mechanical relay systems. But the historical evidence tempers both extremes. Aggregate employment will probably not collapse. But the workers most directly in the path of displacement will bear disproportionate costs. And the new roles that emerge may reward different people, in different industries, with different skills.
Part Two of this series will outline three scenarios for AI implementation and disruption, and examine what firms, regulators, and workers should do in response to the reskilling gap and training bottlenecks.
Key sources
Brynjolfsson, E., Chandar, B. & Chen, R. (2025). Canaries in the Coal Mine? Six Facts about the Recent Employment Effects of Artificial Intelligence. Stanford Digital Economy Lab, November 2025.
Hosseini, S.M. & Lichtinger, G. (2025). Generative AI as Seniority-Biased Technological Change: Evidence from US Resume and Job Posting Data. SSRN Working Paper, October 2025.
Klein Teeselink, B. (2025). AI Exposure and Labour Market Outcomes. King’s College London, October 2025.
Feigenbaum, J. & Gross, D. (2024). Answering the Call of Automation: How the Labor Market Adjusted to the Mechanisation of Telephone Operation. American Economic Review.
Frey, C.B. & Osborne, M.A. (2013). The Future of Employment: How Susceptible Are Jobs to Computerisation? Oxford Martin Programme on the Impacts of Future Technology.
Cazzaniga, M. et al. (2024). Gen-AI: Artificial Intelligence and the Future of Work. IMF Staff Discussion Note.
Eloundou, T. et al. (2023). GPTs are GPTs: An Early Look at the Labor Market Impact Potential of Large Language Models. OpenAI.
Carbonara, E. & Santarelli, E. (2025). The Labor Market Impact of Generative AI: A Critical Survey of the Empirical Literature. University of Bologna.
Goldman Sachs Research (2025). AI and the Labour Market: A Transitional Disruption. August 2025.
Budget Lab, Yale University (2025). Occupational Change in the Generative AI Era. Late 2025.
Anthropic Economic Index (2026). Observed Exposure to AI in the US Workforce. March 2026.
World Economic Forum (2025). Future of Jobs Report 2025.
Acemoglu, D. & Restrepo, P. (2018). The Race Between Man and Machine. American Economic Review.
The Economist, Is AI putting graduates out of work already?, 13 May 2026
