Part One of this series highlighted that, while AI-related corporate layoff announcements suggest a major disruption to the labour market, aggregate statistics remain muted. The evidence, at a macro level, points to a manageable transition.
At a micro level, however, the picture is more troubling. It shows a consistent pattern across different datasets: slower hiring for young workers in AI-exposed occupations, particularly concentrated in higher-wage, analytical and back-office roles, with effects becoming clearer in 2024 and 2025.
The history of technology-driven labour displacement suggests that aggregate employment eventually recovers, but that the workers caught in the transition bear disproportionate costs. The new roles that emerge may reward different people, in different industries, with different skills.
Part One also established a key distinction between AI that substitutes for human labour and specific skills and AI that augments them. Most current deployments are still in the augmentation phase, which helps explain why the aggregate labour market data looks stable.
Companies are also only in the early stages of technology investment and adoption, so any dramatic effects in terms of job replacement or productivity gains are unlikely to be seen.
Previous technological transformations indicate that the pendulum will ultimately swing from augmentation toward substitution. To predict how the transition will play out over the decade ahead, critics and proponents of AI point to several, sometimes competing, theories, which could have significant implications for workers and firms.
Jevons paradox: efficiency breeds demand
The optimistic case for AI and employment rests on Jevons Paradox. This is the 19th-century observation that as technology makes a service more efficient and cheaper, demand for that service tends to expand. The most cited example is the ATM, which did not eliminate bank tellers but instead lowered the cost of opening branches, prompting banks to open more branches and ultimately employ more tellers in a different mix of tasks.
Applied to financial services, the argument runs that if AI allows a law firm, bank or consulting firm to deliver services at a fraction of the current cost, the resulting surge in new clients, including the large reservoir of people and small businesses currently priced out of professional services, may require more human employees, not fewer.
But Jevons Paradox requires time. Time for markets to recognise new demand, time for workers to retrain and for employers to expand rather than simply contract. The ATM analogy is instructive because it played out over decades. Even then, teller employment eventually fell sharply as branch activity continued to shift. Critics also argue that the ATM comparison may underestimate AI’s potential impact because AI affects a much wider range of cognitive tasks than ATMs affected banking operations.
Amdahl’s Law: humans are the binding constraint
The second theory often cited regarding how AI implementation could unfold is Amdahl’s Law, originally formulated by computer scientist Gene Amdahl in 1967 for parallel computing. Its core insight is that when one accelerates certain components of a process, the unaccelerated components become the bottleneck that determines the overall system speed.
Applied to AI and work, this implies that as AI automates more of a task, the remaining human contribution becomes more valuable. If AI handles 95% of a financial analysis, the 5% that requires human judgement, contextual understanding and client relationships becomes the binding constraint on which everything else depends. Humans are no longer doing the volume work; they are doing the work that makes the whole output useful.
At his May 2026 briefing on financial services alongside JPMorgan Chase CEO Jamie Dimon, Anthropic CEO Dario Amodei put it this way: “If you automate 90% of the job, then everyone does the 10% of the job, and the 10% kind of expands to be 100% of what people do and kind of 10xs their productivity.”
There is, however, a critical qualification that Anthropic’s own research surfaces. The January 2026 Anthropic Economic Index found that in some occupations, AI may remove more cognitively demanding components of work, potentially leaving workers focused on lower-skill tasks, although the effect varies considerably across occupations.
The report gave two contrasting examples from adjacent financial roles: travel agents would experience deskilling as AI absorbs complex itinerary planning, leaving behind routine ticket purchasing and payment processing. Property managers, by contrast, would undergo upskilling as AI takes over bookkeeping, leaving them focused on contract negotiations and stakeholder management.
Amdahl’s amplification effect, where the residual human contribution becomes more valuable, only materialises when AI leaves behind the high-judgement, high-skill components. For financial services, the evidence (see Part One) suggests that back-office, codified, data-intensive tasks go first, and that this is also the work through which junior professionals build the tacit knowledge they will need as seniors.
Is this time different?
Like the Jevons Paradox, Amdahl’s Law requires time for rebalancing to take effect and for workers to adapt. A central argument for why this transition may differ from previous technological revolutions is speed. It is often claimed that AI will not unfold over decades.
Amodei acknowledged this at the same briefing: “AI is moving faster than all these previous technologies. And so when you strain a system more than it’s usually strained, it’s possible you get these weird behaviours and this big disruption.”
Standard economics dismisses fears of permanent job loss by invoking the lump-of-labour fallacy. In other words, labour is not finite, and more productive economies have historically employed more people, not fewer, as new demand absorbs displaced supply. But this reabsorption mechanism requires labour to flow from shrinking sectors into expanding ones.
In addition to the time required, the other question raised by economists is what happens when a technology is deployed simultaneously across all sectors, compressing the adjustment period beyond what labour market institutions can handle.
Harvard University’s Daron Acemoglu and Pascual Restrepo developed a reinstatement framework that distinguishes between tasks that AI automates, reducing demand for labour, and entirely new tasks reinstatedinto the process, creating fresh demand. Historically, each wave of automation was accompanied by sufficient task reinstatement to sustain employment. Their concern, supported by evidence that the US economy’s rate of new task creation had already slowed before the generative AI era, is that the current wave may be automating faster than new tasks are being created.
The firm-level incentive problem
Even where augmentation would produce greater total value than substitution, such as a human-AI team outperforming the AI alone, firms face a structural incentive to pursue substitution. Labour is treated as an operating expense, whereas technology investment is treated as capital, with favourable tax treatment and positive signals to equity markets.
In a June New York Times article, Erik Brynjolfsson, the director of Stanford University’s Digital Economy Lab, noted, “A lot of people are under the mistaken idea that the only way that you get productivity from AI is by removing labour costs.” Brynjolfsson pointed to the tax code as a direct driver. Firms that invest in capital pay less tax than those that invest in people.
This was partly reflected in a statement by Standard Chartered CEO Bill Winters in May, announcing plans to cut 7,800 back-office and support roles at the bank by 2030 to integrate AI and automation. Winters explained that this wasn’t about cutting costs, but “replacing, in some cases, lower value, human capital, with the financial capital and the investment capital that we’re putting in”.
The statement sparked widespread backlash over its language, prompting Winters to apologise and walk back the comment, describing it as taken out of context. Curiously, the underlying message was somewhat lost in the outrage over its tone.
The true cost of AI may not yet be priced in
On the flip side, a limiting factor in AI’s impact is equally underappreciated: its cost.
Current AI pricing may reflect a land-grab phase rather than sustainable economics. Token prices for cheaper, less powerful economy-tier AI inference have fallen roughly 600-fold since 2020, faster than Moore’s Law. This has been driven primarily by competition for market share since mid-2024 rather than by fundamental reductions in the cost of building and running the underlying models.
Meanwhile, training costs for frontier models have risen from around $1,000 in 2017 to nearly $200 million today, and the infrastructure required to support them is even more expensive. The major technology companies are collectively committing hundreds of billions of dollars to data centre construction. The gap between what firms currently pay for AI services and what providers will eventually need to charge to recover those investments and earn a return is substantial.
This matters for how we interpret the productivity evidence. Many studies measuring AI’s impact assess gross output improvements, even though AI services are priced well below their long-run economic cost. Once inference costs, infrastructure investment and provider profit requirements are fully reflected in market pricing, a portion of the apparent productivity gains could be transferred from AI users to AI providers, reducing the net benefit to firms and to the broader economy. Workforce restructuring decisions made on the basis of today’s subsidised economics may appear poorly calibrated when customers pay the true price.
There is a further possibility that tends to get lost in a debate polarised between techno-optimism and techno-pessimism: AI can create genuine and measurable value while still disappointing investors and policymakers expecting a transformational productivity revolution. A technology that makes financial analysts 15% more productive, reduces compliance processing times, and improves customer service response rates is genuinely useful. It is not, on current evidence, the kind of step-change in output that would justify the scale of capital committed to it or the pace of workforce restructuring planned for it.
The economic benefits may simply be arriving more slowly than the investment and cost structure would imply, which would be entirely consistent with how previous transformative technologies have behaved. AI will very likely deliver net productivity gains, but the size, timing and distribution of those gains remain uncertain, and the workforce decisions being made today are running well ahead of that uncertainty.
Most economists agree that artificial intelligence will create and destroy jobs, and with so many variables, it remains unpredictable which effect will dominate.
Three scenarios for the decade ahead
Nevertheless, with those competing principles in mind, we can consider three scenarios to describe the probable range of outcomes for financial services employment over the next decade. They are not mutually exclusive as elements of each may emerge simultaneously in different parts of the sector or at different times.
Scenario one: Gradual rebalancing
In this scenario, the augmentation phase continues to dominate through the late 2020s. Hiring for juniors in high-exposure roles slows, but banks and insurers manage transitions through attrition and hiring freezes rather than dramatic redundancies. Overall employment statistics remain broadly benign, consistent with Goldman Sachs’s projection of a 0.5 percentage point rise in unemployment during the transition.
Jevons Paradox operates as the optimists predict. AI lowers the cost of financial advice, new demand emerges from previously underserved clients, and firms expand their service offerings to absorb the efficiency gains. Amdahl’s Law also operates benignly. The judgment-intensive work left to humans is genuinely high-value, and experienced professionals become more productive rather than redundant.
The distributional harm falls primarily on recent graduates and career entrants who find the traditional progression path increasingly inaccessible. This will become visible in hiring data, but below the threshold that triggers institutional responses. The central risk in this scenario is complacency, as the missing-generation problem develops slowly and only becomes apparent in the data when it is already difficult to reverse.
Scenario two: Accelerated structural shift
In this scenario, the gap between observed and theoretical AI exposure narrows faster than current adoption rates suggest. The Anthropic Observed Exposure data show that Claude is currently used for 33% of Computer and Mathematical tasks, even though 94% of these tasks are theoretically feasible. If competitive pressure drives adoption toward that ceiling within five to seven years, the dynamics can change significantly.
Here, Jevons Paradox begins to fail because the speed of transition outpaces the rebalancing mechanism. New demand for financial services does emerge, but the retraining infrastructure, institutional know-how, and new role definitions needed to channel displaced workers to meet that demand do not materialise quickly enough. Firms contract before they expand, and the period between the cost savings and the demand-driven rehiring is too long for the workers caught in the middle.
Amdahl’s deskilling risk becomes acute in this scenario. As AI absorbs the higher-skill components of junior roles at speed, what remains is not the judgment-intensive bottleneck that Amdahl’s Law promises will be amplified. Firms that have shed their junior intake discover a decade later that they cannot develop the next generation of experienced professionals internally. The traditional pyramid-shaped workforce is compressing, and the missing-generation risk materialises as a structural feature rather than a temporary disruption.
Scenario three: New task reinstatement
The economic optimist case rests on what Acemoglu and Restrepo call the reinstatement of new tasks: AI produces entirely new forms of value creation that require human workers in ways not yet predictable. The telephone operators’ successors are today’s data scientists; the displaced financial analysts’ successors may be professionals who combine domain expertise in finance with the ability to direct, audit, evaluate and contextualise AI-generated output.
In this scenario, Amdahl’s Law operates as its most optimistic proponents suggest. The roles that emerge are explicitly organised around the human judgement bottleneck. AI system supervisors, model risk auditors, AI-output interpreters for clients, and ethical oversight professionals whose tacit knowledge and contextual understanding cannot be automated.
It is worth noting that approximately 60% of workers in the US economy today hold occupations that did not exist in 1940. The important caveat is that those transitions took decades, and were accompanied by substantial expansion of public education, social insurance and training infrastructure. Jevons Paradox is a long-run equilibrium claim. Whether the financial services sector and the wider economy are positioned to manage an analogous transition within a compressed timeframe, and whether workers displaced in the near term can truly benefit from the new equilibrium in the long run, are questions that any policy conclusions must address.
Policy and regulatory questions may also insulate certain sectors more than others. Financial services, for instance, may prove more resistant to full automation than other sectors because legal liability, regulatory oversight and client trust create demand for accountable human decision-makers even where AI can technically perform the underlying analysis.
