Beyond the Headlines: How to Think About AI and Jobs

The headline problem

Pick up any report on AI and the labour market and you will find a number. Thirty per cent of jobs could be automated. Eighty per cent of workers have at least some tasks exposed to large language models. Three hundred million jobs globally are at risk. The numbers vary, but the tone is consistent: something very large is about to happen.

These figures are not wrong, exactly, but they are almost always misinterpreted. Most of them measure exposure — the extent to which the tasks that make up a job could in principle be done by AI. They do not measure displacement — whether those jobs will actually disappear. The gap between the two is vast, and it is where most of the interesting questions live.

To see why, consider an analogy. When spreadsheet software arrived in the 1980s, it automated a large share of the tasks performed by bookkeepers and accounting clerks. By the exposure logic, those jobs should have been decimated. Instead, the number of accountants in the US roughly doubled over the following two decades. The technology did not eliminate the work — it changed what the work involved, lowered the cost of financial analysis, and expanded the demand for it.1

That does not mean AI will follow the same benign path. It might. It might not. The point is that moving from "AI can do these tasks" to "these jobs will disappear" requires a theory of how the whole system responds — and that is what most headline figures leave out.

Two forces: displacement and reinstatement

The clearest framework for thinking about this comes from the economists Daron Acemoglu and Pascual Restrepo, whose work on automation and labour markets has become the standard reference in the field.2 They identify two competing forces.

The first is the displacement effect. When AI automates tasks previously performed by workers, it directly reduces labour demand in the affected occupations. This is the channel captured by the headline exposure figures, and it is the one that dominates public discussion. Unlike previous waves of automation that primarily displaced routine manual and cognitive tasks, AI has the potential to automate non-routine cognitive work across a much broader occupational spectrum — which is why the current wave feels different.

The second is the reinstatement effect. Technological change has historically created new tasks and occupations that boost labour demand. Acemoglu and Restrepo identify reinstatement as the dominant force offsetting automation over the past century. The mechanism is intuitive: as machines replace humans in some tasks, the cost of those tasks falls, the relative cost of human labour in other tasks also falls, and humans adapt to provide services that are not in competition with machines. New roles emerge. The economy finds a new equilibrium.

Their model includes a third channel: the rise of the AI industry itself. The development of AI models, the infrastructure required to run them, and the consultancy services associated with advising firms on adoption all create jobs directly. This is a real effect, but in the context of the whole economy it is likely to be modest compared to the displacement and reinstatement channels.

The net impact on employment depends on the relative strength of displacement and reinstatement. Despite their finding that economies have historically self-corrected, Acemoglu and Restrepo argue that AI may generate weaker reinstatement effects than previous technologies, at least in the short to medium term — such that the economy might settle into a new balance with a smaller share of work done by humans. This is not a prediction of mass unemployment, but it is a prediction that the adjustment could be slower and more painful than in previous technological transitions.

The J-curve: why timing matters

If the Acemoglu-Restrepo framework tells us what the end state might look like, it says less about how we get there. A useful complement comes from Erik Brynjolfsson and co-authors, who propose that the economic impact of AI will follow a J-curve pattern.3

In an initial phase, there is heavy investment in AI and considerable hype about its transformative potential, but the impact on actual production is modest. Workers and firms experiment with how best to use the technology. Productivity gains are real but small, because the existing organisation of work was not designed around AI.

Over time, complementary investments in skills and tools allow workers to get more out of AI at an individual level. Then comes a deeper wave of change: firms re-engineer their organisational structures and production processes to make the most of the new technology. This is when the large productivity gains — and the large labour market disruptions — arrive.

This pattern has clear parallels with previous general-purpose technologies. Electricity was available from the 1880s, but it took decades for factories to be redesigned around electric motors rather than the steam-driven belt-and-shaft systems they replaced. The personal computer was widely adopted in the 1980s, but the productivity gains that Solow famously said were "everywhere except in the statistics" did not show up until the late 1990s, after firms had reorganised their workflows around the new technology.

Bringing these two ideas together, we might expect the impact of AI on the labour market to look something like Figure 1. In a first phase, there is excitement about AI and heavy investment in AI technology and infrastructure, but relatively modest impact on workers in other industries. As firms begin to automate existing work, certain jobs in the old economy start to be displaced — beginning with the roles comprising tasks most readily automated, but progressing to a broader set of occupations as innovation makes AI useful for a wider range of tasks and firms make the complementary investments needed to re-engineer their production processes. At this point, employment falls as the displacement effect dominates. Finally, new roles begin to emerge and partially or fully offset the lost jobs.

We are, in early 2026, most likely still in the first phase — or at the very beginning of the transition to the second. The implications for policy are significant: the fact that AI has not yet caused visible unemployment does not mean it will not do so, and waiting for the effects to become obvious before acting may be too late for the workers most affected.

From the economy to the firm — and the region

The framework above describes the impact of AI on an economy as a whole. But economies are made up of firms, and the adjustment process at the firm level is messier than the aggregate picture suggests.

Consider a particular industry — say, banking. AI may be used to replace humans in areas like customer service and compliance. The first steps could happen through experimentation with AI tools by existing staff in existing roles. Some firms may then re-engineer processes more ambitiously, creating new AI-specialist roles and reducing costs. If successful, they compete more effectively, win market share and may even expand their workforce. But it is equally possible that new entrants — built around AI from the start — emerge and rapidly take market share from incumbents, some of which then shrink or fail entirely. This is the logic of creative destruction: the aggregate displacement and reinstatement effects play out through a process of winners and losers at the firm level, not through a uniform adjustment across the industry.

This matters because it is impossible to predict, looking at any individual firm, whether it will be a fast adopter that gains market share or a slow mover that loses it. From the outside, the best one can say is that the industry-level outcome will reflect some mix of both — but the path there will involve considerable turbulence for individual firms and their workers.

At the regional level, this turbulence aggregates differently depending on the local industrial mix. An area with a workforce concentrated in logistics, construction and care faces a very different near-term challenge from one with large concentrations of administrative, financial and professional services work. The national-level framework is a useful starting point, but the specific impact on any given place depends on which industries are present, how exposed those industries' occupations are, and how quickly local firms adapt. The next article in this series explores what a place-based analysis looks like in practice and the third article in the series examines how firms themselves can adapt.

Tasks, not jobs

If the framework above tells us what the broad dynamics might look like, the research that is most useful for practical planning focuses on a more granular question: which specific occupations are most exposed, and in what way?

The key insight underpinning the modern literature is that AI does not automate jobs — it automates tasks. A job is a bundle of tasks, some of which may be highly amenable to automation and others not at all. The risk to any given job depends on the share of its tasks that can be automated, the importance of those tasks to the role, and whether the remaining tasks are valuable enough to sustain the position.

Most studies start from a database of tasks — typically O*NET in the United States or ISCO-08 internationally — and seek to assess the suitability of each task to being performed by AI. The methods vary: some survey job-holders, some consult labour market or AI experts, and some use language models themselves to evaluate task descriptions. Having scored each task, researchers aggregate over all the tasks in a job to produce an occupational exposure score.

There are several important dimensions on which these studies differ.

The first is whether they distinguish between augmentation and displacement. Some studies, such as the widely cited work by Eloundou and co-authors at OpenAI,4 measure exposure without distinguishing whether AI is likely to assist a worker or replace them. Others, notably work by Gmyrek, Berg and Bescond at the International Labour Organisation,5 explicitly score each task separately for its automation and augmentation potential. This distinction is critical: a job in which 60% of tasks can be augmented by AI looks very different from one in which 60% of tasks can be automated.

The second is the scope of AI considered. Some studies focus narrowly on large language models like ChatGPT, while others encompass a broader set of AI technologies including image and speech recognition, machine learning applied to structured data, and robotics. The broader the scope, the more occupations are affected — but at the cost of conflating technologies with very different adoption timelines.

The third is timing. The capabilities of large language models have improved dramatically since the launch of ChatGPT in late 2022. Studies conducted earlier may underestimate current capabilities, while even the most recent studies are working with a moving target.

Across the major studies, a reasonably consistent picture emerges. Clerical and administrative occupations face the highest automation risk — roles like data entry, bookkeeping, and general office administration, in which a large share of tasks involve processing structured information in predictable ways. Professional and semi-professional occupations — particularly in finance, IT and legal services — face high exposure, but a greater share of this is likely to take the form of augmentation rather than outright displacement, at least in the near term. Roles requiring physical presence, manual dexterity or direct human interaction — care work, construction, hospitality — are the least exposed.

There is a striking distributional pattern here. Unlike previous waves of automation, which disproportionately affected lower-paid manual and routine cognitive workers, AI exposure skews towards higher-educated, higher-paid, and disproportionately female workers.6 Administrative and secretarial work, for instance, is roughly 80% female in the UK. This has important implications for the design of policy responses.

A note of caution: tasks are not the whole story

The task-based approach is powerful and has become the workhorse of the field, but it has important limitations. In particular, it treats each task in a job as separable — as if automating 60% of a job's tasks straightforwardly puts 60% of the job at risk. Recent work by Gans and Goldfarb challenges this assumption.7

They argue that many jobs have an "O-ring" structure, in which the value of the whole depends critically on every component being done well. If the remaining human tasks are critical — if they require judgement, accountability or trust that cannot be delegated to a machine — then automating the other tasks does not displace the worker. It makes them more productive. The likely outcome is higher wages for the human, not unemployment.

This is a genuinely important critique, and it may well describe the near-term trajectory for many professional roles. But it does not eliminate the displacement risk — it changes its character. If AI makes individual workers dramatically more productive, firms need fewer of them. And the O-ring logic applies at the level of individual jobs, not at the level of the firm or the economy. A firm that restructures ten jobs into three, each more productive and better paid, has still displaced seven workers.

A later article in this series explores how firms might actually approach this restructuring — not as an exercise in task automation, but as a re-optimisation of how work is organised.

What the evidence says so far

If AI is going to reshape the labour market, is there any sign of it yet? The short answer is: not much in the aggregate data, but some suggestive signals in the detail.

Unemployment rates for workers in highly exposed occupations have not diverged meaningfully from those in less exposed occupations since the launch of ChatGPT. Aggregate wages and employment levels do not yet show a clear AI signature. This is consistent with the J-curve story: we are still in the experimentation phase, before the organisational restructuring that will drive the larger effects.

Look more closely, however, and the picture gets more interesting. There is suggestive evidence that hiring of younger workers into exposed occupations has slowed, even as overall employment remains stable — consistent with firms reducing headcount through attrition rather than redundancy. And the flow data from job advertisements tells a richer story than the stock data from employment surveys.8

A companion article in this series presents our own analysis of vacancy data, examining who is hiring for AI skills, in which roles, and how advertised wages in exposed occupations are evolving.

What this means for policy — and practice

The combination of the Acemoglu-Restrepo framework, the J-curve thesis, and the task-based evidence points to a challenging but navigable landscape. The risks are real but not yet realised. The window for proactive planning is open but will not stay open indefinitely.

For regional and local government, the most valuable step is to understand the specific occupational profile of your area. The same technology will hit different places very differently, depending on the mix of industries and occupations, the demographic profile of exposed workers, and the availability of retraining pathways. Generic national-level exposure figures are a poor guide to local policy. The next article in this series explores what a place-based analysis looks like in practice.

For firms, the question is not whether to adopt AI but how to reorganise around it. The companies that treat AI as a tool to bolt on to existing processes will capture a fraction of its value. Those that rethink how work is structured — which tasks belong together, which can be delegated to machines, and which new capabilities become possible — will capture far more. A later article sets out a framework for approaching this.