Are we worrying about the wrong workers?
Economic activity in the UK is increasingly disembodied from the physical world – and AI isn’t the only culprit
This is Post Haste, our home for swift thoughts, responsive takes, and first attempts at solving problems. Want to write for us? Email editor@britishprogress.org
By Pedro Serôdio (@pdmsero)
Since ChatGPT arrived on the scene, employment in the UK occupations most exposed to generative AI has defied expectations and grown.
Our last report on the impact of AI on the labour market highlighted that the technology’s impact couldn’t yet be seen on aggregate employment figures in the UK. Those jobs supposedly existentially threatened by the new frontier ended up increasing by between 5% and 10%, depending on how we measure exposure. On the other hand, employment in the least exposed occupations has fallen by between 2% and 4%.
Those less exposed occupations include the likes of stonemasons, bricklayers, scaffolders, groundworkers, street cleaners, or refuse collectors. These should be the AI-proof roles of the future. So why are they losing out?
Following last week’s release of updated figures for the ONS’s Annual Population Survey, covering April 2025 to March 2026, we reran the UK labour-market pipeline behind our April report. The conclusions haven’t changed. AI-exposed occupations have added a small number of jobs, while fewer people work in the kinds of jobs that should hold out longer against automation. But this analysis implicitly accepts the framing that “laptop class” jobs are relatively straightforward to replace with AI. As I’ve argued before, that requires very extreme assumptions on the substitutability of tasks, as well as what role jobs have within organisations.
The policy debate about AI and employment often focuses on how we can protect knowledge workers from displacement and manage that transition if it arrives. Those concerns may yet be vindicated, and we have very few satisfying answers ready if they are.
But the past four years of UK data describe a problem that runs counter to prevailing fears. The jobs that should be safest from automation are the ones disappearing. To my mind, the reasons this is happening have very little to do with AI, and far more to do with expensive barriers to real-world activity.
What the latest data shows
In our analysis of the fresh figures, an occupation at the 90th percentile of AI exposure grew between 1.7% and 6.1% faster after ChatGPT’s launch than one at the 10th percentile.
Our analysis also indicates that highly exposed occupations – coders, analysts, brokers – were also growing faster before ChatGPT was introduced. This is consistent with a deeper transformation of the British economy. Exposure to AI is also highly correlated with suitability for remote work.
For those reasons, we cannot reliably conclude that the roll-out of the technology has had a significant impact on job creation.
Software development has been fundamentally transformed by the emergence of this technology and the attendant leaps in AI coding. A reasonable assumption therefore might have been that the sector would be ground zero for job losses. But this hasn’t held up.
Against more gloomy expectations, employment in programmers and software development professionals remains above its 2004 to 2019 trend, though it sits below the peak reached just after ChatGPT launched.
Some of this could have been driven by an increase in the precarity of employment, with companies hiring workers to assist with the transition to full automation. That does not seem to have been the case.
If we decompose overall employment, we can see that part-time work remains between 3.6% and 5.2% of coder employment throughout, around 26,000 jobs in the latest year. The decline from the peak in 2023 was almost entirely driven by full-time employment, and so has the recovery since: full-time coder employment rose 18,000 over the latest year while part-time roles changed by an imperceptible amount.
Meanwhile, cyber security roles have roughly tripled since early 2021, IT business analysts are up about half, and web designers are now below their 2021 level. More than wholesale replacement, software-adjacent work is being repackaged across job titles to a much greater extent than it is shrinking, which is what a model of imperfect substitution with some automation would predict.
This is broadly consistent with most users’ everyday experiences of using AI. Building websites and simple tools is now possible at an unprecedented scale, but complex work and secure environments still require highly specialised skillsets which drive increased demand for labour in these occupations. Recent events suggest that this should remain a fast growing area given the unprecedented threat levels that more capable autonomous systems increasingly pose.
Are productivity effects equally invisible?
If the predicted labour market effects of AI have yet to materialise in large scale transformation of the workforce, at least in the UK, does this mean that there aren’t any measurable impacts? And given that some forecasts projected productivity improvements that were based on labour replacement, what happens to productivity if jobs aren’t displaced?
According to recent research, the firms that grow headcount fastest are also the ones adopting AI most intensively. Researchers at the payments firm Ramp and the workforce-data firm Revelio Labs matched AI spending at 21,600 US firms to workforce records and found that the heaviest adopters grew headcount by about 10% against firms that had not yet adopted.
As for productivity, a study of roughly 12,000 European firms finds that adopting AI raises a firm’s labour productivity by about 4%. The gain came from firms investing in AI as capital rather than from cutting staff, a mechanism economists call capital deepening. The paper finds that overall employment holds up, adopters have higher wages on average, and that gains are concentrated in larger firms.
A separate study estimated that AI’s realised contribution to output would be roughly $878 billion a year in the United States against $138 billion across twelve EU economies. This impact is concentrated in the sectors most exposed to AI – and these findings show, contrary to alternative evidence, that there was some decline in employment.
Indeed, some of what we know so far indicates that large, visible productivity effects are unlikely to materialise without job losses. An IMF study from last year predicts only about a 1.1% cumulative gain in European total factor productivity –a measure of how efficiently inputs are turned into outputs – over the medium term.
This falls below the projections made by consultants a few years ago, which predicted a rise in output that was predicated on a drop in employment across the most AI-exposed sectors. Investors, too, are banking on bigger effects. US share prices have already baked in a permanent 30.5% productivity gain for software engineers since late 2022.
Evidence for the impact on jobs is therefore mixed and does not yet suggest the dramatic effects expected in some quarters, while productivity has been moderately boosted If we’re becoming a bit more productive but the expected decline in exposed jobs hasn’t yet materialised, does that mean we have been worrying about the wrong activities?
The jobs AI cannot yet do, and humans cannot find
If employment for coders and analysts continues to grow, policy attention should shift back towards bricklayers and scaffolders. For these types of jobs, there has been a longer-term transformation, accelerated by increasing costs in carrying out economic activity in the physical world.
Employment in the least AI-exposed occupations held roughly level through 2024, and then resumed its long-term decline. By the year to March 2026 it is 2% to 4% below its 2021 level, depending on the exposure measure, with the decline concentrated in the last eighteen months.
Almost the entire fall is in work that happens in the physical world: building trades, machine and process operatives, warehousing and delivery. Other service jobs also at the low end of the exposure ranking, such as waiting tables, care work, coffee shops, grew over the same period. So ‘manual work’ in general is not disappearing, but rather work tied to building and physical infrastructure.
This isn’t a new phenomenon. If we split occupations into those which are most and least exposed to AI today, it is clear from the data that the highest exposure occupations drove employment growth over the past 20 years to a much greater extent than the least exposed, well before the advent of generative AI.
That points to an interesting phenomenon: economic activity in the UK is increasingly disembodied from the physical world.
Much of this is the natural result of an increasing share of the services sector, and the relative decline of primary and secondary activities like agriculture and industry.
But part of it is self-inflicted. Planning constraints that restrict construction, high prices for production inputs such as industrial electricity, statutory floors on other input costs, or tax increases have all made many of these activities more expensive to carry out in the UK. That has translated into fewer of the kinds of jobs that are least exposed to automation. All of these barriers together cap how much building, infrastructure and industrial work there is to hire for.
This is most obvious in construction, because output has fallen alongside overall employment. Detailed planning approvals in England are down roughly 37% since 2019. Brick deliveries across Great Britain are down about 30% since 2018. Housing starts in London fell from 21,700 in 2022-23 to 4,200 just two years later. When fewer homes are approved, even fewer are started, so fewer bricks are laid, and fewer bricklayers are hired.
And it’s not just construction or the private sector where these consequences can be felt. According to research we carried out with Britain Remade, the UK pays 65% more than comparator countries in the OECD for its infrastructure.
Higher costs mean the same budgets deliver less physical output, which means fewer people employed carrying out the work. Meanwhile, more people are employed in the relatively unproductive jobs of navigating the regulatory and bureaucratic constraints produced by the UK’s increasingly complex web of approvals and vetoes.
The reduction in the number of roles and the relatively low value added, in comparison with other occupations, are clear signs that the UK currently places expensive barriers to real-world activity. This state of affairs is limiting how much physical private and public infrastructure we can afford to install.
Not only does this result in fewer opportunities for workers, it expands the compensation of more knowledge-intensive roles that do not add physical capacity. We should remind ourselves that projects like the Lower Thames Crossing have already cost north of £1bn before any bit of actual construction has happened. A costly process for deploying any kind of infrastructure compensates precisely the kind of work which does not actually build anything.
This should force us to consider whether our current concerns are well calibrated. Almost four years since the release of ChatGPT, AI has added to or at least not materially affected the demand for bit-intensive or knowledge intensive work: analysts, coders, cyber security, the occupations with underlying capabilities that are enhanced by the technology.
But under our current set of heavy constraints on physical work, very little of the value this is generating has resulted in greater demand for work that exists in tangible materials and activities. The labour market for these occupations depends to a much greater extent on how much gets built, connected and maintained in the physical world, and that quantity is set far less by technology than by planning committees, grid queues and payroll costs. UK policy over the last few decades has consistently and inadvertently raised the barriers on this kind of work, which happen to also constrain knowledge work much less.
Worry about making work pay, not automation
Southern newspapers in the 1930s predicted that the new mechanical cotton-picker would throw hundreds of thousands of sharecroppers into idleness. Some even called for the machine to be banned.Ironically, on this occasion the displacement turned out to be real, but was not caused by technological transformation. Federal acreage-reduction payments and the high factory wages of the war years had already emptied the cotton fields before mechanical pickers were deployed at scale. Opposition to technological progress laid at the feet of these new tools a transformation that had already taken place because of broader economic forces.
We are still making this same mistake. We look to technology as the culprit, and in the process forget that broader economic forces continue to shape the fabric of work across industries.
There is a tempting, and misguided, instinct to conclude that we should simply let these forces play out, and that displaced workers across all industries will naturally flow to where opportunities arise. In the current context, that is an exceptionally complacent and dangerous response.
Absence of evidence is not evidence of absence, and it remains very plausible that large-scale labour market disruption will arise if capabilities continue to evolve.
But more importantly, we are already failing to enable the flow of workers to opportunity at scale, and that is a much more urgent and pressing concern. The routes into knowledge work are through entry-level jobs, which tend to concentrate in the most productive cities. Access to these occupations by younger workers in employment will continue to be limited while market rates for housing remain close to their current levels, which is likely to continue if we build as little as we have.
Barriers to real world infrastructure also severely limit both jobs and wages in those roles that are least exposed to AI.The latest data release offers brief reassurance on the future of work in general, but it should raise our alarm at how much constraints to activity in the real economy have depressed economic opportunity for a large number of workers. While we worry about how to address the problem of technological unemployment, we should devote closer attention to how difficult we make building and rebuilding our physical infrastructure.
 | A guest post by
|