How the WFH factor complicates the AI jobs story

What's really eroding the first rung of the career ladder?

This is a special multi-part edition of Post Haste, exploring AI’s effect on the labour market. You can read part one here and part two here.

Background image from The Parable of the Unmerciful Servant by Jan Sanders van Hemessen

By Dr Pedro Serôdio (@pdmsero)

The automation of all knowledge work by artificial intelligence would be the single most transformational event in the history of human civilisation.

But three years on from the release of ChatGPT, UK data does not yet show signs of a sharp and radical disruption. Employment in the occupations most exposed to AI has grown faster than in less exposed occupations. Declining wages in these roles precede the emergence of enterprise grade generative AI systems. Across a wide range of empirical approaches, geographies and outcomes, economists are in somewhat unusual agreement: aggregate effects are small or null – at least on the metrics we can easily measure.

So what story should we be telling about AI in Britain?

The erosion of the first step on the career ladder offers a seemingly straightforward answer. AI, some conclude, is already leading to fewer opportunities for young people. This particular concern is easier to map onto what the data shows, but is also more complicated than the headlines suggest.

It may be that we are asking the wrong question. Instead of fretting over whether AI will disrupt labour markets, or when, we should focus on whether labour markets are ready to absorb the shock if it comes.

For the UK, it is not obvious that this would be the case. The country has historically had a more flexible labour market than its European peers, and that flexibility was the channel through which young people moved into work faster than many of its counterparts in the OECD. Three structural changes have eroded it in the past five years: the post-pandemic shift to hybrid work, the compression of the youth wage floor, and the layering of new regulatory protections that raise the cost of trial employment.

This piece sets out what the evidence says on what AI has done so far, and asks whether structural changes to the UK labour market have left it well-prepared to handle the economic consequences of artificial intelligence.

Aggregate signals

If a country were going to show the labour market fingerprint of generative AI early, it would be Denmark. Technology adoption is high; the labour market is relatively flexible.It has the kind of high-quality administrative data that economists usually only dream about. A study of 11 exposed Danish occupations published last year found essentially no impact: earnings are within 1% of the baseline and hours unchanged, with the pass-through from AI-driven productivity gains to wages running at 3-7%.

At Anthropic, researchers found that unemployment for workers in the most exposed quartile, three years into the AI boom, looks indistinguishable from that of workers with no exposure. They concluded that AI’s labour market effects so far resemble the slow erosion of the China trade shock of the 2000s more than the sudden disruption of a pandemic.

At the business level, the first internationally representative firm survey on AI across the UK, US, Germany and Australia asks firms themselves to put a number on AI’s employment impact over the past three years. The UK figure is −0.14%; and “essentially zero” across all four countries.

Absence of evidence is not evidence of absence

It is tempting therefore to declare victory over doomsayers and confidently predict that concerns are overblown, but on looking deeper it becomes apparent we should be more circumspect.

Why? Because career ladders thin from the bottom.

Labour markets are not like the market for oranges. Because they are built on long-term relationships between employers and employees, the first sign of a shock shows up in vacancies rather than in headcount.

Firms first stop posting jobs that AI might be able to do instead of engaging in costly rounds of layoffs. Hours per worker fall before workers are let go, because firms find it easier to adjust how much they require from workers rather than restructuring.

Erosion at the bottom rung

A weakening of entry-level roles and reduction in the number of vacancies seems to now be a relatively well-accepted feature of the data across a number of countries.

For the United States, the United Kingdom, Canada and Australia, the junior share of new hires has fallen by 8-11 percentage points below its 2019 baseline. Entry-level hires are down 14-29%; senior hires are up 5-21%. An observable impact across four countries with different institutions and economic trajectories provides us with the strongest piece of evidence for a common underlying shock.

The working assumption, supported by research, has been that this shock was generative AI.

Using ADP payroll data on 25 million US workers, researchers at Stanford’s Digital Economy Lab last year found a 16% relative employment decline among 22-25 year olds in the most AI-exposed quintile of occupations, against 6-9% growth among workers aged 35 and above in the same occupations. Another working paper from Harvard economists used résumé and posting data on 65 million US workers, finding junior employment at AI-adopting firms fell 9% within six quarters of adoption, with senior employment unaffected and an 80% drop in junior hires per quarter from the pre-period mean. And for the UK, Dr Bouke Klein Teeselink’s analysis of 75 million employment ‘spells’ found maximally exposed firms cut total employment by 4.5% 18 months after ChatGPT, with the heaviest effects in high-wage occupations and worst in London.

So there does appear to be a weakening of entry-level openings and transitions. That is where we would expect labour market disruption from technological transformation to first emerge. The question is whether this narrative is as straightforward as it first appears

The remote work confound

Entry-level roles are a better candidate for identifying any potential effect of AI on employment, because they are particularly susceptible to shocks. But that same sensitivity makes them more responsive to every other shock that has hit labour markets over the last few decades, and there have been several: the Great Recession and the sluggish recovery across many countries, followed by the Eurozone crisis, Brexit, the COVID-19 pandemic, the commodity-price shock from the invasion of Ukraine, the ensuing inflationary spikes and exit from near-zero interest rates, and finally the emergence of generative AI. Many of these shocks arose in the last six years alone, and most of them would entail a weakening of economic conditions that would translate into weaker vacancies and fewer entry-level posts.

Some of these are red herrings. Why would a weaker economy or higher interest rates penalise occupations exposed to AI more than other jobs?

But one of the more interesting recent papers in this literature, authored by two LSE academics, takes up exactly that question and offers another reason. In The Broken Ladder: AI, Remote Work, and Early-Career Hiring, Peter John Lambert and Yannick Schindler argue that there is a better explanation for the effect on these occupations which hasn’t been sufficiently factored into these early explanations: working from home.

They find that occupations with WFH exposure are highly correlated with those that have generative-AI exposure. Software developers, data scientists, accountants and management consultants tend to be both remote-friendly and susceptible to AI substitution. Janitors, construction labourers, roofers and electricians, on the other hand, all work on location, and are therefore less exposed on both margins.

Viewed this way, both look like plausible drivers of the decline of entry-level roles. A one-standard-deviation increase in either WFH exposure or AI exposure is associated with a 1.4-1.9 percentage-point fall in the junior share of new hires by 2025, alongside a smaller fall in the share of postings asking for three or fewer years of experience. But when researchers account for both these changes at once, the picture changes. The effect of working from home remains stable, but the effect of generative AI collapses to between −0.4 and +0.3 percentage points - statistically indistinguishable from zero.

Once more, we must resist the temptation of jumping to conclusions: even if the increased prevalence of remote work provides a better account of the decline of entry-level jobs than generative AI, that does not negate the potential threat of automation to human labour. What we can say is that a large share of the observed impact on entry-level roles in this decade so far is better explained by the post-2020 reorganisation of working arrangements than by the post-2022 release of ChatGPT.

How hybrid arrangements can cost young workers

For those already established in their careers, hybrid arrangements turn out to be a desirable equilibrium. In a randomised controlled trial led by the economist Nick Bloom at Trip.com, a Chinese technology firm with 35,000 employees, hybrid workers had 33% lower attrition than fully office-based ones. They also reported higher job satisfaction, with the largest effects on the workers one might most expect to value flexibility: non-managers, women, workers with long commutes.

Predictably, this story is quite different for workers still acquiring deep expertise. For them, the office provides something that remote work cannot.

In a study which followed 61,000 Microsoft employees through the first six months of 2020, when the company went fully remote, findings indicated that internal collaboration networks became markedly more static and more siloed. This came down to fewer ties bridging disparate parts of the organisation, less synchronous interaction, more written asynchronous communication, and (the authors argue) less of the cross-team movement of tacit knowledge on which the company had been relying.

It may have become a cliché rolled out by executives, but the “water-cooler effect” of in-person learning and collaboration has been repeatedly observed. The office has been found to boost productivity among Indian data-entry workers and Turkish call centre operatives. At Fortune 500 companies, software engineers who sit near teammates receive more feedback on their code – a crucial learning opportunity for newer employees.

The Broken Ladder gives us a clean theoretical exposition of this mechanism. Firms hire workers not only for their immediate contributions, but also for their potential. For early-career employees, that involves a substantial degree of training and investment in their capability. The investment return depends on two parameters: how quickly a young worker learns on the job, and how much senior time is required to make that learning happen. Working from home worsens both of these dimensions at once, slowing the rate at which juniors pick things up and raising the cost in senior calendar time of each new hire who needs supervision. Neither change requires anything to be different about the technology involved in the work itself.

AI might also play a role here. There is some compelling modelling suggesting that generative AI flattens the learning curve by absorbing the easier tasks that made more junior workers useful from the start, raising the threshold of what they have to know on day one.

What this leaves us with is the conclusion that the flexible-work equilibrium, though it may be appreciated by established workers and measured as roughly productivity-neutral by firms, imposes a real cost on younger cohorts. This cost appears somewhat outside the firm rather than inside it, and is expressed through reduced hiring decisions that make early-career roles harder to justify. From far enough back, the pattern in the data looks like entry-level AI displacement, but closer up, we can see that much of what is actually happening is an erosion of entry-level work. This in turn may well accelerate a degree of automation that further worsens terms for young workers.

That brings us back to AI. Recall that in the first piece in this series, we explored the importance of tacit knowledge as a barrier that automation has greater difficulty overcoming because much of that information is difficult to codify. The clear implication of this is that younger workers will necessarily also struggle to learn on the job if they have to perform much of it away from where critical tacit information is passed along. Young workers who build up human capital lacking in these important dimensions will not just struggle to progress in their careers at the same rate as older workers: they will also face greater risk of automation as a result of their working arrangements. Increasingly, flexible working may turn out to represent a greater benefit to more senior and established workers, who enjoy a better work-life balance and can leverage their impact through automation, while giving younger workers a more rickety ladder to climb.

The geography of who lives where

From a UK perspective, there is an added dimension to the problem of entry-level work and the potential of automation. There is a clear geographic pattern to occupations most exposed to both AI and WFH: they are concentrated in London and the South East of England. Klein Teeselink’s UK study identifies the same set of occupations (software, finance, professional services) as the ones where companies are cutting entry-level posts most aggressively.

These are also the urban labour markets where housing costs are highest and the cost-of-living pressure on a 23-year-old in their first job is most acute. These features produce a somewhat unique confluence of factors that make hiring younger workers more challenging: many of these roles can be performed remotely and are at least partly automatable, while working on location is made pricier by cost of living pressures that effectively make many of these roles unviable to young workers.

London has the highest natural change and the highest net international migration of anywhere in the country, yet its overall net internal migration is negative: on balance, more people leave for the rest of the UK than arrive. Among 18-to-29-year-olds the city is still a net draw, but that draw is weakening. In 2019 it gained about 33,000 young adults on net from the rest of the UK; by 2024 that had fallen to around 24,000, as outflows rose faster than inflows. The narrowing is suggestive of younger workers finding it harder to bear the cost of living and foregoing the career opportunities that a move to London used to offer.

Entry-level white-collar jobs are being re-engineered around a hybrid working equilibrium that entails additional disadvantages for new entrants, just as alternative entry points into the labour market more broadly through the hospitality and retail sector are eroded by increases in the cost of hiring younger workers.

Post Haste: Why can’t young Brits find work?

The labour market entry path that characterised the UK economy before the pandemic for new graduates would entail something like moving to London after a degree for an entry-level role at a professional-services firm, learning quickly on the job by being in the office every day, and gradually developing the knowledge, skills, and network that would support a higher wage over the next sequence of moves.

Those steps are less straightforwardly available in 2026. Pinning this state of affairs on remote work or AI alone would be to oversimplify it. Instead we have a complex interplay of hybrid arrangements, a capital that is more expensive to live in, and hiring costs that shrink entry-level opportunities.

The UK has lost its main advantage

If AI adoption accelerates as much as financial markets suggest, the structural problems facing the UK will only worsen. We are now beginning to live through the kind of youth employment slump that the AI displacement story has been expected to generate.

The 18-24 NEET (not in employment, education or training) rate was at 15.2% in the fourth quarter of 2025, and the roughly 220,000 payroll jobs the country has lost since the mid-2024 peak are accounted for entirely by under-35s. The longer view is the same: since August 2019, payrolled employment among 18-to-24-year-olds has fallen about 2%, while among those 25 and over it has risen nearly 6%.

The recent Milburn review of the increase in NEETs sets these numbers in context: what looks like a stable headline rate actually conceals a slow drift from cyclical unemployment, which falls during times of economic recovery, towards persistent inactivity, which does not. This belies a structural deterioration of the youth labour market that has rapidly accelerated.

Looking at where those job losses are concentrated makes clear that much of this is unrelated to AI adoption. The steepest employment falls for 16-24 year olds have happened in transport, retail, hospitality and communications, sectors that score lowest on every available AI exposure index. The occupations highest on those indices are still showing the strongest employment and pay growth in the UK economy: software engineering, finance, professional services.

An AI displacement story would predict the opposite distribution of losses, which is hard to reconcile with observed impact.

What the data does fit much better is the gradual deterioration of one of the strengths of the UK economy and its labour market: a flexibility which made transitions into and out of roles easier and lent it considerable dynamism.

For much of the last three decades, the UK’s youth employment record was among the very best in the developed world: comparable to Denmark, Switzerland and the US on most of the relevant measures, and some way ahead of France, Italy, or Spain. This labour market was characterised by light regulation around trial employment, wage floors which did not aggressively bind in less economically prosperous regions, and a willingness across employers and policymakers to let entry-level pay be genuinely entry-level.

Taken together, those three helped create a channel through which young people moved into work faster in the UK than in most peer economies. For different reasons, all three have weakened sharply in the same five-year window, and while AI technology is not expected to play a similar role for sectors with low automation potential in the near future, transitions into the labour market will only become harder across all occupations for the median young worker.

White collar work is experiencing a version of this through the combination of changing working arrangements and the wage floors imposed by high costs of living in its most economically dynamic cities. The post-pandemic shift to hybrid working would suggest that this, more than ChatGPT, is what drives weakness in entry-level vacancies in the UK data, as well as the decline in the junior-share across the four economies the WFH literature looks at.

Heavily restricted access to housing and the high cost of living compound that shift, by making the same young workers less able to actually live in the cities where those firms are concentrated. Across both ends of the youth labour market, the UK’s current entry-level problem has been produced from a combination of policy choices and behavioural shifts that have very little to do with AI itself, but could be exacerbated by its impacts.

Is the UK prepared?

On an initial reading, the evidence regarding AI’s effect on the labour market is reasonably reassuring.

AI may not necessarily displace human labour, and throughout this series we’ve explored how the underlying economics can be compatible with continued employment and wage growth. What we have seen so far is consistent with that, with limited impacts on occupations that have high levels of exposure to the technology.

But thinking this settles the question would be complacent. There are good reasons to worry about AI’s impact even if it only ends up driving significant reorganisation of the labour market. This reorganisation will require flexibility and the kind of dynamism that can tolerate risk and uncertainty, and countries that lack the mechanisms for enabling this kind of dynamism will be exposed to the costs of disruption through international competition but lack the means to capitalise on the opportunities it creates.

Even if AI’s macro effects turn out to be very modest after all, the frictional adjustment that any large technological shift produces can remain a real problem. Industry transitions of any size historically take a decade or more to play out. The workers most exposed to them have always tended to be those at the start of their careers, with the most years of working life ahead of them and the least sunk cost in any specific firm or role. The UK is going into the next decade with its youngest cohort facing one of the worst entry-level labour market conditions in a long time, for reasons that have very little to do with any new technology. The case for fixing that part of the picture is the same, whether or not AI ends up doing what its proponents say it will.

Is the UK prepared for large AI labour market effects if they do arrive? A detached reading of the evidence and of recent changes to its labour and housing markets suggests that it is not.

The country’s traditional advantage in moving young people into work has weakened on each of the three components that produced it. A youth labour market that is already struggling to absorb its current cohort is in no shape to handle the kind of reallocation shock a large AI effect might impose.

Artificial intelligence promises to completely reshape economic activity, and it may start by changing how we work. The next few years will be a sharp test of whether the UK stands to gain from these changes, big or small, or whether it will be increasingly forced to watch from the sidelines as other countries enjoy those benefits.