What the AI transition means for the UK labour market — beyond the headlines

Executive Summary

The UK labour market faces significant structural disruption from AI over the next five to ten years. The picture is more nuanced than either the optimists or pessimists suggest. Drawing on DSIT’s 2025 AI Activity in UK Business survey,[1] ONS labour force microdata,[2] OECD exposure indices disaggregated to UK Standard Industrial Classification codes,[3] and an original AI Policy Exchange survey of 420 UK employers,[4] this brief identifies three structural dynamics that current policy is failing to address.

First, the most exposed sectors are not the ones making headlines. Professional services (SIC 69–71), financial services (SIC 64–66), and public administration (SIC 84) face higher near-term disruption than manufacturing or logistics. Roughly 61% of task-hours in professional services are technically exposed to current-generation large language models and adjacent AI systems, compared with 28% in manufacturing.[3] Public discourse, and policy attention, remains fixated on blue-collar automation scenarios inherited from the robotics era.

Second, we introduce an AI Exposure–Adoption Matrix, which plots sectors along two dimensions: technical exposure to AI capability and organisational readiness to adopt. The matrix shows something important. The sectors where AI could deliver the greatest productivity gains, public administration, health, and education, are precisely those where institutional friction, procurement complexity, and risk aversion produce the slowest adoption. We call it the “procurement paradox.” It is a market failure only coordinated policy intervention can fix.

Third, the distributional effects of AI adoption follow a regressive pattern we describe as the “mid-skill squeeze.” AI augments the productivity of high-skill workers who can direct and interpret its outputs. It automates routine low-skill tasks that were already low-wage. It eliminates mid-skill roles (paralegal, junior analyst, claims assessor, procurement officer) that have historically been entry points into professional careers. The ONS Annual Survey of Hours and Earnings already shows that median earnings growth in the most AI-exposed mid-skill occupations has fallen 1.8 percentage points below the economy-wide average since 2023.[5] Unless policy intervenes, the dynamic will narrow the pipeline of social mobility at exactly the moment the government claims to be widening it.

This brief runs in six sections. Section 1 maps the exposure landscape across UK sectors and occupations. Section 2 runs granular sector-level analysis of the five most affected industries. Section 3 examines the adoption gap between theoretical capability and organisational deployment, and introduces the Exposure–Adoption Matrix. Section 4 analyses distributional effects, with particular attention to the mid-skill squeeze. Section 5 benchmarks UK preparedness against comparator economies. Section 6 closes with six policy recommendations calibrated to the evidence.

1. The exposure landscape

When the government talks about AI and the labour market, it tends to default to aggregate statistics: “AI could affect 10–30% of jobs.” These headline numbers, typically drawn from studies by Goldman Sachs, McKinsey, or the IMF,[6] are almost useless for policymaking. They conflate exposure with displacement, ignore the gap between tasks and jobs, and treat the economy as a homogeneous mass. Policy needs to know which jobs, in which sectors, at what speed, and with what distributional consequences. This section tries to provide that granularity.

Our analysis draws on four data sources: the OECD’s 2025 AI exposure index disaggregated to two-digit UK SIC codes using ONS workforce composition data;[3]the ONS Business Insights and Conditions Survey (BICS) waves 120–128, which now include questions on AI adoption;[7]DSIT’s AI Activity in UK Business survey, covering 2,000 firms across all sectors;[1]and an original AI Policy Exchange dataset surveying 420 UK employers on their AI deployment timelines, use cases, and perceived barriers.[4]The employer survey was conducted between September and November 2025 and deliberately oversampled mid-sized firms (50–249 employees), which make up the bulk of UK private-sector employment but are underrepresented in existing research.

The results show a sharp mismatch between public narrative and empirical reality. The five sectors with the highest task-level AI exposure scores: financial and insurance activities (SIC 64–66) at 64% of task-hours exposed; professional, scientific, and technical activities (SIC 69–75) at 61%; information and communication (SIC 58–63) at 59%; public administration and defence (SIC 84) at 53%; administrative and support services (SIC 77–82) at 48%.[3]By contrast, manufacturing (SIC 10–33) scores 28%, construction (SIC 41–43) scores 19%, and accommodation and food services (SIC 55–56) scores just 14%.

“Exposure” is not the same as “displacement.” A task being technically performable by an AI system does not mean it will be automated in practice. Whether exposure translates into job loss, transformation, or augmentation depends on a web of factors: the cost of the AI system relative to labour, the regulatory environment, the degree of client or patient trust required, the presence of complementary human skills, and the organisation’s capacity to integrate new tools. Our framework distinguishes three modes of AI impact on any given occupation: substitution (the task is automated and the role eliminated or consolidated), augmentation (the task is AI-assisted and the worker becomes more productive), and restructuring (the task is automated but the role is redesigned around higher-value activities).

Apply that three-way framework to the OECD exposure data and the picture gets more actionable. Of the 61% of task-hours exposed in professional services, we estimate that roughly 15% fall into substitution, 32% into augmentation, and 14% into restructuring.[8]The substitution share is highest in administrative and support services (22% of task-hours), where routine document processing, scheduling, and data entry face direct AI replacement with little need for human judgement. It is lowest in health and social care (7%), where regulatory requirements, patient trust, and physical-world interaction constrain automation even where the technical capability exists.

The geographic distribution of exposure adds a further dimension. London and the South East, with their concentration of financial and professional services, have the highest aggregate exposure: 52% of task-hours across all sectors, compared with 34% in the North East and 31% in Wales.[9] The capacity to absorb disruption is also highest in London, where labour market dynamism, retraining infrastructure, and job creation in AI-adjacent roles partially offset displacement risk. The regions most at risk of net negative outcomes are those with moderate exposure concentrated in a narrow range of sectors: the East Midlands, and Yorkshire and the Humber, where public administration and administrative services account for a disproportionate share of mid-skill employment. A purely national policy response will miss these regional asymmetries entirely.

2. Sector-level analysis

To get beyond aggregate exposure scores, we ran detailed analysis on the five sectors most affected by current AI systems. Each one combines exposure, adoption speed, workforce composition, and policy constraint differently. The analysis below draws on our employer survey data alongside published ONS and DSIT statistics.

Financial and insurance activities

UK financial services employs roughly 1.1 million people and contributes 8.3% of economic output.[10]It is the most exposed sector, and it is adopting fastest. DSIT’s survey found that 68% of financial services firms with more than 50 employees had deployed at least one AI system by mid-2025, against a cross-economy average of 34%.[1]The dominant use cases are fraud detection (used by 81% of adopters), customer service automation (74%), credit risk assessment (62%), and regulatory compliance monitoring (58%). Our employer survey found the median large financial services firm expects a 12–18% headcount reduction in operations and middle-office functions by 2029, partly offset by growth in AI engineering, data science, and compliance.[4]The net employment effect is estimated at −7 to −9% of current sector headcount. The roles being cut are disproportionately in the £28,000–£45,000 mid-skill band that acts as the entry ramp into financial careers for graduates outside the Russell Group.

Professional, scientific, and technical activities

This sector covers legal services, accounting, management consultancy, architecture, and engineering, and employs around 2.9 million people across the UK. AI exposure is high (61% of task-hours) but adoption is bifurcated. Large firms in the “Big Four” and Magic Circle have invested heavily. PwC alone has committed £1 billion to AI deployment across its global operations.[11]The sector is dominated by SMEs, though. 94% of firms have fewer than 50 employees, and our survey found that only 19% of small professional services firms had deployed any AI system beyond basic productivity tools like email summarisation.[4] The most affected occupations are junior legal researchers, trainee accountants, and associate-level management consultants, where AI can now do in minutes what used to take days of document review, data compilation, or slide assembly. The Law Society reported in January 2026 that training contract offers from Top 50 firms fell 14% year-on-year, the first decline not attributable to an economic downturn.[12]

Public administration and defence

The public sector is the starkest example of what we call the procurement paradox. 53% of task-hours are exposed to AI, overwhelmingly in document processing, casework management, correspondence handling, and internal reporting. The productivity gains from adoption would be enormous: the National Audit Office estimated in its February 2026 report that AI-driven automation of routine admin tasks across central government could yield £3.6–5.2 billion in annual efficiency savings.[13]And yet adoption is glacial. Only 22% of central government departments reported operational AI deployment in DSIT’s survey, and for local authorities the figure is just 11%.[1]The barriers are institutional rather than technical: procurement frameworks designed for traditional IT contracts with fixed specifications, the Government Digital Service’s cautious approach to algorithmic decision-making after the A-level grading controversy, data sharing restrictions between departments, and a civil service workforce where 67% of employees over 50 report low confidence in using AI tools, according to the Cabinet Office’s 2025 People Survey.[14]

Information and communication

The technology sector is both the creator and a significant target of AI disruption. The 59% exposure score reflects how much of the work (software development, content creation, data analysis, customer support) current AI systems can do partially or fully. What sets this sector apart is the speed of adoption: 79% of firms in our survey had deployed AI in core production workflows.[4]The workforce effects are already visible. Stack Overflow’s 2025 Developer Survey found that 43% of UK developers reported their teams had shrunk over the past year despite stable or growing output.[15] Junior software engineering roles, historically the most common entry point for computer science graduates, are being restructured around AI-assisted workflows that need fewer people with more experience. Graduate hiring in the UK tech sector fell 23% between 2024 and 2025, per the Institute of Student Employers.[16]

Health and social care

Health is a paradox of a different kind. Technical exposure is moderate (37% of task-hours), concentrated in diagnostic imaging, clinical documentation, appointment scheduling, and back-office admin. The clinical case for AI adoption is overwhelming. NHS England’s own analysis suggests AI-assisted triage and diagnostics could cut average wait times by 18–22% in high-volume specialties.[17]But adoption faces the most formidable barriers of any sector. MHRA regulatory pathways for AI-as-medical-device remain slow, with a median approval time of 14 months. NHS procurement cycles average 24–36 months from business case to deployment. Clinical staff resistance is significant: the BMA’s 2025 survey found that 52% of consultants were “concerned” or “very concerned” about AI in clinical decision-making, even while 71% acknowledged its potential to reduce administrative burden.[18] The result is a sector where the back-office productivity gains are real and achievable, but the transformative clinical applications remain years from deployment at scale.

3. The adoption gap

A common and consequential mistake in AI labour market analysis is to conflate what AI can theoretically do with what organisations will actually deploy. The gap between capability and adoption is where policy has the most leverage, and where current UK strategy is weakest. We propose a structured framework for thinking about that gap: the AI Exposure–Adoption Matrix.

The matrix plots sectors along two axes. The horizontal axis measures technical exposure: the share of task-hours in a sector that current AI systems can perform at or above the level of a median human worker. The vertical axis measures adoption readiness, a composite index combining five factors (digital infrastructure maturity, management AI literacy, regulatory permissiveness, procurement agility, and workforce receptiveness) each scored on a 0–100 scale using data from our employer survey and published indices.[4] The four quadrants that fall out of this each pose a different policy challenge.

The upper-right quadrant (high exposure, high adoption readiness) contains financial services and the technology sector. These sectors will largely manage their own transitions, driven by competitive pressure and existing digital capabilities. The policy priority here is distributional: making sure the gains from AI-driven productivity are not captured entirely by capital and senior employees, and that displaced mid-skill workers have viable transition pathways. The lower-right quadrant (high exposure, low adoption readiness) is where the procurement paradox is most acute. Public administration, health, and education sit here, with enormous potential for AI-driven productivity improvement but institutional friction that delays deployment by years or even decades.

The upper-left quadrant (low exposure, high adoption readiness) includes sectors like retail and hospitality, where AI deployment is concentrated in narrow applications (inventory optimisation, dynamic pricing, chatbot-based customer service) but most of the workforce performs physical-world tasks that remain out of AI reach. The policy challenge here is modest. The lower-left quadrant (low exposure, low adoption readiness) contains construction, agriculture, and parts of manufacturing. These sectors face minimal near-term AI disruption, though robotics and autonomous systems may shift them rightward over a longer time horizon.

Our employer survey gives granular data on the barriers driving low adoption readiness. Among firms that had identified at least one high-impact AI use case but had not yet deployed, the reported barriers were: lack of internal AI/data science skills (73% of respondents), uncertainty about regulatory requirements (61%), procurement and contracting processes not designed for AI systems (58%), concerns about data quality or availability (54%), inability to build a compelling business case given uncertainty about benefits (49%), and senior leadership scepticism or risk aversion (41%). Cost was cited by only 29% of respondents. The barrier is institutional, not financial.[4]

The temporal dimension of the adoption gap is critical for workforce planning. Our survey asked firms with active AI deployment plans to estimate the gap between initial pilot and full operational deployment. The median response was 22 months, but the variance was enormous: technology firms reported a median of 8 months, financial services 14 months, professional services 19 months, and public sector bodies 38 months.[4] Which means the labour market effects of the same underlying AI capability will land in different sectors at very different times. It is a rolling wave of disruption, not a single shock. Current policy treats AI labour market disruption as one event to prepare for, rather than a staggered process to manage in real time.

4. Distributional effects

The most concerning finding in our analysis is about the distributional impact of AI adoption across the skill and income distribution. The optimistic story says AI will “lift all boats” by raising economy-wide productivity. The evidence points the other way, to a pattern of regressive disruption that we call the “mid-skill squeeze.”[19] Any policy response that takes equity seriously needs to start here.

Here is how the mid-skill squeeze works. At the top of the skill distribution, AI acts mainly as an augmentation tool. Senior lawyers use AI to review contracts faster. Experienced analysts use it to process larger datasets. Consultants use it to generate first drafts that they then refine with expert judgement. These workers become more productive, and early evidence suggests their compensation reflects this. ONS data shows that earnings in the top decile of AI-exposed professional occupations grew 6.2% in real terms between 2023 and 2025, against 2.1% across all professional occupations.[5] At the bottom of the skill distribution, AI automates routine tasks (data entry, basic scheduling, form processing) that were already low-paid. The workers displaced from these roles usually move laterally into other low-skill service work. The impact is real, but the absolute earnings loss is modest, and the workers affected were already covered by the existing social safety net.

The mid-skill band, roughly the second and third earnings quartiles (annual salaries of £25,000–£50,000), bears the concentrated impact. These roles combine routine cognitive work, which AI can automate, with judgement and client interaction, which AI cannot yet replicate. In previous technological transitions that combination made such roles relatively secure. But current AI systems are specifically good at the routine cognitive component: summarising documents, drafting correspondence, doing standardised analysis, managing workflows. Once that component is automated, the remaining tasks often do not justify a full-time role. Or they get absorbed by the senior professionals whom AI has made more productive.

You can already see this in early-adopting sectors. In financial services, the ratio of junior to senior analysts at large UK investment banks has shifted from roughly 4:1 in 2022 to 2.5:1 in 2025, according to data compiled from Financial Conduct Authority regulatory filings.[20] In the legal sector, the number of paralegal positions advertised on the three largest UK legal recruitment platforms fell 31% between Q1 2024 and Q1 2026, even as demand for qualified solicitors remained stable.[12] The Big Four accounting firms have collectively cut their UK graduate intake by about 18% since 2023, while hiring more for AI and data roles.[21]

The social mobility implications are serious. Mid-skill professional roles (trainee accountant, junior solicitor, associate consultant, junior analyst) have historically been the main way graduates from non-elite backgrounds enter professional careers. The Social Mobility Commission’s 2024 report found that 62% of professionals from working-class backgrounds entered their current sector through exactly these mid-skill entry points.[22] If AI eliminates or drastically reduces these roles, the career ladder does not just lose a rung. It loses the rung that the broadest range of people could reach. The professional labour market that results looks more like an hourglass: a large base of low-skill service work, a large top of high-skill AI-augmented professionals, and a hollowed-out middle that used to be the bridge between the two.

The geographic dimension makes the inequality worse. DWP Stat-Xplore data shows that the mid-skill roles most exposed to AI (administrative officers, accounting technicians, legal secretaries, insurance underwriters) are disproportionately concentrated outside London.[23]44% of high-skill AI-augmented roles (data scientists, AI engineers, senior analysts) are located in London and the South East, but only 27% of the mid-skill roles they are displacing are based in the same region. The net effect is a geographic transfer of labour market value from the regions to the capital, layered on top of an already severe regional productivity gap.

5. International comparisons

The UK is not dealing with AI-driven labour market disruption in isolation. Its policy response should be informed by what comparator economies are doing, and also by what they are getting wrong. We benchmark the UK against five comparator jurisdictions: the United States, the European Union (with particular attention to Germany and France), Singapore, Canada, and South Korea. The comparison runs across four dimensions: exposure profile, adoption speed, institutional response, and workforce transition infrastructure.

On exposure, the UK’s profile is among the most concentrated in the OECD. Its services-heavy economy means 71% of employment is in sectors with above-median AI exposure, against 63% in Germany, 58% in France, 67% in the United States, and 74% in Singapore.[24]Only Singapore, with its even more concentrated services-and-finance economy, has a higher share. The OECD’s 2025 Employment Outlook ranks the UK third among G7 nations for aggregate AI task exposure, behind only the United States and Canada.[24]This is a direct consequence of the UK’s post-industrial economic structure. Decades of shifting away from manufacturing towards knowledge-intensive services has left the economy disproportionately exposed to exactly the capabilities that large language models and adjacent systems now have.

On adoption speed, the UK sits in an uncomfortable middle position. It lags behind the United States, where adoption is driven by aggressive private-sector investment and a permissive regulatory environment. The Stanford HAI AI Index 2026 reports that 52% of US firms with over 250 employees had deployed AI in core business processes by the end of 2025, against 41% of equivalent UK firms.[25]But the UK is ahead of the EU average of 29%, where the AI Act’s classification and compliance requirements have created what European Commission surveys describe as a “regulatory chill” in high-risk application domains. Singapore leads all comparators at 61%, reflecting its aggressive National AI Strategy 2.0, which includes direct co-investment in enterprise AI deployment.[26]The UK’s position (faster than Europe, slower than the US and Singapore) means it captures a moderate share of AI productivity gains while still copping the full displacement effects from competitor economies that adopt faster.

Institutional responses vary dramatically. The United States has largely treated AI labour market adjustment as a private-sector problem, with limited federal workforce transition support beyond existing programmes. The EU has focused on regulation and worker protection, embedding AI workforce provisions in the AI Act and proposing a directive on algorithmic management. Germany has taken a sectoral approach. Its Federal Employment Agency funds 47 sector-specific AI transition programmes in partnership with industry associations and trade unions.[27]France has invested heavily in AI skills through its France 2030 programme, allocating €2.2 billion to AI training and reskilling between 2023 and 2028.[28]South Korea’s Digital New Deal includes a dedicated AI Workforce Transition Fund of $1.3 billion over five years, with individual training accounts for workers in exposed sectors.

The UK’s institutional response, by contrast, has been fragmented and reactive. Responsibility is split across multiple departments: DSIT leads on AI policy, DWP on employment support, the Department for Education on skills, and HM Treasury on fiscal implications. No single body has a mandate to monitor, anticipate, and coordinate the workforce transition. The Apprenticeship Levy, the main existing instrument for employer-led training, is widely acknowledged to be poorly suited to AI reskilling. Its rigid standards framework, 12-month minimum duration requirement, and limited eligibility for modular or short-course training make it ill-adapted to the rapid, iterative skill development that AI adoption demands.[29] The UK Commission for Employment and Skills was abolished in 2017 and has not been replaced with any equivalent strategic body.

The most instructive comparison is probably Singapore’s SkillsFuture programme, which gives all citizens aged 25 and above a $S500 credit (topped up periodically) for approved training courses, with sector-specific transition support for workers in AI-exposed industries on top. The programme has achieved 67% uptake among workers in high-exposure sectors, against roughly 12% participation in equivalent DWP-funded skills programmes in the UK.[30]The gap is partly structural. Singapore’s programme is universal and individual-led, while UK programmes are employer-mediated and means-tested. It is also partly cultural: Singapore’s government has invested heavily in public communication that frames AI workforce transition as an opportunity rather than a threat. The UK can learn from this model without copying it wholesale, and we come back to this point in our policy recommendations.

6. Policy recommendations

Current UK policy treats AI workforce disruption mainly as a skills problem, to be solved by retraining programmes. That framing is incomplete. Left uncorrected, it will produce a policy response that is permanently behind the curve. Skills investment is necessary, but it is not enough without institutional infrastructure for monitoring, anticipating, and managing transition at sector level. Our six recommendations are designed to fill that gap. They are sequenced by urgency and calibrated to the fiscal and institutional constraints any realistic policy proposal has to acknowledge.

Recommendation 1: Establish the AI Labour Market Observatory

DSIT should create a standing analytical unit combining real-time employer survey data, ONS labour force microdata, HMRC PAYE real-time information, and sector-level AI exposure modelling. The Observatory should publish quarterly dashboards disaggregated by sector, region, and occupation, and feed early-warning indicators to DWP Jobcentre Plus and local authorities. Estimated annual cost: £8–12 million. That is trivial compared with the cost of reactive employment support after displacement has already happened. Germany’s IAB (Institute for Employment Research) is a proven model for this kind of standing labour market intelligence, and its AI-focused analysis has demonstrably shaped the Bundesagentur für Arbeit’s transition programmes.[27]

Recommendation 2: Mandate AI workforce impact assessments for public sector procurement

All public sector procurement contracts above £10 million that involve AI systems should be required to include a workforce impact assessment, modelled on environmental impact assessments. The assessment should estimate the number and type of roles affected, the timeline for impact, and the mitigation measures being taken: redeployment, retraining, and transition support. This does two things at once. It makes sure government itself accounts for the workforce consequences of its own adoption decisions, and it creates a data stream for the Observatory. The Cabinet Office’s existing Social Value Model is a foundation that could be extended to cover AI workforce considerations without building an entirely new compliance framework.

Recommendation 3: Reform the Apprenticeship Levy into an AI-era Skills Levy

The current Apprenticeship Levy should be reformed to let employers spend on short-course, modular AI reskilling programmes accredited by a new fast-track quality assurance process. The 12-month minimum duration requirement should be replaced with a competency-based framework that lets workers upskill in 8–16 week blocks while remaining in employment. The reformed levy should also let firms fund training for workers at risk of displacement, not only those in stable roles. The Confederation of British Industry has estimated that levy reform alone could increase employer-funded AI reskilling by 40–60% within two years of implementation.[29]

Recommendation 4: Create sector-specific transition compacts

For the five most exposed sectors identified in this brief, the government should convene tripartite transition compacts bringing together employer associations, trade unions and professional bodies, and relevant government departments. Each compact should build a sector-specific transition plan with a five-year horizon, updated annually from Observatory data. The compacts should have authority to direct a portion of sectoral levy funds towards agreed transition priorities. There is precedent for this in the UK’s Sector Skills Councils (too weak and unfocused to succeed) and Germany’s Transformation Councils (more effective, thanks to genuine tripartite governance and dedicated funding).

Recommendation 5: Introduce AI transition individual learning accounts

Drawing on the Singaporean SkillsFuture model and the French Compte Personnel de Formation, the government should pilot individual learning accounts for workers in the most AI-exposed occupations. Each account would carry £2,000–£3,500 in training credit, topped up annually, usable at any accredited provider for courses aligned with the transition compact priorities for the worker’s sector. The accounts should be portable across employers and accessible without means-testing. DWP’s existing Universal Credit conditionality framework could be adapted to include AI reskilling as a recognised “work preparation” activity for claimants in exposed occupations, so the safety net supports transition rather than just cushioning displacement.

Recommendation 6: Address the procurement paradox directly

The Government Digital Service, working with the Crown Commercial Service, should create a dedicated AI procurement pathway for public sector bodies. This should include pre-approved AI vendor frameworks with streamlined due diligence, standardised data processing agreements for common AI use cases, a “regulatory sandbox” for AI deployment in low-risk public sector applications, and a dedicated £150 million AI Adoption Fund for local authorities and NHS trusts, structured as matched funding to incentivise institutional commitment. The current procurement framework was built for buying software licences and consultancy. It is structurally incapable of supporting the iterative, experimental deployment model AI adoption requires. Without procurement reform, the public sector will sit in the lower-right quadrant of the Exposure–Adoption Matrix indefinitely, forgoing billions in productivity gains while the private sector races ahead.