Written submission from Dr Bouke Klein Teeselink (AIB0081)

 

Artificial Intelligence, Business and the Future of the Workforce

King’s College London

Coordinated by Dr Bouke Klein Teeselink

Assistant Professor, Department of Political Economy

King’s Institute for Artificial Intelligence

April 2026

 

Introduction 

  1. King’s College London houses researchers studying AI’s effects on labour markets, workplaces, and regulation across a number of disciplines and departments, including Political Economy, King’s Business School, Informatics, Digital Humanities, the Dickson Poon School of Law, and Culture, Media and Creative Industries. The coordinating author (Klein Teeselink) is an economist specialising in AI’s labour market effects, appointed Chief Economist at the AI Objectives Institute in 2025. The evidence below draws on peer-reviewed publications, working papers, and reports.  

Theme 1. The future of AI 
 
How is AI likely to evolve over the next decade? What are the scenarios we should plan for? 

  1. Over the next decade, AI is likely to disrupt a substantial number of current hours with the shift to automation. UK needs to prepare sustained automation pressure on professional, technical roles, clerical and manual roles.  
  1. It is helpful to classify four categories of AI application. Predictive (forecasting outcomes from data), generative (producing text, code, and images), agentic (autonomous multi-step decision-making), and embodied (robots acting in physical environments). Paragraphs 4-6 address these, drawing from work at King’s and collaborators (Chalmers et al.,  2026). 
  1. The first mode, and the one most extensively examined in early organizational scholarship, is predictive and extractive AI. These systems rely on pattern recognition to forecast outcomes, classify entities, or detect anomalies in large datasets.   
  1. A second and more recent mode is generative AI. These systems generate synthetic text, images, code and other multimodal artefacts, shifting attention from the analysis of existing data to the creation of new material. In many workplaces, generative AI is likely to have a structural impact as it makes many established job roles redundant, forcing a change to the shape of organizations. 
  1. A third, still emerging mode involves reasoning and agentic AI. These systems go beyond the generation of artefacts and begin to participate in more structured forms of problem solving. They can break down tasks, invoke external tools such as search functions or code execution environments, retrieve information, and construct multi-step analyses. These are reasoning-capable systems. There is a growing literature on how the introduction of these systems is influencing the way professionals in a range of contexts interpret and solve problems. As AI assumes greater involvement in reasoning processes, this raises concerns around the future of work once agentic AI becomes widely implemented.  
  1. A final, and still comparatively underdeveloped, mode concerns embodied AI. This refers to the integration of computational intelligence with physical form, i.e. robotic systems. We anticipate that embodied AI will become one of the most significant areas of organizational research in the coming decade as societies confront the radical implications of near zero-marginal-cost physical labour. 
  1. Evidence suggests that AI’s occupational impact will extend beyond routine tasks, with mid-qualification white-collar roles in healthcare, IT, and manufacturing among the most affected, although to be clear, this may not necessarily always mean that the effect takes the form of significant job replacement. This is because in many cases, AI applications augment rather than replace roles, and others will apply to roles where there are currently job shortages meaning that there is less scope for job replacement (Septiandri, Constantindes, Quercia, 2024).  
  1. Estimates of the scale of disruption vary, but converge on a substantial share of current work hours being automatable within this decade. McKinsey estimates that up to 30 per cent of US work hours could be automated by 2030. 

Theme 2. AI Adoption 

Where is AI currently being deployed across the UK economy?

  1. Using data on millions of UK job postings, I measure AI adoption by identifying firms that post job advertisements requiring generative AI skills (such as prompt engineering or LLM integration) after the release of ChatGPT in November 2022. Each posting is verified using a language model to eliminate false matches. Firm size is the strongest predictor of AI adoption. This likely reflects the ability of large firms to spread fixed adoption costs (training, infrastructure, and workflow redesign) across more workers (Klein Teeselink, 2025).
  2. AI adoption also varies by industry, with the information sector leading, followed by professional services and finance. Agriculture, accommodation and food services, and construction have the lowest adoption rates (Klein Teeselink, 2025).

What are the main barriers to AI adoption? 

  1. It is helpful to categorise different types of barriers. Four key barriers identified in King’s research include: financial constraints due to firm size; lack of, or inadequate infrastructure; regulatory uncertainty; and skills/knowledge gaps.  
  2. As noted above, firm size is the largest predictor of AI use. That the largest UK firms have adopted AI at rates many times higher than the smallest. This suggests that one barrier is financial constraints, especially faced by smaller firms.
  3. Data gathered since the release of ChatGPT in November 2022, indicates that firm size is the largest predictor of AI use. The largest UK firms have adopted AI at rates many times higher than the smallest. This pattern remains the case where more profitable firms adopt AI at higher rates. (Klein Teeselink, 2025), 
  4. This also holds in relation to the second barrier, infrastructure. This barrier is explicit in place like retail where implementation issues are linked to legacy systems that have limited data availability and processing capacity. With these barriers in mind, it becomes clear that without targets support, small and medium-sized enterprises (SMEs) risk falling behind, widening productivity dispersion.  
  5. In some specific contexts, regulatory uncertainty can be a barrier. For example, firms are avoiding using generative AI in selection decisions during recruitment processes due to lack of regulatory clarity and suggests that the uncertainty surrounding potential regulations may inhibit wider adoption of AI into the workplace (Nyberg et al., 2025).  
     
  6. A fourth key barrier that we have noticed from our research is a skills and knowledge gap for AI use. The evidence highlights that these challenges, along with resistance to change and worker mistrust of AI, are among the most frequently cited barriers. These concerns and lack of knowledge are reported as ranking above technological constraints in some research. This suggests that even if firms have the necessary infrastructure, a lack of acceptance, and understanding of AI can significantly impact its adoption into the workplace. (Robertson, Botha, Oosthuizen & Montecchi, 2025). 

Theme 4. Impacts on work & workers 

Which sectors, regions and groups are most likely to be affected by AI, and what are the implications for inequality?

  1. Previous waves of automation displaced middle-skill, middle-wage workers, creating a “hollowing out” of the wage distribution. LLMs appear to do the opposite. A KCL labour market study analysing job postings between 2021 and 2025 found that firms whose workers perform tasks that large language models (LLMs) can automate have reduced employment and curtailed hiring.

High compensation firms show employment reductions roughly three times larger than the full sample, while low-compensation firms show near-zero effects.

  1. Whether this compression reduces inequality depends on which tasks AI removes.  When AI displaces an occupation’s low-expertise tasks, leaving specialised work intact, advertised wages rise because the bar for entry increases. When AI removes high-expertise tasks, wages fall.
     
  2. Policy responses should account for this distinction: the labour market consequences depend not only on how many tasks are exposed but on which tasks.

How is AI used in recruitment and HR processes, and what broad safeguards should be in place to ensure fairness and transparency in AI-influenced recruitment and employment decisions? 

  1. Research highlights that AI systems used in recruitment and workplace management introduce significant risks for bias, discrimination, and opacity, which existing legal frameworks were not designed to address.   
     
  2. AI is already widely embedded in employment processes, particularly through applicant tracking systems (ATS), which are used by 98 per cent of Fortune 500 companies. These systems automate CV screening by matching candidates to keywords, which can reject up to 75% of applicants before any human review. Although this increases efficiency it does so at the cost of fairness. KCL research shows that these systems can engage in proxy discrimination. Proxy discrimination occurs in these systems through inferring protected characteristics from seemingly neutral data such as postcodes or career breaks as proxies for ethnicity or gender, respectively (Kelan, 2024).   

 

  1. More broadly, AI hiring systems exhibit three key sources of bias. First, system training data can encode historical discrimination. Second, bias can be embedded into systems by predominately non-diverse development teams. For example, systems optimised for organisational fit risk reinforcing workplace homogeneity. Third, AI decision-making is opaque which makes its outcomes, like hiring decisions, challenging to audit or challenge. (Kelan, 2024).
      
  2. Current regulation struggles to address these issues. The Equality Act 2010 was not designed for algorithmic decision-making and may be insufficient to tackle AI-driven discrimination. Similarly, the EU AI Act does not clearly define discrimination in algorithmic contexts or set out specific measures to prevent it, and its risk-based approach relies heavily on self-identification by developers, weakening enforcement (Albrengues & Lu, 2025).   
     
  3. Overall, AI in employment creates risks of bias and opacity that current UK regulation does not adequately address. Stronger legal definitions of algorithmic discrimination, alongside enforceable workplace safeguards, are needed to ensure fairness and transparency. 

 Recommendations 

  1. Mandatory impact assessments for AI in hiring and employment decisions. Current voluntary approaches are insufficient. AI systems used in recruitment, performance management, and termination decisions should require independent audits, with results disclosed to workers and regulators.
     
  2. Enforceable standards for algorithmic management. The Fairwork AI Principles (Fair Pay, Fair Conditions, Fair Contracts, Fair Management, Fair

Representation) provide a ready-made framework. Workers subject to algorithmic management should have the right to an explanation of AI-driven decisions that affect them, and the right to appeal.

 

  1. Preserving entry-level training pipelines. The Government should consider structured apprenticeship or mentoring requirements, or tax incentives

for firms that maintain training pipelines alongside AI adoption.
 

  1. Strengthening the evidence base.
    Job posting data can serve as an early warning indicator. Hiring responds to AI exposure roughly eight months before actual headcount reductions materialise. The Government should invest in real-time labour market monitoring systems that track AI-driven shifts as they occur.
     
  2. Regulating AI use cases, not the technology itself. Senior business leaders concur that policy and laws should target GenAI use cases rather than the

technology itself (Nyberg et al., 2025). A use-case approach allows proportionate

regulation that does not stifle beneficial adoption while protecting workers in

high-risk applications.

 

  1. I would welcome the opportunity to provide oral evidence to the Committee on any of the themes addressed above.

References

Albrengues A, Lu L. Weight of gender in artificial intelligence models’ implementation in the European Union non-discrimination laws. Law Ethics Technol. 2025(3):0007, https://doi.org/10.55092/let20250007. 

Chalmers, D., Hunt, R. ‘., Pachidi, S., Potočnik, K., & Townsend, D. (2026). The Acceleration of Artificial Intelligence: Rethinking organization and work in an era of rapid technological change. Journal of Management Studies, 63(2), 285–314. https://doi.org/10.1111/joms.70063

Kelan, E. K. (2023). Algorithmic inclusion: Shaping the predictive algorithms of artificial intelligence in hiring. Human Resource Management Journal, 34(3), 694–707. https://doi.org/10.1111/1748-8583.12511

Klein Teeselink, B. (2025). Generative AI and labor market outcomes: Evidence from the United Kingdom. Working Paper https://papers.ssrn.com/sol3/papers.cfm?abstract_id=5516798

Nyberg, A. J., Schleicher, D. J., Bell, B. S., Boon, C., Cappelli, P., Collings, D. G., Molle, J. E. D., Feuerriegel, S., Gerhart, B., Jeong, Y., Korsgaard, M. A., Minbaeva, D., Ployhart, R. E., Tambe, P., Weller, I., Wright, P. M., & Yakubovich, V. (2025). A brave new world of human resources research: navigating perils and identifying grand challenges of the GenAI revolution. Journal of Management, 51(6), 2677–2718. https://doi.org/10.1177/01492063251325188

Robertson, J., Botha, E., Oosthuizen, K., & Montecchi, M. (2025). Managing change when integrating artificial intelligence (AI) into the retail value chain: The AI implementation compass. Journal of Business Research, 189, 115198. https://doi.org/10.1016/j.jbusres.2025.115198

Septiandri, A. A., Constantinides, M., & Quercia, D. (2024). The potential impact of AI innovations on US occupations. PNAS Nexus, 3(9), pgae320. https://doi.org/10.1093/pnasnexus/pgae320