Code, Capital, and Clusters: Understanding Firm Performance in the UK AI Economy

The UK has established a distinctive position in the global AI landscape, driven by rapid firm formation and strategic investment. However, the interplay between AI specialisation, local socioeconomic conditions, and firm performance remains underexplored. This study analyses a comprehensive dataset of UK AI entities (2000 - 2024) from Companies House, ONS, and glass.ai. We find a strong geographical concentration in London (41.3 percent of entities) and technology-centric sectors, with general financial services reporting the highest mean operating revenue (33.9 million GBP, n=33). Firm size and AI specialisation intensity are primary revenue drivers, while local factors, Level 3 qualification rates, population density, and employment levels, provide significant marginal contributions, highlighting the dependence of AI growth on regional socioeconomic ecosystems. The forecasting models project sectoral expansion to 2030, estimating 4,651 [4,323 - 4,979, 95 percent CI] total entities and a rising dissolution ratio (2.21 percent [-0.17 - 4.60]), indicating a transition toward slower sector expansion and consolidation. These results provide robust evidence for place-sensitive policy interventions: cultivating regional AI capabilities beyond London to mitigate systemic risks; distinguishing between support for scaling (addressing capital gaps) and deepening technical specialisation; and strategically shaping ecosystem consolidation. Targeted actions are essential to foster both aggregate AI growth and balanced regional development, transforming consolidation into sustained competitive advantage.

💡 Research Summary

This paper provides a data‑driven mapping of the United Kingdom’s artificial‑intelligence (AI) sector from 2000 to 2024, focusing on how firm‑level characteristics, technical specialisation, and local socioeconomic conditions jointly shape firm performance and sector dynamics. The authors assembled a longitudinal dataset of 4,392 AI‑related entities by merging Companies House registration records, Office for National Statistics (ONS) postcode‑level census data, and web‑scraped information from glass.ai. After filtering for “companies” and “universities” that develop core AI products, the final sample comprises 3,143 active firms and 398 dissolved firms, with a subset of 529 firms (451 after cleaning) providing complete revenue and employee data.

Methodology
Two analytical strands are pursued. First, the authors employ CatBoost, a gradient‑boosting decision‑tree algorithm well‑suited for heterogeneous tabular data, to model operating revenue and revenue per employee as functions of (i) firm‑level variables – employee count, years of operation, sector (SIC code), and a quantitative AI‑specialisation score derived from TF‑IDF weighting of website‑derived keywords (e.g., “machine learning”, “natural language processing”, “predictive analytics”) – and (ii) postcode‑level socioeconomic indicators – population density, proportion of residents with Level 3 (A‑Level) and Level 4+ (higher education) qualifications, and employment density in information‑communication, finance, and professional occupations. The models are trained on an 80 %/20 % train‑test split, hyper‑parameter tuned via iterative optimisation, and evaluated using R² and RMSE. To interpret the black‑box predictions, SHapley Additive exPlanations (SHAP) are calculated, yielding marginal contribution values for each predictor.

Second, the authors forecast sector‑wide trajectories (total AI entities, active entities, dissolved entities, and dissolution ratio) for 2025‑2030 using four univariate time‑series techniques – ARIMA, Theta, Exponential Smoothing (ETS), and Median Fourier Linear Exponential Smoothing (MFLES). Hyper‑parameters are optimised with a Tree‑structured Parzen Estimator (TPE), and model selection is based on RMSE and MAE on a validation window (2020‑2024). Conformal prediction is applied to generate 95 % prediction intervals, providing distribution‑free coverage guarantees.

Key Findings

1. Geographic Concentration – London hosts 41.3 % of all AI entities in 2024, and accounts for 39.5 % of new registrations between 2020‑2024. The remaining firms are dispersed across England, with a notable cluster in the South‑East.

2. Sectoral Revenue Patterns – General financial services firms (n = 33) exhibit the highest mean operating revenue (£33.9 million), followed by computer software and IT services. Cumulative lifetime revenue for firms with ≥10 years of operation totals approximately £284.7 million, with sector‑specific contributions ranging from £13.2 million (software) to £1.5 million (financial services).

3. Drivers of Firm Performance – The CatBoost models identify firm size (employee count) and AI‑specialisation intensity as the dominant predictors of operating revenue. Larger firms (20‑200 employees) generate disproportionately higher revenues, while firms with higher TF‑IDF‑derived specialisation scores see revenue growth rates 15‑20 % above the baseline.

4. Local Socio‑Economic Effects – Postcode‑level Level 3 qualification rates, population density, and employment density in high‑skill sectors contribute positively, albeit modestly, to revenue. A one‑percentage‑point increase in Level 3 qualification prevalence correlates with a 5‑7 % uplift in average firm revenue, suggesting that regional human‑capital endowments lower recruitment costs and foster knowledge spillovers.

5. Sector Evolution and Outlook – Registrations accelerated sharply until 2020 (peak new registrations 1,607 in 2020‑2024), then dipped by 7.3 % in the most recent year. Dissolution ratios rose from 0.85 % (2020) to 2.29 % (2024). Forecasts indicate total AI entities will reach 4,651 by 2030 (95 % CI 4,323‑4,979), while the dissolution ratio is projected to increase to 2.21 % (95 % CI –0.17 % to 4.60 %). This suggests a transition from rapid expansion to slower growth accompanied by consolidation.

Policy Implications

• Regional Balance: To mitigate systemic risk from London‑centric clustering, targeted investments in AI infrastructure, talent development, and incubator programs should be directed to high‑potential regions with favorable Level 3 qualification rates and skilled‑labor density.

• Differentiated Support: Capital‑gap interventions (e.g., venture financing, loan guarantees) are most effective for early‑stage SMEs, whereas firms that have achieved scale should receive incentives for deepening technical specialisation (R&D tax credits, collaborative research grants).

• Managed Consolidation: The rising dissolution ratio need not be viewed solely as a negative signal; it can reflect market‑driven pruning of low‑productivity firms. Policymakers should therefore facilitate smooth M&A processes, technology transfer mechanisms, and “smart cluster” governance to preserve knowledge assets during consolidation.

Conclusion
The study demonstrates that the UK AI economy is characterised by strong geographic concentration, sectoral heterogeneity, and a clear hierarchy of performance drivers: firm size and AI‑specialisation dominate, while local human‑capital and economic density provide meaningful but secondary boosts. Forecasts point to a maturing sector where growth slows and consolidation intensifies. Effective policy must therefore blend place‑sensitive ecosystem building with nuanced financial and technical support, ensuring that the UK not only expands its AI output but also distributes its benefits across regions, turning consolidation into a source of sustained competitive advantage.