Understanding Interfirm AI Talent Flow Networks through Online Professional Profiles

1 Introduction

• The study addresses limited attention to interfirm AI talent pathways by constructing annual mobility networks and testing whether network position predicts outcomes beyond workforce scale.AI roles are identified from job text, and firm transitions are tracked across the United States, Europe, and Southeast Asia from 2010–2022.Source: §1 · p. 2

2.1 Constructing and characterizing the AI talent mobility network

• Position-level AI labels and career transitions form directed, weighted firm networks; in the U.S. projection, highly connected firms link many less-connected partners.Connectivity was uneven across 2,255 firms, and clustering and mean neighbour degree declined with degree.Source: §2.1 · p. 3

2.2 AI talent mobility exhibits distinct regional sectoral patterns

• Regional central firms differ: technology dominates the U.S. ranking, Europe is more industrially diverse, and Southeast Asia combines technology and finance.AI flows also exceeded non-AI destination-share baselines into technology in the United States, into technology and business services in Europe, and into business services and technology in Southeast Asia.Source: §2.2 · pp. 3–6

2.3 AI talent inflows exhibit growing inequality, cumulative advantage, and a dynamic core

• AI inflows became more concentrated over time, although their distribution remained less concentrated than non-AI inflows in 2022.The pooled AI Gini rose from 0.602 in 2010 to 0.785 in 2022; the top decile received 73.0% of AI inflows in 2022, up from 55.8% in 2010.Source: §2.3 · p. 6
• AI hiring showed stronger preferential attachment than non-AI hiring, while AI core firms had lower retention and more upward entry across all three markets.AI talent-volume exponents exceeded one in every market; AI core retention ranged from 32.0% to 72.6%, versus 79.8% to 91.0% for non-AI cores.Source: §2.3 · p. 7

2.4 Network position carries valuation information beyond AI workforce size

• A one-standard-deviation increase in standardized log AI PageRank was associated with 0.023 log points higher next-year enterprise value, conditional on workforce and firm controls.The estimate uses 20,413 firm-years from 3,329 firms; next-year EBITDA was also positively associated, but less precisely estimated.Source: §2.4 · p. 7

2.5 Valuation and firm activity rise after large improvements in network position

• After network jumps, AI employment and revenue increased, but earnings gains and reductions in costs relative to revenue were not clearly detected.R&D spending increased among firms reporting it; the cost-ratio findings were not significant.Source: §2.5 · pp. 8–10
• 0.100 log points higher enterprise value followed large AI PageRank jumps relative to matched firms, with no detectable pre-event valuation divergence.The pooled post-event estimate was 0.100 log points (s.e. 0.048, P = 0.036).Source: §2.5 · p. 8

3 Discussion

• AI talent circulation adds a relational dimension to organizational AI advantage: firms differ in network position as well as workforce size.Across the United States, Europe and Southeast Asia, concentration coexists with turnover among central firms.Source: §3 · p. 10
• Strong preferential attachment concentrates AI inflows, yet the network core is less persistent than its non-AI counterpart, leaving room for firms outside it to improve their positions.Firms with larger prior inflows tended to attract disproportionately more subsequent talent.Source: §3 · p. 10
• The industries occupying central positions differ by market: Big Tech dominates in the United States, Europe is broader, and Southeast Asia centers on technology and finance.The authors interpret these differences as consistent with AI talent circulation being embedded in existing regional structures.Source: §3 · p. 10
• AI network position was associated with subsequent enterprise value after accounting for workforce size and financial characteristics, and network advances preceded value gains relative to peers with similar pre-event trajectories.The event-study comparison is described as reinforcing the valuation association.Source: §3 · p. 10
• Knowledge carried by mobile workers is one possible source of the association, but centrality may also signal employer attractiveness, reputation or expected growth.Weighted PageRank can indicate the attractiveness of destinations receiving workers from important nodes.Source: §3 · p. 10
• The valuation association was clearer than the EBITDA association; event-study results also linked network advances to higher revenue and greater R&D spending among reporting firms, without a detectable cost-ratio reduction.These patterns are described as compatible with capability building and business expansion.Source: §3 · p. 11
• For managers, recruitment origins may reveal capabilities missed by headcount; for policymakers, concentration raises questions about access, while core turnover means concentration alone misses who holds advantageous positions.The authors also point to continued turnover as relevant to monitoring network advantage.Source: §3 · p. 11
• Interpretations are bounded by text-based AI-role identification, uneven profile coverage, listed firms and within-market flows in three markets; strategy and prospects may jointly affect network position and performance.Financial outcomes do not directly measure knowledge transfer, productivity or innovation, and the study period largely predates widespread generative AI deployment.Source: §3 · p. 11

4.1 Data and sample; 4.2 Identifying AI human capital

• The sample draws on about 535 million employment records, links employers to ultimate-parent firms and matches listed firms to Orbis financials; analysis covers 2010–2022 and three headquarters-based markets.The panel covers 23,596 listed firms, with 15,391 matched to Orbis; the enterprise-value specification uses 20,413 firm-years from 3,329 firms. AI roles are classified using a curated 213-term, six-category dictionary with GPT-4o-assisted assignments and human review.Source: §4.1 · pp. 11–12, §4.2 · p. 12

4.3 AI talent mobility networks

• A transition is assigned by the AI classification of the source position, dated at the destination position’s start year, and aggregated into annual directed, weighted firm networks.Networks are built separately by market and year; the AI and non-AI transition sets are mutually exclusive.Source: §4.3 · p. 12

4.4 Network position measures

• Weighted PageRank is the focal position measure, complemented by degree, strength, net flow, closeness, betweenness and eigenvector centrality.PageRank scores firms highly when inflows come from firms that are themselves central.Source: §4.4 · p. 12

4.5 Sectoral mixing residuals

• Sectoral mixing residuals compare observed AI flows with expected flows that preserve each source sector’s AI outflow but use the non-AI destination distribution.Positive residuals indicate sector pairs where AI mobility exceeds the non-AI-based expectation.Source: §4.5 · p. 13

4.6 Concentration and preferential attachment

• Concentration uses annual in-strength, while core status is the market-year top decile of PageRank; the study measures core retention, upward mobility and preferential attachment.Preferential attachment is estimated separately by market and network type using regularized PAFit, with α > 1 indicating super-linear attachment.Source: §4.6 · p. 13

4.7 Panel regressions

• The outcome models relate next-year enterprise value or EBITDA to AI PageRank, AI and non-AI employment, assets and cash over assets, with firm and year fixed effects.PageRank is logged and standardized within market-year; standard errors are clustered by firm.Source: §4.7 · p. 13

4.8 Centrality-jump event study

• The event study follows firms after their first qualifying AI PageRank-rank jump during 2013–2019 and matches them to non-jump firms in the same market, event year and industry.Matching accounts for prior PageRank trajectories, employment, assets and enterprise value; the model traces outcomes from three years before to up to five years after the event.Source: §4.8 · p. 14

Supplementary information.

• The supplementary materials provide classification and network-measure details, full regression tables, and heterogeneity and robustness analyses.Robustness checks vary fixed effects, samples, end year, centrality coding, controls and clustering.Source: Supplementary information · p. 14, §4.7 · p. 14

Data Availability.

• Study data were obtained primarily from Revelio Labs and Moody’s through institutional WRDS subscriptions, subject to provider licensing terms.Access is governed by the respective data providers’ terms and conditions.Source: Data Availability · p. 14

Code Availability.

• The analysis used Python and R, and reproduction scripts are available from the corresponding author upon reasonable request.Source: Code Availability · p. 15

Author Contributions Statement.

• The statement assigns writing, methodology, analysis, data curation and conceptualization roles across the named authors.Source: Author Contributions Statement · p. 15

Competing Interests Statement.

• The authors declare no competing interests.Source: Competing Interests Statement · p. 15

• Across 2010–2022, AI talent inflows became more concentrated and showed stronger preferential attachment than general labor mobility, while central positions changed hands more often than in non-AI networks.The network is reconstructed from approximately 535 million employment records across 58 countries.Source: Supplementary Information:… · p. 18
• AI network centrality was associated with higher enterprise value after accounting for workforce scale, assets, and fixed effects.The abstract also reports that sharp network-position improvements were followed by higher enterprise value relative to matched firms with similar pre-event trajectories.Source: Supplementary Information:… · p. 18
• The findings suggest that corporate AI advantage depends on firms’ positions in talent circulation as well as the amount of talent they accumulate.Source: Supplementary Information:… · p. 18

1 Literature review

• Most prior studies treat AI as an internal stock, while this study addresses limited attention to the pathways through which AI capability circulates across firms.The review organizes the contribution across AI adoption, knowledge diffusion, labor-flow networks, and technological concentration.Source: §1 · pp. 18–19

1.1 AI adoption and firm performance

• Firm-level AI research links AI workers, patents, and products with performance gains, but generally frames AI capability as accumulating within firm boundaries.The paper instead examines firms’ positions in interfirm mobility networks as another dimension of economic exposure to AI.Source: §1.1 · p. 19

1.2 Labor mobility and knowledge diffusion

• Labor-mobility research describes tacit-knowledge transfer and learning by hiring, motivating attention to hiring relationships alongside AI workforce size.The study constructs AI talent mobility networks to test whether network position provides information about subsequent performance beyond workforce scale.Source: §1.2 · p. 19

1.3 Inter-firm networks and firm performance

• Research on interfirm networks links centrality with access to information and resources, and labor-flow studies emphasize workers’ embodied knowledge.The paper applies network centrality measures to AI hiring networks and examines their association with performance beyond AI workforce scale.Source: §1.3 · p. 19

1.4 Concentration and technology-driven inequality

• The review connects technological concentration with leading firms and networks; the study points to concentrated AI inflows and central positions as a possible network-based channel of firm-level inequality.Source: §1.4 · pp. 19–20

2.1 Translation and language detection

• For records not predominantly in English, job titles were language-detected and non-English titles and descriptions were translated to English; translated titles still detected as non-English were excluded from keyword classification.Source: §2.1 · p. 21

2.2 Markets and final sample

• The 2010–2022 network panel covers 23,596 listed firms across the United States, Europe, and Southeast Asia, with cross-market flows excluded.Firms are assigned to markets by headquarters location; the enterprise-value specification uses 20,413 firm-years from 3,329 firms.Source: §2.2 · p. 21

• The six-category taxonomy distinguishes core technical fields from task- and product-based roles, and positions may receive multiple category assignments.Tables 1–3 cover category definitions and the 213-term dictionary.Source: §3 · p. 21, Table 1 · p. 22, Table 2 · p. 22
• A position is classified as AI-related when available job text matches at least one dictionary term, and the overall binary measure counts it once even if it matches multiple categories.Source: §3 · p. 21

4.1 Structural characterization of the 2022 United States AI network

• The 2022 US projection included 2,255 edge-observed firms, and its most connected firm had 711 distinct partners.Clustering declined with degree (slope −0.30129), while degree assortativity was −0.16607.Source: §4.1 · p. 23
• Adding an exponential cutoff did not significantly improve the pure power-law fit to the selected US degree tail.The selected threshold was kmin = 34, retaining 159 firms (7.05% of the graph); the comparison is conditional on that threshold.Source: §4.1 · p. 23

4.2 Network Dynamics

• By 2022, the bottom 80% of firms received approximately 14.6% of AI inflows in the United States, 26.2% in Europe and 44.2% in Southeast Asia.Lorenz curves move further below equality across the displayed years; Southeast Asia’s eight positive-inflow firms began at equality in 2010.Source: §4.2 · pp. 23–24, Figure 1 · p. 24

4.3 Computation of network measures

• Measures use the full annual networks; inward closeness reflects incoming reachability, while high-betweenness firms bridge otherwise weakly connected parts of the flow network.Betweenness is computed exactly, and eigenvector centrality gives high scores to firms receiving talent from other high-scoring firms.Source: §4.3 · p. 24, Closeness centrality · p. 25, Betweenness centrality · p. 25

4.5 Cross-Market Comparison

• Closeness and eigenvector centrality indicate core embeddedness, whereas betweenness captures brokerage; firms ranking highly on one measure need not rank highly on another.Path-based measures use inverse-weight distances so stronger flows correspond to shorter distances.Source: §4.3 · p. 25, §4.5 · p. 26
• US closeness and eigenvector rankings are dominated by major technology firms, Europe’s leading firms span finance, services and manufacturing, and Southeast Asia’s include platforms, finance and telecommunications.In the United States, betweenness rankings are more sectorally varied than closeness and eigenvector rankings.Source: §4.5 · pp. 26–27

4.6 Preferential Attachment Estimation

• AI attachment estimates were 1.0458 in the United States, 1.0350 in Europe and 1.2393 in Southeast Asia, while non-AI estimates were below one in all three markets.AI confidence intervals lie above one and non-AI intervals below one; the estimates use the regularized PAFit estimator.Source: §4.6 · p. 29

5.1 Descriptive CCDF analysis

• The analysis fits ordinary least-squares lines to log-transformed CCDFs, weighting each distinct positive inflow value equally when at least two values are observed.The slope and R2 summarize the displayed distribution.Source: §5.1 · p. 29
• The US fitted slope changed from −1.28 in 2010 to −1.01 in 2022, and the European slope from −1.74 to −1.34, consistent with relatively greater representation of large inflows.Across the eight US and European panels, R2 ranged from 0.939 to 0.988; the lowest value was Europe in 2018.Source: §5.1 · p. 29, Figure 5 · p. 30
• Southeast Asia’s 2010 distribution had eight positive-inflow firms but only one distinct value, while the 2014 fit used two values and had R2 = 1.The 2014 R2 therefore provides little evidence about distributional form.Source: §5.1 · p. 29, Figure 5 · p. 30

6 Descriptive Analysis

• AI adopters averaged 7,370.0 million in enterprise value and 11,749 employees, compared with 481.3 million and 723 employees among non-adopters.Adopters also had higher turnover and R&D investment; the reported financial and firm-characteristic differences were statistically significant at p < 0.001.Source: §6 · p. 30, Table 5 · p. 31
• AI adopters received an average of 1.65 AI transitions per firm-year, versus 0.01 for non-adopters, and accounted for 37,553 of 38,314 AI-network firm-years.AI flows are classified by the source position, so a firm without AI-related positions can still receive them by hiring former AI workers into other roles.Source: §6 · pp. 30–31
• Among firm-years in the AI network, mean PageRank did not differ significantly between adopters and non-adopters (p = 0.10).The comparison included only 761 non-adopter firm-years.Source: §6 · p. 31

7 Heterogeneity Analysis

• Subgroup PageRank point estimates were positive in every partition and largest among technology firms (0.047) and in Southeast Asia (0.060).The Southeast Asia estimate comes from a small sample of 405 firm-years.Source: §7 · p. 32, Table 6 · p. 33
• Interaction tests did not reject equality of PageRank coefficients by technology status, market, firm size or period.The reported tests were technology versus non-technology (P = 0.12), across markets (P = 0.76), large versus small firms (P = 0.40), and early versus late periods (P = 0.69).Source: §7 · p. 32, Table 6 · p. 33

Alternative fixed effects and lagged outcome

• 0.021–0.023 (P ≤ 0.005) under country- or industry-by-year fixed effects; with a lagged outcome, the coefficient falls from 0.017 to 0.012 (P = 0.090).The lagged-outcome specification is reported as a sensitivity analysis because lagged dependent variables can introduce bias in short fixed-effects panels.Source: Table 7 · p. 34, §8 · p. 34, Table 8 · p. 35

Sample and period sensitivity

• 0.018–0.027 (P ≤ 0.018) across five sample restrictions, while varying the predictor end year from 2019 to 2022 yields 0.016–0.023 (P ≤ 0.029).Restrictions include trimming firms by assets, requiring at least five panel years, retaining AI-employing firms, excluding exceptionally high-PageRank firms, and excluding Southeast Asia.Source: §8 · p. 34, Table 9 · p. 35, Table 10 · p. 36

Alternative centrality measures

• 0.050 (P = 0.010) for PageRank percentile rank; in-strength, closeness, and eigenvector centrality are positive, whereas betweenness is −0.060 (P = 0.009).The pattern is consistent with valuation tracking proximity to the network core rather than brokerage between otherwise disconnected parts.Source: §8 · p. 34, Table 11 · p. 36

Controls and clustering

• 0.026–0.027 when cash holdings are dropped or leverage and R&D intensity are added; the coefficient remains significant at 5% under industry and firm-year clustering.In control-set and clustering comparisons, columns share a sample; sample-restriction and end-year checks change it by design.Source: §8 · p. 34, Table 12 · p. 37, Table 13 · p. 37

Event definition.

• A qualifying event is a firm's first within-panel rank jump above the 2013–2019 90th-percentile threshold, without a sharp preceding decline.The threshold is 0.1127 rank units; 139 of 246 eligible events were matched.Source: Event definition · p. 38

Matching.

• 139 of 246 eligible events found a match: controls were exact-matched on market, event year, and industry, then selected by distance on pre-event characteristics.No Southeast Asian event had an eligible control within the industry and rank constraints.Source: Matching · p. 38

Estimation.

• Estimates span relative years −3 to 5, omit year −1, and include firm-by-event and event-by-relative-year fixed effects; years 4–5 are observed for fewer events.The specification also estimates a pooled post-event association separately; that coefficient is not an arithmetic average of dynamic estimates.Source: Estimation · p. 38

Enterprise value.

• 0.100 log points (s.e. 0.048, P = 0.036) is the pooled post-event enterprise-value estimate; years 2, 4, and 5 are significant after near-zero pre-event coefficients (joint P = 0.75).The event-year estimate is 0.015 (95% CI −0.061 to 0.092); the years 1–5 estimates are 0.103, 0.142, 0.112, 0.181, and 0.233 log points.Source: Enterprise value · pp. 38–39

Comparison with the panel estimates.

• 0.87 standard deviations is the event-year PageRank jump; its implied next-year enterprise-value difference is about 0.02 log points, versus event-study estimates of 0.10–0.23 in years 1–5.PageRank remains elevated by 0.29 to 0.44 standard deviations in years 1–5, with a pooled post-event difference of 0.45.Source: Comparison with the panel estimates · p. 39

Sensitivity.

• Enterprise-value and operating-revenue estimates remain positive across alternative comparison groups and baseline adjustments, while gross-profit and EBITDA evidence remains less conclusive.The alternatives include different numbers of never-treated controls and not-yet-treated firms as controls.Source: Sensitivity · p. 39, Table 15 · p. 41

Income statement and costs.

• 0.084 log points for operating revenue (P = 0.039) and 0.143 for R&D expenditure (P = 0.008), while cost-to-turnover ratios show no detectable post-event differences.Gross profit and EBITDA estimates are positive but imprecise; operating and net profit are also imprecise and observed for fewer events.Source: Income statement and costs · p. 39, Table 16 · p. 42