From Clerks to Agentic-AI: How will Technology Change Labor Market in Finance?

Abstract: Financial firms have gone through three major technological waves: computerization in the 1980s and 1990s, the rise of indexing and passive investing in the 2000s and 2010s, and the AI and automation wave from roughly 2015 to the present. This project studies how much labor is required to manage capital across those waves by tracking a simple productivity measure: assets under management per employee. Using a small panel of representative firms, we compare changes in AUM per employee, revenue per employee, and operating expense intensity over time. The goal is not to identify causal effects, but to document stylized facts about how technology changes the scale of asset management work.

• The paper develops a filing-based automation-intensity framework and finds that productivity coefficients rise across computerization, indexing, and AI eras, while causal claims remain limited.
• The paper finds that higher AI disclosure correlates with 58.4% greater AUM per employee but 5.4% lower revenue per employee, indicating reorganization and output-mix changes rather than uniform productivity gains.
• The paper argues that AI will initially reallocate tasks instead of immediately reducing labor costs, strengthening large institutions and small teams while placing middle-layer coordination roles under the greatest pressure.

Overview and Motivation

"From Clerks to Agentic AI: How Will Technology Transform the Labor Market in Finance?" by Lu Yu (Georgetown University) and Xiang Li (Pittsburgh Supercomputing Center, Carnegie Mellon University) examines how successive waves of task automation—computerization, indexing, and agentic AI—reshape labor productivity and labor demand in U.S. financial firms. The paper's central claim is that automation reorganizes workflows, scale, and the distribution of capability across firms rather than simply eliminating jobs, producing a more polarized industry in which large institutions retain structural advantages, small teams become more capable, and middle-layer coordination functions face the greatest pressure.

Finance is a useful setting for this question because it combines standardized workflows, information processing, client service, and judgment-intensive decision making within the same firms, so technology affects tasks unevenly: some activities become cheaper immediately while others remain constrained by supervision, trust, interpretation, and accountability. The authors frame the relevant question not as whether AI reduces headcount, but as how automation changes the internal allocation of tasks and what that implies for productivity, scale, and labor demand.

Three Waves of Automation

The paper organizes technological change into three stylized eras. The computerization era (1985–2000) saw personal computers, spreadsheets, Bloomberg and Reuters terminals, and electronic trading infrastructure transform information processing and risk management without eliminating finance; instead, these technologies changed who could operate effectively and which firms could scale an edge. The indexing era (2000–2015) codified a large set of previously labor-intensive investment and rebalancing tasks through ETFs, benchmark replication, passive allocation, and automated portfolio implementation. The AI/automation era (2015–present), accelerating sharply with generative and agentic systems, differs from both predecessors: unlike computerization, which mainly expanded computation and connectivity, it pushes into cognition-adjacent tasks such as monitoring, summarization, drafting, triage, and workflow coordination; unlike indexing, which standardized portfolio construction in a narrow domain, it has the potential to affect a broader range of occupations inside financial firms.

The authors are explicit that these eras are descriptive and interpretive rather than a stand-alone identification strategy—they provide a historically grounded framework for comparison, not sharp causal partitions of technological change.

Conceptual Framework

The core conceptual variable is task automation intensity: the extent to which routine, codifiable, or repetitive work is shifted from human labor to software, rules, machines, or AI systems. Although the vocabulary changes across eras, the economic concept is stable. The framework yields three testable implications:

• Productivity should rise more in later technology eras than in the pre-computerization benchmark.
• Higher automation intensity should be associated with higher output per worker or greater scale per worker.
• Labor expense per worker may lag output measures if automation operates primarily through reorganization and augmentation rather than immediate labor shedding.

The second implication deserves emphasis: the paper predicts that technology can change the composition of work well before it changes headcount, as firms reassign workers toward supervision, exception handling, and client-facing tasks while software absorbs repetitive work in the background.

Measurement and Data

The empirical setting is a panel of publicly listed U.S. financial firms observed in filing and accounting data, targeting roughly 100 firms with a possible extension to 300 for robustness. In the current draft, however, the AI-specific filing analysis uses only five large institutions: Bank of America, BNY Mellon, JPMorgan Chase, S&P Global, and State Street—a limitation the authors acknowledge directly.

Era-specific automation measures are constructed from annual reports and 10-K filings. Computerization-era language includes computers, terminals, spreadsheets, electronic trading, digitization, workflow systems, and back-office automation; indexing-era vocabulary covers index funds, ETFs, passive management, benchmark replication, systematic rebalancing, and rules-based allocation; AI-era terms include artificial intelligence, machine learning, generative AI, autonomous agents, robotic process automation, workflow automation, and LLMs. The final ai_exposure variable counts direct AI terms at higher weight alongside adjacent automation language, normalized by total filing words.

Productivity outcomes include revenue per employee (the main measure), assets under management per employee (informative for asset managers and custodians where scale expands without proportional headcount growth), and labor expense per employee (a direct but incomplete proxy for near-term labor-cost compression).

Empirical Design

The design proceeds from descriptive cross-era comparisons to a fixed-effects panel specification. The baseline AI-era regression takes the form:

$$\log(1 + \text{outcome}_{it}) = \beta \cdot \text{ai\_exposure}_{it} + \text{firm FE} + \text{year FE} + \text{error}_{it}$$

Firm fixed effects absorb time-invariant firm characteristics and year fixed effects absorb common macro shocks. The authors state plainly that this supports descriptive within-firm comparisons but does not support strong causal interpretation. Heterogeneity analysis—banks versus asset managers, passive-heavy versus active-heavy firms, large versus mid-sized institutions—is flagged as important because automation may affect scale, coordination, and labor demand differently across business models.

Results

Productivity across eras

Descriptive regressions comparing four broad eras show that both revenue per employee and AUM per employee are substantially higher in later eras than in the pre-computerization benchmark, controlling for firm-group fixed effects. Era coefficients rise from 1.4070 (computerization) to 2.0543 (indexing) to 2.3688 (AI) in the revenue specification, and from 1.3638 to 2.4197 to 3.3944 in the AUM specification. Deflating the productivity ratios preserves the pattern, indicating the increase is not an artifact of nominal growth. These coefficients are descriptive, but they suggest productivity growth in finance accumulated through successive waves of technological and organizational change rather than arriving all at once.

AI exposure and operating outcomes

The fixed-effects estimates on AI exposure are mixed and, notably, point in opposite directions across outcomes:

Outcome	Coefficient	Robust SE
Revenue per employee	−0.0535	0.0117
AUM per employee	+0.5843	0.1646
Labor expense per employee	+0.0107	0.0219

Greater AI disclosure intensity is associated with lower revenue intensity per worker but higher AUM per worker. The authors interpret this combination as more consistent with reorganization, reinvestment, and output-mix change than with a single-margin productivity story—an internally somewhat contradictory pattern that they do not resolve definitively.

Labor cost response

The coefficient on labor expense per employee is small and statistically indistinguishable from zero. Firms do not respond to rising automation intensity through immediate, visible labor-cost compression; AI may initially affect scale, task allocation, and work organization before those changes appear in compensation or headcount. Wage-trend evidence from the broader panel is consistent with this reading: labor costs evolve over time but not in lockstep with output-per-worker measures, as expected if automation operates through augmentation, workflow consolidation, and task reallocation.

Adoption heterogeneity

Event-study style figures around each firm's first substantive increase in disclosure intensity reveal staggered rather than synchronized adoption: Bank of America appears earlier, S&P Global accelerates later, JPMorgan and State Street move gradually, and BNY Mellon crosses the threshold very late. This points to heterogeneous organizational adoption rather than a single sector-wide break, reinforcing the paper's polarization thesis: large institutions retain advantages in infrastructure, proprietary data, compliance depth, and execution quality, while firms whose edge depended on labor-heavy coordination face pressure if AI lowers the cost of standardizing information flows.

Interpretation

Taken together, the evidence supports transformation rather than simple replacement. Productivity gains appear before labor-cost reductions, suggesting automation initially operates by changing how work is organized rather than by eliminating jobs. The likely outcome is a polarized industry: large institutions preserve structural advantages; small teams become more capable because AI compresses the cost of research support, monitoring, and operational coverage; and the middle layer faces the greatest pressure because its historical value came from coordinating information flows that increasingly can be standardized or partially automated.

This interpretation aligns with the broader automation literature. Autor's task-based framework holds that technology substitutes for routine tasks while complementing abstract and interpersonal ones; Acemoglu and Restrepo distinguish automation, augmentation, capital deepening, and task creation as separate channels; Brynjolfsson and McAfee emphasize that organizational redesign matters as much as the technology itself. The results fit that logic: automation in finance aligns more clearly with reorganization of scale and workflow than with an immediate decline in labor expense per employee.

Limitations

The authors concede several limitations candidly. First, the filing-based exposure measure captures disclosure intensity as much as underlying operational adoption—firms may differ in how much they discuss AI even when actual implementation differs less. Second, the AI-focused sample contains only five firms, so the results constitute a structured pilot rather than a definitive sector-wide estimate. Third, the fixed-effects design cannot cleanly separate causality from reverse causality or omitted organizational change, and productivity ratios are influenced by business mix, firm strategy, and market conditions. Finally, the three-era classification is stylized and should not be read as a sharp causal partition of technological change. An open question the paper leaves unresolved is why AI exposure correlates negatively with revenue per employee while positively with AUM per employee—the output-mix explanation is plausible but untested.

Conclusion

The paper embeds agentic AI within a longer history of computerization and indexing in finance and contributes a transparent, filing-based automation-intensity measure linked to firm-level operating outcomes. Its descriptive evidence indicates that finance has become progressively more scalable and productive across technology eras, that higher AI exposure is associated with greater AUM per employee but weaker revenue per employee, and that labor expense shows no clear immediate response to automation intensity. Adoption is staggered and organizationally mediated. The overall conclusion is that technology will likely transform financial labor markets through task reallocation, workflow redesign, and uneven gains in capability rather than uniform labor replacement—with the middle layer of the industry facing the greatest pressure.