Generative AI forces a rethink of professional expertise and firm design: existing theories of work and organizations need extension to capture new, evolving interdependencies between professionals and generative machine learning, the paper argues.

Main Finding

The paper argues for a targeted research agenda to theorize how generative machine learning (GML) — especially large language models (LLMs) and related tools — is reshaping professional work and professional service firms (PSFs). Rather than treating GML as a standalone technology, the authors call for studying emergent interdependencies between professionals and GML that require extending and reimagining theories of expertise, professional identity, task allocation, organization and governance.

Key Points

• Distinctive features of GML
  • GML is generative (creates new content) and often built on unsupervised learning and LLMs, not just extractive/predictive ML. Outputs are produced by remixing large corpora and can mimic interpretative/advisory professional outputs.
  • Natural-language interfaces and ease of use democratize access (professionals, paraprofessionals, and lay users).
  • GML exhibits a “jagged frontier” of capabilities: it can unexpectedly reach into some tasks traditionally seen as uniquely human while failing at others; outputs can hallucinate, be biased, and are often opaque.
• Why professions and PSFs are a special case
  • Professional work centers on individually-embedded expertise (diagnosis, inference, treatment). GML challenges how expertise is produced, recognized, and defended (jurisdictions).
  • Professional identity and apprenticeship: mastery of tasks underpins identity and boundary work; GML’s automation/mimicry could disrupt training, identity formation, and career trajectories.
  • Organizing and governance: PSFs rely on collegial/autonomous forms of organizing and low capital intensity; GML may push different governance, business model and investment choices (e.g., centralized platforms, productization).
• Empirical and conceptual uncertainties
  • Adoption is emergent and contextual: risk concerns (privacy, bias, liability), environmental costs, and regulatory responses shape uptake.
  • Productivity effects are uneven: studies suggest GML can raise outputs for less-skilled workers (reducing variance), but it’s unclear which skill-groups will form the strongest human–GML relations (junior, mid-career, or senior).
  • Estimates of automation potential are substantial (papers cited estimate ~40% of professional labour time could be automated), but context matters.
• Research orientation recommended
  • Focus on situated human–machine interdependencies (relations, networks, and interactions), not just technology capabilities.
  • Extend prior literatures on digital/algorithmic/AI tech while foregrounding GML’s particular agency and affordances.
  • Ask contextualized questions about expertise, identity, task allocation, governance, business models, ethics, and wider societal effects.

Data & Methods

• Type of paper: conceptual/theoretical research agenda built from synthesis of existing literatures (digital, algorithmic, AI in professional contexts), not new empirical data.
• Methods used by the authors: literature review and theoretical extension — drawing on studies of earlier technologies (algorithmic systems, extractive/predictive ML), recent work on GML, and organizational/professional theory (e.g., Abbott on jurisdictions, apprenticeship/identity literatures, PSF governance).
• Recommended empirical approaches for future work (implied/suggested):
  • In-depth, contextualized case studies and ethnographies of GML adoption and work practices in PSFs.
  • Comparative and longitudinal field studies to capture evolving interdependencies.
  • Mixed-methods designs including qualitative process tracing, interviews, and quantitative analyses of firm-level productivity, billing, time-use and employment data.
  • Experimental or quasi-experimental designs where feasible (e.g., randomized deployment of GML tools, difference-in-differences exploiting staggered rollouts).
  • Task-based measurement and detailed microdata (task-level, time allocation, output quality) to identify complementarities/substitution and distributional effects.

Implications for AI Economics

• Labor market impacts and task reallocation
  • GML accentuates task-based substitution/complementarity debates: it can automate or augment parts of expert tasks (drafting, summarizing, screening), changing demand for different skill levels within PSFs.
  • Potential to compress productivity dispersion (raises outputs of less-skilled workers) while shifting value toward tasks that remain uniquely human (judgment, complex client relationships). This has implications for wage structure and within-firm skill premia.
  • Apprenticeship and career progression effects: reduced on-the-job training opportunities for junior staff could alter human capital accumulation and long-term wage trajectories.
• Firm behavior, business models and market structure
  • GML may lower marginal costs of producing advisory outputs and enable productization of services; this can change pricing, bundling, and the tradability of professional services.
  • Lower capital intensity barriers may change entry dynamics (paraprofessionals, platform entrants) and could either increase competition or generate winner-take-most effects if firms with proprietary data/models gain advantages.
  • Governance and investment: PSFs may face choices between decentralized adoption (individual use) and centralized, platform-style investments (capturing scope economies, quality control, liability management).
• Measurement and identification challenges for economists
  • Need granular task-level measures, quality-adjusted output metrics, billing and time-use data. Hallucinations and correctness impose measurement noise on productivity gains.
  • Opaque model behavior (black box outputs) complicates attributing changes in outcomes to GML rather than complementary organizational changes.
  • Adoption endogeneity: firms with specific unobserved traits may be early adopters; causal designs should exploit plausibly exogenous variation (pilot rollouts, vendor constraints, regulation changes).
• Policy and market failures to study
  • Externalities: data ownership, privacy, model training environmental costs — these create social costs not internalized by firms.
  • Liability and regulation: professional standards, malpractice risk, and accountability regimes influence adoption and the distribution of gains/losses.
  • Distributional policy responses: upskilling, re-training incentives, and accreditation reforms to manage transitions in professional labor markets.
• Suggested empirical research directions for AI economists
  • Estimate causal effects of GML deployment on productivity, wages, employment composition, and billable rates using firm-panel or administrative data combined with task-level measures.
  • Structural models of task allocation within PSFs to quantify complementarities/substitution, compute welfare and distributional impacts, and simulate policy responses.
  • Market-level analyses of prices, entry/exit, and concentration in professional services pre/post-GML diffusion.
  • Experiments and RCTs on human–GML workflows (cyborg vs centaur vs automated regimes) to identify optimal task divisions and impacts on quality, training, and decision-making.
  • Cost–benefit analysis incorporating environmental and societal externalities of large-scale GML adoption.

Short takeaway for AI economists: GML changes the unit of analysis from “automation of jobs” to evolving, situated interdependencies at the task and relational level. Empirical work should combine micro-level task and quality data with firm- and market-level measures to identify complementarities, distributional effects, and the role of organizational choices and regulation in shaping economic outcomes.