pith. sign in

arxiv: 2604.16465 · v1 · submitted 2026-04-08 · 💻 cs.AI · econ.GN· q-fin.EC

Healthcare AI for Automation or Allocation? A Transaction Cost Economics Framework

Ari Ercole This is my paper

Pith reviewed 2026-05-10 18:13 UTC · model grok-4.3

classification 💻 cs.AI econ.GNq-fin.EC
keywords transaction cost economicshealthcare AItask analysisO*NET databaseclinician rolescoordination frictionsinformation searchdecision coordination
0
0 comments X

The pith

Clinician roles carry substantially higher transaction-cost intensity than non-clinician roles, driven by information search and decision coordination.

A machine-rendered reading of the paper's core claim, the machinery that carries it, and where it could break.

The paper applies transaction cost economics to map coordination frictions at the task level across healthcare occupations using O*NET data. Each task is assigned to one dominant category of friction and scored for intensity via constrained LLM coding, then aggregated by occupation. Clinicians show markedly higher overall intensity than non-clinicians, mainly from greater needs to search information and coordinate decisions, while the spread of costs inside each occupation stays similar. This matters because it indicates that the structure of coordination work, rather than raw technical difficulty, determines where digital tools can most reduce wasted effort. The result points to uneven returns from AI depending on whether the goal is task automation or better allocation of coordination burdens.

Core claim

Aggregating task statements and frequency weights from the O*NET occupational database after coding each unique task into one of four transaction-cost categories shows clinician roles with substantially higher transaction-cost intensity than non-clinician roles. The difference arises primarily from greater burdens of information search and decision-related coordination. Dispersion of transaction costs within occupations does not differ across clinician and non-clinician groups. These patterns demonstrate systematic heterogeneity in the nature of coordination work and imply that opportunities for digital and AI interventions are shaped less by technical task complexity than by underlying fric

What carries the argument

Transaction-cost intensity score obtained by classifying each O*NET task into one dominant friction category (information search, decision and bargaining, monitoring and enforcement, or adaptation and coordination) via constrained LLM coding, then weighting and summing at the occupation level.

If this is right

  • AI tools focused on automation will likely produce larger productivity gains in non-clinician roles that exhibit lower transaction-cost intensity.
  • Interventions that reduce information search and decision coordination burdens will disproportionately improve output in clinician roles.
  • Uniform deployment of AI across healthcare occupations will underperform because coordination structures differ systematically by role.
  • Similar dispersion of costs within occupations means high-friction tasks exist in every role and can be targeted consistently.

Where Pith is reading between the lines

These are editorial extensions of the paper, not claims the author makes directly.

  • The same task-coding method could be applied to other service sectors to identify where AI yields the largest coordination-cost reductions.
  • AI development priorities in healthcare should emphasize features that lower information search and decision coordination loads for clinicians.
  • Empirical validation through direct observation of coordination activities could refine the occupation-level intensity rankings.
  • Resource allocation for technology training or procurement could be guided by these transaction-cost maps to maximize impact.

Load-bearing premise

The constrained LLM coding of tasks into the four transaction-cost categories accurately and consistently reflects real-world coordination frictions in healthcare work.

What would settle it

Time-motion studies or direct measurement of coordination time and error rates showing that clinicians do not actually spend more time on information search and decision coordination than non-clinicians, or that the LLM-derived scores fail to predict observed frictions.

Figures

Figures reproduced from arXiv: 2604.16465 by Ari Ercole.

Figure 1
Figure 1. Figure 1: Transaction-cost friction across healthcare occupations. Occupations are plotted by mean transaction-cost intensity and within-occupation dispersion across tasks, capturing coordination burden (search, decision, monitoring, adaptation) rather than task complexity; clinician and non-clinician roles occupy distinct coordination regimes, implying uneven digital opportunity. underpins common practices such as … view at source ↗
read the original abstract

Healthcare productivity is shaped not only by clinical complexity but by the costs of coordinating work under uncertainty. Transaction-cost economics offers a theory of these coordination frictions, yet has rarely been operationalised at task level across health occupations. Using task statements and frequency weights from the O*NET occupational database, we characterised healthcare work at task granularity and coded each unique task using a constrained large language model into one dominant transaction-cost category (information search, decision and bargaining, monitoring and enforcement, or adaptation and coordination) together with an overall transaction-cost intensity score. Aggregating to the occupation level, clinician roles exhibited substantially higher transaction-cost intensity than non-clinician roles, driven primarily by greater burdens of information search and decision-related coordination, while dispersion of transaction costs within occupations did not differ. These findings demonstrate systematic heterogeneity in the nature of coordination work across healthcare roles and suggest that the opportunities for digital and AI interventions are unevenly distributed, shaped less by technical task complexity than by underlying coordination structure.

Editorial analysis

A structured set of objections, weighed in public.

Desk editor's note, referee report, simulated authors' rebuttal, and a circularity audit. Tearing a paper down is the easy half of reading it; the pith above is the substance, this is the friction.

Referee Report

2 major / 2 minor

Summary. The paper operationalizes transaction cost economics (TCE) at the task level for healthcare occupations by drawing on O*NET task statements and frequency weights. Each task is classified by a constrained LLM into one of four TCE categories (information search, decision and bargaining, monitoring and enforcement, or adaptation and coordination) and assigned an intensity score. Occupation-level aggregates show that clinician roles have substantially higher TCE intensity than non-clinician roles, driven mainly by information search and decision-related coordination burdens, while within-occupation dispersion does not differ. The authors conclude that opportunities for AI interventions in healthcare are shaped by underlying coordination structure rather than technical complexity alone.

Significance. If the empirical results are robust, the manuscript supplies a scalable, theory-grounded method for mapping coordination frictions across occupations using public data. This could inform targeted AI deployment decisions (automation versus task reallocation) by distinguishing roles with high search/decision costs from those dominated by monitoring or adaptation. The approach is novel in its task-granular application of TCE to healthcare and leverages existing occupational databases without introducing new fitted parameters.

major comments (2)
  1. [Methods] Methods section (task coding procedure): The central empirical claims rest entirely on LLM-generated labels for TCE category and intensity. No validation is reported—no human-expert benchmark, inter-rater reliability statistics, prompt-ablation results, alternative-model comparisons, or sensitivity checks. Because the four categories are interpretive, even modest systematic bias (e.g., over-assignment of clinical judgment to “decision and bargaining”) would directly inflate the clinician/non-clinician gap and the reported driver breakdown.
  2. [Results] Results section (aggregation and statistical reporting): Task-level scores are aggregated to the occupation level, yet the manuscript supplies no details on weighting (e.g., by O*NET frequency), handling of multi-category tasks, or statistical tests for the reported differences. No confidence intervals, robustness checks across coding variants, or dispersion metrics are shown, leaving the “substantially higher” and “did not differ” claims without visible quantitative support.
minor comments (2)
  1. [Abstract] Abstract: The phrase “constrained large language model” is used without specifying the model, prompt constraints, or temperature settings; these details belong in the methods section for reproducibility.
  2. [Methods] The manuscript would benefit from a table or figure showing example task statements and their LLM-assigned categories to illustrate the coding scheme.

Simulated Author's Rebuttal

2 responses · 0 unresolved

We thank the referee for the constructive comments on our manuscript. We agree that the methods and results sections require additional details and validation to fully support the empirical findings. We have revised the paper to address these points and respond to each comment below.

read point-by-point responses

  1. Referee: [Methods] Methods section (task coding procedure): The central empirical claims rest entirely on LLM-generated labels for TCE category and intensity. No validation is reported—no human-expert benchmark, inter-rater reliability statistics, prompt-ablation results, alternative-model comparisons, or sensitivity checks. Because the four categories are interpretive, even modest systematic bias (e.g., over-assignment of clinical judgment to “decision and bargaining”) would directly inflate the clinician/non-clinician gap and the reported driver breakdown.

    Authors: We acknowledge the validity of this concern. The original manuscript did not report validation for the LLM coding procedure. In the revised version, we have added a dedicated subsection on validation, including a human-expert benchmark on a random sample of 200 tasks coded independently by two healthcare professionals (Cohen's kappa = 0.82 for category assignment), prompt sensitivity analyses, and comparisons with an alternative open-source model. These checks indicate low systematic bias and support the reported differences. We have also clarified the prompt constraints used. revision: yes

  2. Referee: [Results] Results section (aggregation and statistical reporting): Task-level scores are aggregated to the occupation level, yet the manuscript supplies no details on weighting (e.g., by O*NET frequency), handling of multi-category tasks, or statistical tests for the reported differences. No confidence intervals, robustness checks across coding variants, or dispersion metrics are shown, leaving the “substantially higher” and “did not differ” claims without visible quantitative support.

    Authors: We agree that more statistical detail is needed. The revised manuscript now specifies that aggregation uses O*NET frequency weights for each task, with multi-category tasks assigned to their highest-intensity category. We report two-sample t-tests for clinician vs. non-clinician differences, including 95% confidence intervals. Within-occupation dispersion is quantified using standard deviation of task-level scores, showing no significant difference. Additional robustness checks varying the LLM temperature and category thresholds are provided in the appendix, confirming the stability of the main results. revision: yes

Circularity Check

0 steps flagged

No circularity: empirical aggregation of external O*NET data via LLM coding

full rationale

The paper's central results are produced by taking public O*NET task statements and frequency weights, applying a single constrained LLM call to assign each task to one of four TCE categories plus an intensity score, then aggregating the resulting labels by occupation. No equations, fitted parameters, or derived quantities are defined in terms of the target results and then reused as inputs or predictions. No self-citations are invoked to justify uniqueness or load-bearing premises. The derivation chain is therefore a straightforward data-processing pipeline whose outputs are not equivalent to its inputs by construction.

Axiom & Free-Parameter Ledger

0 free parameters · 2 axioms · 0 invented entities

The central claim rests on the assumption that LLM-assigned transaction-cost categories validly measure coordination frictions and that O*NET task statements and weights adequately represent actual healthcare work; no free parameters or invented entities are visible from the abstract.

axioms (2)
  • domain assumption LLM coding under constraints produces reliable and unbiased assignment of tasks to transaction-cost categories
    The method depends on this without reported validation or human inter-rater agreement.
  • domain assumption O*NET task statements and frequency weights accurately capture the coordination demands of real healthcare occupations
    Aggregation to occupation level assumes these inputs reflect actual work.

pith-pipeline@v0.9.0 · 5465 in / 1346 out tokens · 37269 ms · 2026-05-10T18:13:42.213616+00:00 · methodology

discussion (0)

Sign in with ORCID, Apple, or X to comment. Anyone can read and Pith papers without signing in.

Reference graph

Works this paper leans on

7 extracted references · 7 canonical work pages

  1. [1]

    Principles for artificial intelligence (ai) and its application in healthcare

    British Medical Association. Principles for artificial intelligence (ai) and its application in healthcare. Report, 2024. Accessed via BMA publication

    work page 2024

  2. [2]

    Raza, Kaushik P

    Marium M. Raza, Kaushik P. Venkatesh, and Joseph C. Kvedar. Generative ai and large language models in health care: pathways to implementation.npj Digital Medicine, 7:62, 2024. doi: 10.1038/s41746-023-00988-4. 7

    work page doi:10.1038/s41746-023-00988-4 2024

  3. [3]

    Abramoff, Noelle Whitestone, Jennifer L

    Michael D. Abramoff, Noelle Whitestone, Jennifer L. Patnaik, et al. Autonomous artificial intelligence increases real-world specialist clinic productivity in a cluster-randomized trial.npj Digital Medicine, 6:184, 2023. doi: 10.1038/s41746-023-00931-7

    work page doi:10.1038/s41746-023-00931-7 2023

  4. [4]

    Ronald H. Coase. The nature of the firm.Economica, 4(16):386–405, 1937. doi: 10.1111/j. 1468-0335.1937.tb00002.x

    work page doi:10.1111/j 1937

  5. [5]

    Williamson.The Economic Institutions of Capitalism: Firms, Markets, Relational Contracting

    Oliver E. Williamson.The Economic Institutions of Capitalism: Firms, Markets, Relational Contracting. Free Press, New York, 1985

    work page 1985

  6. [6]

    North.Institutions, Institutional Change and Economic Performance

    Douglass C. North.Institutions, Institutional Change and Economic Performance. Cambridge University Press, Cambridge, 1990. doi: 10.1017/CBO9780511808678

    work page doi:10.1017/cbo9780511808678 1990

  7. [7]

    O*net database

    O*NET Resource Center. O*net database. https://www.onetcenter.org/database.html. Accessed 2026-03-16. 8

    work page 2026