AI automates contiguous blocks of workflow steps, forcing firms to redefine tasks and jobs and producing nonlinear organizational shifts as AI quality improves. Occupations where AI-suitable steps are dispersed see far less automation, while a step is more likely to be AI-executed when its neighbors are also automated.

Main Finding

AI adoption leads firms to bundle adjacent production steps into contiguous “AI chains,” altering task and job boundaries in ways that can overturn standard comparative-advantage predictions. Improvements in AI quality produce non-linear, threshold-driven reorganizations (a J-curve-like pattern) in task allocation, labor demand, and wages. At the aggregate level, firm-level Leontief production with heterogeneous AI deployment can aggregate to a CES representation with distinct manual and AI-assisted labor inputs. Empirical evidence — linking O*NET tasks, human AI-exposure ratings, GPT-derived workflow orderings, and realized AI execution from Anthropic’s index — supports the model’s three central predictions about chaining, fragmentation, and local complementarities.

Key Points

• Steps vs. tasks vs. jobs:
  • Primitive units are ordered production steps. Firms endogenously group contiguous steps into tasks; tasks are bundled into jobs.
  • In the pre-AI baseline, tasks collapse to single steps; AI enables larger multi-step tasks (chains).
• Three execution modes:
  • Manual: human-only.
  • Augmented: human requests AI and verifies output (human oversight required).
  • Automated: AI executes end-to-end without direct human intervention.
  • Crucial distinction: verification is tied to augmented steps; when multiple adjacent steps are automated and only the final step is augmented, the verification cost is a fixed cost per chain rather than per step.
• Chaining and comparative advantage:
  • Avoiding additional human checkpoints can make it optimal to append neighboring steps to an AI chain even when humans would have comparative advantage on those steps in isolation — comparative-advantage assignment logic can fail.
• Fragmentation:
  • The spatial (sequence) dispersion of AI-suitable steps within a workflow matters. More fragmented exposure (AI-suitable steps scattered) reduces realized AI execution versus concentrated exposure (adjacent automatable steps).
• Short run vs. long run:
  • Short run: fixed job designs and skills — AI reduces step times and raises within-job productivity.
  • Long run: firms can redesign tasks/jobs and reassign skills; discrete reorganization thresholds appear as AI quality improves.
• Non-linear effects and threshold behavior:
  • Continuous improvements in AI quality can produce little change until they cross thresholds that permit longer chains or new job designs, after which reorganization and productivity jumps occur.
• Aggregation:
  • Despite firm-level Leontief structure, cross-firm heterogeneity in AI deployment yields a macro-level CES production with manual labor, AI-assisted labor, and capital as inputs.
• Computation:
  • Short-run optimization over AI deployment for an m-step job solvable via dynamic programming in O(m^2). Long-run joint optimization over job design and AI deployment is polynomial-time solvable up to arbitrarily small error.
• Empirical support (three tests):
• AI-executed steps co-occur in contiguous blocks (chains), not randomly dispersed.
• Conditional on the share of AI-exposed steps, greater dispersion (fragmentation) of exposed steps predicts a lower share of steps actually executed by AI.
• A step is more likely to be AI-executed in an occupation when its neighboring steps are AI-executed there (local complementarities).

Data & Methods

• Theoretical framework:
  • Production modeled as a Leontief technology over an ordered sequence of exogenous steps; tasks are endogenous contiguous blocks; jobs are bundles of tasks assigned to workers.
  • Cost components: manual execution time and skill requirements, augmented execution time/skills, hand-off (verification/coordination) costs between workers, AI success probabilities (affecting chain success).
  • Optimization:
    • Short run: choose which steps are manual, augmented, or automated given fixed job boundaries and skills; dynamic programming yields O(m^2) solution.
    • Long run: jointly choose AI deployment and task/job design; polynomial-time algorithm with controllable approximation error.
  • Aggregate implications derived by allowing firm heterogeneity in AI deployment and applying classic aggregation results to obtain a CES macro-production representation with separate manual and AI-assisted labor inputs.
• Empirical strategy and datasets:
  • Step/Task mapping: O*NET tasks mapped to occupations and ordered workflows using GPT-generated workflow orderings (to recover step sequence).
  • AI exposure measures: human assessments of AI exposure (Eloundou et al., 2024).
  • Realized AI execution: Anthropic’s Economic Index (Handa et al., 2025) records which steps are actually executed by AI versus humans.
  • Key empirical tests:
    • Measure contiguity/co-occurrence of AI-executed steps within workflows and test against random or null patterns.
    • Define a fragmentation metric (dispersion of exposed steps across workflow), regress realized AI execution share on fragmentation controlling for total exposure and other covariates.
    • Compare the execution probability of identical/similar steps across occupations conditional on the AI status of neighboring steps (testing local complementarities).
  • Results: the three predicted patterns hold in the assembled datasets (paper reports statistically significant relationships consistent with the chaining mechanism). The authors control for step-level exposure intensity when identifying sequencing effects.

Implications for AI Economics

• Theory and measurement:
  • Sequencing matters: standard task-based substitution models that ignore order and inter-step complementarities can mispredict which work gets automated. Exposure indices should incorporate workflow order and fragmentation.
  • Local complementarities and fixed verification costs generate discontinuities; small AI quality improvements can produce large reorganizations once thresholds are crossed.
• Labor markets and jobs:
  • AI can reshape the unit of work (tasks and jobs), altering skill demands non-monotonically: automation of chains can reduce demand for some hands-on tasks while increasing demand for monitoring/verification and broader oversight roles.
  • Wage and labor-demand effects may be lumpy and heterogeneous across firms and occupations depending on workflow structure and how clustered automatable steps are.
• Macro modeling:
  • Micro-founded CES with separate manual and AI-assisted labor inputs gives a tractable way to incorporate organizational responses to AI into macro models while preserving substitution/complementarity features.
• Policy and firm strategy:
  • Firms: investing to enable chaining (interfaces, data plumbing, reliability) yields outsized gains when it allows longer AI chains; organizational redesign and retraining may be optimal when AI crosses quality thresholds.
  • Policymakers: prepare for discrete, occupation-specific transitions — policy should focus on measuring workflow structure, supporting re-skilling for oversight/validation roles, and monitoring concentrated disruption where automatable steps cluster.
• Empirical research direction:
  • Future empirical work should use ordered workflow data (not only task exposure counts), investigate firm-level heterogeneity in chain adoption, and quantify thresholds in AI capability that trigger reorganizations.