Automation and substitution

Whether AI substitutes for inputs (labor, capital) or augments them — and what scales, what shrinks.

2026-kim-mapping-ai-into-production finds that treated firms grew output (tasks, customers, revenue) while labor demand stayed flat and capital demand fell ~39.5% — a “do more with less” pattern where AI substituted for external capital rather than for headcount, at least in young firms over a short window. Founders described AI as doing the “analyst or intern version” of tasks.

2026-ermakov-designing-post-role-enterprise frames automation as a design-boundary problem, not simple task replacement. Some workflow nodes can be handed to agents, but others are “Human-Critical Nodes” where ethical judgment, relational trust, irreversible consequences, or visible accountability require human presence. The design problem is which workflow parts can be delegated to digital labor, which require human authority, and how accountability is assigned before failure.

2026-dellacqua-jagged-technological-frontier shows task substitution is jagged, not linear. In a BCG consultant experiment, GPT-4 assistance improved productivity and quality for tasks inside the frontier but reduced correctness on a task outside it. The substitution question is therefore not “can AI do consulting work?” but which task components are currently inside the model’s capability frontier and how that frontier is detected.

2026-dellacqua-cybernetic-teammate demonstrates that substitution can occur at the team-function level before it shows up as job-level replacement. In a P&G field experiment, individuals with AI matched the quality of non-AI teams and produced more functionally balanced ideas, suggesting AI can substitute for some short-run benefits of cross-functional collaboration in bounded ideation work.

2025-dellacqua-super-mario-ai shows that substitution can fail at the team-system level even when the automated agent is individually high-performing. In a laboratory team game, replacing one human with AI lowered firm performance, increased coordination failures, and reduced trust and effort; capability at the replaced task was not enough to preserve the routines and motivation that made the human team work.

2025-brynjolfsson-generative-ai-at-work reports that GenAI augmented customer-support agents rather than replacing them in the observed rollout. Productivity rose 15% on average, especially for lower-skill and less experienced workers, but the paper cannot observe wages, overall labor demand, or hiring composition. It is evidence for within-job task augmentation, not direct labor substitution.

2026-tomei-what-jobs-can-ai-learn reframes exposure around what AI systems may be able to learn, not only what current LLMs can already do. Its RL Feasibility Index finds that some monitoring and control occupations score higher on future automation feasibility than text-centered exposure indices suggest, because the relevant tasks have verifiable outcomes and simulable environments.

2023-nyc-automated-employment-decision-tools-faq brings the substitution question into employment decisions. Covered AEDTs are defined around tools that substantially assist or replace discretionary decision-making, and screening counts as an employment decision. algorithmic-assessment can therefore substitute for parts of human judgment before the final hiring or promotion decision is made.

2022-dellacqua-falling-asleep-at-wheel shows substitution can happen behaviorally before it is formalized. Recruiters remained the decision-makers, but higher-quality AI advice made them more likely to follow recommendations and lower-quality AI induced more effort. The substitution boundary is therefore partly psychological and workflow-shaped, not only a question of whether the human is still “in the loop.”

2025-bastani-human-ai-contracting-paradox shows substitution can also follow from contracting failure rather than technical superiority. In the model, a human-AI team can be best for total system value, yet the principal may prefer AI-only operation or forbid AI because motivating human oversight becomes too costly when failures are rare and effort is hard to observe.

2023-landers-machine-learning-psychometric-assessment implies that algorithmic substitution should be evaluated at the level of assessment-design components: combining scale composites, scoring item responses, dropping predictors, or handling novel data formats. Replacing OLS with machine learning is not the same thing as improving the selection system.

2025-mazeika-remote-labor-index adds a direct end-to-end automation benchmark for remote freelance work. Across 240 real commissioned projects, current agents completed at most 2.5% of tasks at a reasonable-client acceptance standard, despite the same models often looking strong on narrower knowledge, coding, or computer-use evaluations. This is evidence against treating benchmark progress as realized labor substitution without checking whether the whole deliverable can actually be completed.

2025-shukla-ai-assisted-design-ironies brings in the substitution myth from human-automation research. The paper argues that assigning UX tasks to AI based on a simple “machines are better at” logic misses how automation changes the human role: designers may do less direct drafting but more monitoring, validation, troubleshooting, and accountability work.

2025-handa-economic-tasks-ai estimates that 43% of observed Claude.ai usage looked automation-like and 57% looked augmentation-like. That makes automation and augmentation empirical usage patterns rather than mutually exclusive futures: both appear in current task-level AI use, often before there is evidence about employment, wages, or durable substitution.

2025-marguerit-augmenting-automating-labor adds labor-market outcome evidence for keeping automation and augmentation separate. Using Stack Overflow-derived exposure measures for U.S. occupations and industries from 2015–2022, the paper finds automation AI exposure reduced hourly wages in aggregate 2SLS models and reduced new work, employment, and wages for low-skilled occupations, while augmentation AI exposure increased new work and employment overall and raised wages for high-skilled occupations.

2025-agarwal-what-work-ai-doing suggests current conversational GenAI use is more concentrated in dynamic, cognitively demanding, creative, non-routine tasks than in standardized operational tasks. That complicates the old routine-task automation story: low conversational use for standardized work may mean those tasks are already automated through other systems, better suited to API/workflow automation, or simply less attractive for current LLM chat use.

2025-tomlinson-working-with-ai estimates AI applicability from observed Copilot conversations, but explicitly warns against converting high AI-action applicability into automation claims or high user-goal applicability into augmentation/wage claims. Applicability indicates where AI may change work activities; employment and wage effects depend on task recomposition, business decisions, and organizational design.

2025-shao-future-work-ai-agents adds a worker-desire constraint to substitution claims. Workers wanted automation for 46.1% of audited tasks, especially low-value and repetitive work, but the Human Agency Scale results show that many occupations prefer collaboration or sustained human involvement rather than removal. Substitution potential therefore has to be read against desired agency, not only technical capability.

2026-doshi-human-ai-task-tensor clarifies why augmentation/substitution is not a binary label. Its Task Augmentation/Automation Scale breaks decision-making authority into 20 levels, and the broader tensor asks whether output definition, audit needs, AI contribution, AI structure, and human persona differ across tasks. Substitution claims therefore need dimensional task evidence before they are converted into labor-demand or human-agency conclusions.

2026-zhu-choose-your-agent gives early behavioral evidence that autonomous execution can outperform advisory or coaching modes in a strategic group task, but only under a specific authority arrangement. The Delegate condition substituted for human proposal/action at each turn and improved group surplus more than Advisor or Coach access, yet participants still preferred Advisor control. This is substitution at the decision-turn level, not proof of job-level labor substitution.

2026-zhang-fatigue-aware-learning-defer adds a fatigue-aware routing case. In learning-to-defer, substitution is not a fixed property of the task: a system may route a hard case to a human when the expert is fresh and to AI later when accumulated workload predicts degraded human performance. That makes substitution contingent on human state and task sequence, not just AI capability.

2026-massenkoff-labor-market-impacts-ai adds early labor-market evidence for treating substitution as observed exposure plus outcome movement, not capability alone. The report weights automated, work-related Claude use more heavily than augmentative use, finds that observed exposure is still far below theoretical exposure, and reports no systematic unemployment increase in highly exposed occupations since late 2022, with only suggestive evidence of slower young-worker entry into exposed jobs.

2026-richmond-ai-jobs-transition-framework adds a demand-side and human-necessity filter to substitution claims. The report argues that highly exposed jobs can sort into automation risk, reorganization, or AI-driven growth depending on whether human involvement remains necessary and whether lower effective cost expands demand enough to offset labor savings; in the Drive version, 18% of jobs are classified as higher short-term automation risk, 24% as reorganizing, and 12% as potential growth jobs.

2026-wang-agent-development-real-world-work adds a benchmark-representativeness warning. If agent benchmarks overrepresent programming-heavy, easily specified, and easily verified tasks, then rising benchmark performance is weak evidence for broad labor substitution until the evaluated tasks resemble the domains, skills, complexity, and verification conditions of real work.

2025-friis-performance-principle adds a social-feasibility boundary to substitution claims. In a U.S. survey across 940 occupations, support for full automation nearly doubled when AI was described as outperforming humans at lower cost, but roughly 12% of occupations remained morally repugnant for AI use. Substitution forecasts therefore need moral-boundaries-of-ai-automation alongside capability, cost, worker desire, and realized adoption evidence.

2025-landers-ethical-use-ai-iop adds an ethical deployment boundary to replacement claims. Worker replacement is not settled by technical feasibility, cost, or owner rights under capitalism: deployment decisions still need to consider worker dignity, rights, reskilling, reactions, and the organizational duty to monitor harms when AI changes or eliminates roles. This makes responsible-ai-deployment a governance layer for substitution decisions rather than an after-the-fact sentiment check.

2026-rodriguez-when-not-to-automate turns that governance boundary into a proposed pre-decision protocol. Its PHP-AIO framework asks decision makers to score tacit-knowledge erosion, resilience reduction, regulatory exposure, and socio-institutional capital degradation before committing to automate, then returns automate, augment, hybrid, or preserve. The useful contribution is the decision timing: capability, cost, exposure, and even post-hoc compliance are incomplete if the organization has not asked which human capability, redundancy, and recovery capacity a workflow would remove. The formal scores are a conceptual proposal, not validated evidence that these gates predict better automation outcomes.

Connections

• Relates to human-ai-task-allocation because substitution is often decided at the routing level — which task, which turn, which human state — rather than at the job level, and allocation policies can move the boundary within a single work episode.
• Relates to human-ai-task-taxonomy because substitution claims need dimensional task evidence — decision authority, audit requirements, output definition — before they can be converted into labor-demand conclusions.
• Relates to moral-boundaries-of-ai-automation because a share of occupations remains morally off-limits for automation even when AI outperforms humans at lower cost, so social feasibility bounds technical substitution.
• Relates to automation-complacency because substitution can happen behaviorally — through reliance and effort migration — while the human formally remains the decision-maker.
• Relates to agentic-delegation because delegation modes determine whether autonomous execution substitutes for human action turn-by-turn, and users may refuse the mode that performs best.
• Relates to ai-workforce-impact-measurement because the gap between theoretical exposure, observed use, and realized labor-market movement is a measurement problem before it is a substitution fact.
• Relates to work-redesign because automating a task often recomposes the job — new monitoring, validation, and accountability work — rather than removing the worker.
• Relates to responsible-ai-deployment because replacement decisions carry governance obligations — worker dignity, reskilling, harm monitoring — that are not settled by technical feasibility or cost.

Contradictions & open questions

• Replacement vs. protected human judgment. Automation can reduce dependence on some human task execution while increasing the importance of human accountability at specific nodes. The wiki needs empirical sources on whether organizations can actually identify and govern these nodes reliably.
• The labor-demand question is still thin. Disagreement on labor substitution should be surfaced here, not averaged away.
• 2026-dellacqua-jagged-technological-frontier is task-performance evidence, not labor-demand evidence. It supports selective task delegation, but not claims about headcount substitution.
• 2026-dellacqua-cybernetic-teammate is team-function evidence, not labor-demand evidence. It shows that AI-supported individuals can match non-AI teams on a bounded product-innovation task, but 2025-dellacqua-super-mario-ai shows the opposite pattern for interdependent coordination work: even a high-performing AI replacement can reduce system output by disrupting coordination, trust, and effort. Substitution claims therefore need to specify which team function is being replaced.
• Short-run augmentation vs. equilibrium substitution. 2025-brynjolfsson-generative-ai-at-work shows higher output per hour inside an existing job, but explicitly cannot answer whether firms later hire fewer agents, hire different agents, change wages, or automate the role more fully.
• Exposure vs. realized substitution. 2026-tomei-what-jobs-can-ai-learn is a forward-looking exposure index, not evidence that listed occupations have already been automated. Its job-posting analysis is suggestive; the stronger contribution is measurement of task learnability.
• Support vs. substitution in hiring. Employers may describe algorithmic assessment as decision support, while regulatory language treats some uses of simplified outputs, rankings, or classifications as substantial assistance or replacement of discretion. The boundary is practical, not merely semantic.
• Formal human-in-the-loop vs. behavioral delegation. 2022-dellacqua-falling-asleep-at-wheel suggests humans can remain formally responsible while attention and effort migrate to the model. That is substitution through reliance, not job removal.
• Task exposure vs. end-to-end deliverables. 2025-mazeika-remote-labor-index cuts against simple exposure narratives: a task can be computer-based, remote, and economically valuable while still resisting autonomous completion because quality, file integrity, multimodal artifacts, and project integration fail at the deliverable level.
• Substitution vs. transformed responsibility. 2025-shukla-ai-assisted-design-ironies suggests that even when AI substitutes for visible task execution, the human may inherit harder-to-see responsibilities for interpretation, correction, rationale, and consequences.
• Conversation automation vs. labor substitution. 2025-handa-economic-tasks-ai can classify some AI conversations as automation-like, but that does not establish job-level substitution. The human may still be selecting the task, evaluating the output, integrating it into a larger workflow, and bearing accountability.
• Exposure evidence with outcome movement. 2025-marguerit-augmenting-automating-labor moves beyond exposure-only claims by linking separate automation and augmentation exposure measures to new work, employment, and wages, but it still infers exposure from developer-question traces rather than directly observing firm adoption or internal task redesign. Its strongest contribution is the directional split: automation exposure and augmentation exposure should not be collapsed into one AI-impact coefficient.
• Routine-task theory vs. conversational GenAI use. 2025-agarwal-what-work-ai-doing finds more use in complex and creative task clusters than in standardized operational tasks. This does not refute routine-task automation theory; it may show that chat-based LLM traces capture a different adoption channel than embedded automation systems.
• Applicability vs. substitution. 2025-tomlinson-working-with-ai strengthens the evidence that AI is useful across information-work activities, but usefulness is not the same as headcount substitution. The paper’s own ATM example is the right caution: automating a task can lower costs, change service models, and recompose jobs rather than simply eliminate workers.
• Capability vs. consent-to-automate. 2025-shao-future-work-ai-agents separates tasks that AI experts believe agents can automate from tasks workers want agents to automate. The Automation Red Light zone matters because labor substitution can be technically plausible while still violating worker preferences, agency, trust, or the meaning of the task.
• Taxonomy vs. evidence. 2026-doshi-human-ai-task-tensor helps describe augmentation and automation arrangements, but classification is not outcome evidence. The wiki still needs studies showing when particular authority/audit/output configurations actually change performance, effort, learning, accountability, or labor demand.
• Static vs. state-contingent substitution. 2026-zhang-fatigue-aware-learning-defer suggests the automation boundary may move within a single work episode as human fatigue accumulates. That is useful for system performance, but it also raises work-design questions about whether organizations should optimize around fatigue, redesign workload upstream, or preserve human agency over routing decisions.
• Observed exposure vs. realized displacement. 2026-massenkoff-labor-market-impacts-ai makes substitution evidence stronger by using observed work-related usage and automation weighting, but it still finds little labor-market movement so far. The plausible early channel may be slower hiring into exposed occupations, especially for young workers, rather than unemployment among incumbent workers.
• Exposure plus demand is still not displacement. 2026-richmond-ai-jobs-transition-framework improves on exposure-only substitution logic by adding human necessity and elasticity, but its archetypes remain pressure categories. They do not show whether firms reduce headcount, redesign roles, expand output, or move work into new occupations.
• Benchmark success vs. real-work substitution. 2026-wang-agent-development-real-world-work implies that substitution claims need a validity check: the benchmark may measure convenient agent tasks rather than the job domains, interpersonal work, ambiguity, and end-to-end accountability that determine whether labor can actually be replaced.
• Contracting failure vs. technical failure. 2025-bastani-human-ai-contracting-paradox implies that organizations may substitute toward AI-only systems or away from AI not because humans add no value, but because the value of human vigilance is hard to contract for. That creates an adoption/substitution mechanism the wiki should keep separate from capability, worker preference, and complacency.
• Performance resistance vs. principle resistance. 2025-friis-performance-principle cuts against the assumption that better AI eventually dissolves all resistance: much public opposition is performance-contingent, but a smaller moral floor remains for care, therapy, spiritual leadership, and other human-presence roles.
• Protocol determinism vs. judgment validity. 2026-rodriguez-when-not-to-automate makes a protocol reproducible once structured inputs are set, but its input judgments, weights, thresholds, and financial-services calibration are expert-elicited and not yet validated against realized knowledge loss, resilience failure, regulatory action, or trust outcomes. A repeatable automate/preserve recommendation is not yet an empirically warranted one.