O*NET Intermediate Work Activities

• O*NET IWAs are empirically coded, occupation-linked behavioral task units that capture granular work activities and serve as a bridge between job descriptions and skill requirements.
• They are integrated into mathematical and network models to quantify job complexity, measure AI applicability, and reveal phase transitions in workforce performance.
• IWAs underpin empirical analyses by informing training, reskilling strategies, and the design of human–AI integration in diverse occupational settings.

O*NET Intermediate Work Activities (IWAs) are empirically coded, occupation-linked behavioral task units that serve as the backbone of contemporary analyses of labor, skill composition, and the implications of technological change—including AI—for workforce structure and economic value. IWAs bridge the granular, actionable interface between job descriptions and the skill requirements that drive occupational task performance, training, and reskilling. Recent research has leveraged IWAs as a central unit of analysis in modeling worker–job fit, quantifying AI applicability, and mapping labor market complexity through task–skill–occupation networks (Celis et al., 29 May 2025, Lee et al., 15 Jun 2025, Tomlinson et al., 10 Jul 2025).

1. Definition, Structure, and Position within O*NET

O*NET (Occupational Information Network) encodes occupational information for U.S. jobs in a multi-level schema. At the intermediate level, “Work Activities” (synonymously, IWAs) capture behavioral units that characterize what workers do on the job but are more granular than occupational titles and less fine-grained than atomic actions or skills. Examples include “Gather information,” “Write material,” “Edit documents,” and “Provide customer assistance.” Each occupation is defined by a weighted vector of IWAs, calibrated by relevance and frequency. These weights reflect detailed empirical measurement and allow for quantitative analyses that are robust to both cross-occupation and intra-occupation variability (Celis et al., 29 May 2025, Tomlinson et al., 10 Jul 2025). The demarcation of IWAs enables their use as nodes in occupational–task–skill bipartite graphs and as loci for mapping skill complexity, applicability of technology, and economic impact.

2. Mathematical and Network Models Leveraging IWAs

Recent formal frameworks have modeled jobs as sets of tasks (proxied by IWAs), themselves functions of underlying skills with varying complexity and weight. In (Celis et al., 29 May 2025), every job is treated as a set of $m$ tasks $\{T_i\}_{i=1}^m$, each of which depends on a specific subset of $n$ skills with associated difficulty parameters $s_j$. A bipartite “task–skill dependency graph” encodes these relations—tasks (typically mapped to IWAs) are linked to required skills, facilitating quantitative analysis of performance, fit, and the effects of worker heterogeneity or augmentation (such as via GenAI). This schema is extensively used to mathematically link observational workforce data (from O*NET or real-world system logs) with theoretical models of performance, success thresholds, and phase transitions in occupational task completion.

In occupational complexity research (Lee et al., 15 Jun 2025), the occupation–skill network is constructed as a bipartite graph $M_{os}$ (occupations to skills, with elements binarized at an importance cutoff). Community detection and complexity indices (e.g., Occupational Complexity Index, OCI; Skill Complexity Index, SCI) are computed by iterative methods (Method of Reflections). Although IWAs are not always distinguished from individual skills in every network analysis, the IWA level provides a natural aggregation unit that captures both the diversity of skill requirements and the practical structure of daily work.

3. Skill Decomposition: Decision-Level and Action-Level Subskills

A key innovation of recent frameworks is the systematic decomposition of each skill underlying an IWA into two subskills: decision-level (problem-solving, reasoning, planning) and action-level (execution, implementation) (Celis et al., 29 May 2025). Mathematically, for each skill $j$ in IWA $T_i$, overall skill difficulty $s_j$ is partitioned by a “decision-level degree” parameter $\lambda_j$, yielding $s_{j1} = \lambda_j s_j$ (decision) and $s_{j2} = (1-\lambda_j) s_j$ (action). Subskill-specific ability profiles (denoted $\alpha_1,\ \alpha_2$) are measured separately for each agent (worker or AI system). This decomposition reflects empirical findings that generative AI systems (GenAI) are typically stronger in action-level functions (e.g., text generation, code filling), while humans outperform in complex reasoning, diagnosis, and unstructured problem-solving. Applying this decomposition to O*NET IWAs supports nuanced, occupation-specific analyses of the complementarity between human and AI labor.

4. Quantitative Analysis: Success Metrics, Phase Transitions, and Merging Gains

The mathematical apparatus in (Celis et al., 29 May 2025) demonstrates that job success probability (the likelihood that error on all required IWAs falls below a tolerance $\tau$) is a sharp function of average decision-level ability, with a “phase transition” at a critical value $\mu_1^c$. Small increases in decision-level skill can push a worker from near-zero to near-unit probability of success, especially when noise is low and task complexity is well-aligned. Specifically, the width of the transition $\gamma_1$ is bounded by parameters such as the Lipschitz constant of error functions and the smallest derivative (“MinDer”), with representative bound:

$$\gamma_1 \approx \frac{L \cdot \sqrt{ n(\sigma_1^2 + \sigma_2^2) \ln(1/\theta) } }{ \text{MinDer}_{\mu_1} }$$

where $n$ is the number of relevant skills (often reflecting the number of IWAs linked to the job), and $\sigma_1,\,\sigma_2$ represent noise in ability distributions.

When two agents (e.g., a human and an AI system) are combined by taking the superior subskill ability for each skill—IWA pair, formal results (Theorem 2 in (Celis et al., 29 May 2025)) guarantee that the merged team’s job success significantly exceeds that of the weaker solo agent, provided success thresholds are straddled. This mathematically underpins observed “productivity compression,” in which technology augmentation can disproportionately benefit lower-skilled workers when the merging lifts them past the critical phase transition for IWA completion.

5. Empirical Analyses Using O*NET IWAs and Generative AI

Empirical studies assign quantitative scores to IWAs, aligning them between O*NET-coded activities and observed human–AI interactions (Tomlinson et al., 10 Jul 2025). The “AI applicability score” evaluates, for each occupation, the extent to which IWAs are successfully addressed by generative AI (Copilot), via:

$$a_i^{user} = \sum_{j \in IWA(i)} w_{ij} \cdot 1\{ f_j^{user} \geq 0.0005 \} \cdot c_j^{user} \cdot s_j^{user}$$

where $w_{ij}$ is the occupational weight for IWA $j$, $f_j^{user}$ is the observed activity share, $c_j^{user}$ the completion rate, and $s_j^{user}$ the impact scope. Parallel measures are computed for actions taken by the AI and then averaged to obtain an occupation-level exposure score. These scores validate that AI applicability—high for knowledge, communication, and writing-intensive occupations—concentrates in IWAs centered on information gathering, content creation, and customer/service interaction.

Task–skill analysis further utilizes IWA importance weights, O*NET proficiency scores, and empirically derived subskill splits (via NLP/GPT-4 analysis) to parameterize decision- and action-level demands per IWA (Celis et al., 29 May 2025). Real-world evaluation of completion and impact scope demonstrates which IWAs are most susceptible to automation or augmentation, highlighting alignment between theoretical models, forecasted AI impact, and actual usage patterns.

6. Labor Market Complexity and Economic Implications

Economic complexity research (Lee et al., 15 Jun 2025) frames occupations as bundles of IWAs and their underlying skills. Using structural network analysis—community detection (e.g., Louvain algorithm), binarized occupation–skill matrices, and the Method of Reflections—researchers derive the OCI and SCI, quantifying the depth and connectivity of occupations and skills. Notably, “general skills” (e.g., coordination, listening, literacy), which undergird many IWAs, form the core of the network, bridging the more specialized cognitive and physical skill communities. Regression analyses show that high general skill content in IWAs amplifies the wage premium from cognitive skills and offsets the wage penalty of physical skills. This suggests that policies focused solely on technical (specialized) skill or narrow IWA upskilling may be suboptimal; enhancing the general skill content of IWAs multiplies the value of specialized competencies within complex occupational environments.

The taxonomy and weighting of IWAs—together with their skill composition—thus serve as central analytical constructs in understanding occupational structure, wage formation, and the potential for skill-based reskilling or technology-driven transformation.

7. Broader Implications for Human–AI Workforce Integration

O*NET IWAs are the locus for practical human–AI integration. Quantitative modeling (Celis et al., 29 May 2025) and empirical studies (Tomlinson et al., 10 Jul 2025) both point toward an optimal paradigm in which GenAI augments rather than replaces human labor—specifically, by pairing action-level strengths of AI with decision-level strengths of human workers at the IWA/task level. The phase transition and merging theorems formalize the synergistic potential in this integration. Further, the structure of dependency networks among IWAs and skills, and their role in facilitating coordination, knowledge transfer, and adaptability, are critical in designing training, hiring, and performance evaluation practices that remain robust as technology reshapes the distribution and content of work.

In sum, O*NET Intermediate Work Activities operate both as empirical nodes for data-driven occupational science and as mathematically rigorous variables in modern models of skill, complexity, and AI-human complementarity, making them an indispensable foundation for both theoretical and applied research in labor economics, human capital modeling, and technology workforce integration.