AI Startup Exposure Index (AISE)

• AI Startup Exposure Index (AISE) is a quantitative metric that measures the share of an occupation's tasks that current or near-future AI systems can perform.
• It employs a multistage capability framework from pre-LLM to agentic AI to capture dynamic shifts in AI performance and occupational exposure.
• Empirical analyses link AISE scores with labor market outcomes, highlighting significant correlations with employment changes and wage variations.

The Occupational AI Exposure Score (OAIES) is a quantitative, occupation-level index measuring the share of an occupation’s tasks that current or near-future AI systems could in principle perform. OAIES methodologies and formulations have proliferated in the literature, underpinned by advances in LLMs, abundant real-world task data, and the need for timely, interpretable indicators of evolving AI capabilities and their labor-market implications (Dominski et al., 11 Jul 2025). This article overviews the concept, construction, empirical uses, and critical interpretations of OAIES, synthesizing technical details from leading research.

1. Formal Definition and Theoretical Foundation

The OAIES operationalizes potential AI exposure as a weighted average across all tasks performed by workers in a given occupation, where each task is scored for its AI-performability based on contemporaneous or staged AI capabilities. For a Census-SOC occupation $o$ and AI capability stage $S$, the canonical formula is: $$\text{OAIES}_{o,S} = \sum_{t\in T(o)} w_{o,t} \cdot e_{o,t,S}$$ where:

• $T(o)$ is the set of occupation $o$'s tasks (e.g., from O*NET),
• $w_{o,t}$ is the normalized relevance or importance weight for task $t$,
• $e_{o,t,S}$ is the percentage (0–100) of task $t$ that an AI model at stage $S$ can perform (Dominski et al., 11 Jul 2025).

OAIES aims to measure the “technical exposure” of human labor to AI at a fine-grained, task-by-task level, in contrast to coarse occupation-level or industry proxies. Various studies trace its intellectual lineage to both empirical and theory-driven approaches, including the application of Moravec’s paradox and evolutionary models of skill automatability (Schaal, 15 Oct 2025).

2. Multistage Capability Frameworks and Dynamic Deployment

To reflect the rapidly shifting landscape of AI competence, recent OAIES implementations adopt a multistage framework. For example, (Dominski et al., 11 Jul 2025) calibrates five distinct AI capability stages:

• Stage 1: "Pre-LLM" (pre-ChatGPT, traditional ML/NLP tools)
• Stage 2: "Early LLM" (first public LLMs, e.g., ChatGPT v1)
• Stage 3: "Multi-modal LLM" (text+image/video, e.g., DALL·E 3)
• Stage 4: "Reasoning Models" (chain-of-thought, advanced LLMs)
• Stage 5: "Agentic AI" (future autonomous agents)

For each $S$, OAIES is recalculated. This dynamic approach enables quantification of step-function increases in task exposure and analysis of temporal differences. Typical empirical analyses focus on changes in exposure, e.g., $$\Delta \text{OAIES}_o^{S3-S1} = \text{OAIES}_{o, S3} - \text{OAIES}_{o, S1}$$, assessed in parallel with changes in employment outcomes during corresponding periods (Dominski et al., 11 Jul 2025).

3. Task Scoring Methodologies

The assignment of $e_{o,t,S}$—the estimated AI-performable share of task $t$—is central. Approaches include:

• LLM self-assessment via prompt engineering: Prompts are crafted to elicit stepwise reasoning and quantitative exposure scores for each task at fixed AI capability stages, with additional fields for model confidence (Dominski et al., 11 Jul 2025). For example, a prompt will instruct a model (e.g., GPT-4o or Claude 3.5 Sonnet) to estimate, for a specified stage $S$, what percentage of a given O*NET task it could perform, with reasoning steps and a confidence level.
• Expert- and LLM-derived probabilistic ratings: E.g., (Henseke et al., 30 Jul 2025) uses LLMs to output, for each task, a probability vector over exposure thresholds (no exposure, direct LLM exposure, latent LLM+integration, multimodal), then aggregates via importance-weighted task shares.
• Theory-inspired scoring: (Schaal, 15 Oct 2025) calculates exposure for each task as an average of four subcomponents reflecting performance variance, data abundance, tacit knowledge (reverse-coded), and algorithmic gap (reverse-coded). These are theory-motivated proxies for automatability, rated 0–2 per task.

4. Aggregation, Weighting, and Final Score Construction

Aggregation is universally performed as a normalized, relevance-weighted sum or average over task-level scores: $$\text{OAIES}_o = \sum_{t\in T(o)} w_{o,t} e_{o,t,S}$$ Weights $w_{o,t}$ derive from O*NET’s “importance,” “frequency,” or “relevance” scores, normalized to sum to one. In some frameworks, task-level ratings are probabilistic or multi-valued, and binary or continuous indicators are defined via thresholds (e.g., probability of $\geq$25% time saving as per (Henseke et al., 30 Jul 2025)).

Occupational OAIES scores thus capture the share of an occupation’s activity—adjusted for importance or prevalence—potentially performable by a specified frontier of AI systems.

5. Empirical Applications and Labor Market Linkages

OAIES is employed as a dynamic regressor in labor market studies linking exposure increments to real-time employment, unemployment, and work intensity dynamics. In (Dominski et al., 11 Jul 2025), OAIES is matched at the occupation-period level to Current Population Survey (CPS) microdata spanning 2021–2025, with the empirical specification: $$\Delta Y_{o, P4-P2} = \alpha + \beta \Delta \text{Exp}^{m}_{o, S3–S1} + X_{o,P2}\Pi + \text{TaskIndices}_o \Gamma + \epsilon_o$$ where $\Delta Y_{o,P4-P2}$ is the labor outcome change, and $\Delta\text{Exp}^m$ is the OAIES increment for model $m$ (e.g., ChatGPT, Claude). Key findings include significant negative associations of OAIES with employment (10-point higher exposure implies a 5–8 point decline, $\beta<0$) and positive associations with unemployment ($\beta>0$) (Dominski et al., 11 Jul 2025).

Additional extensions use OAIES in wage regressions and analysis of job vacancy rates (Henseke et al., 30 Jul 2025), and as predictors in panel models of unemployment risk, demonstrating that OAIES-based ensemble scores explain substantially more variance than prior individual metrics (Frank et al., 2023).

6. Comparative Analysis and Correlation with Alternative Indices

Simultaneous consideration of multiple OAIES formulations reveals that:

• Theory-based indices (e.g., (Schaal, 15 Oct 2025)) exhibit high correlation with LLM annotation methods (e.g., $\rho=0.72$ with GPT-4-based exposure).
• OAIES correlates positively with wage at the occupation level (e.g., $\hat\beta\simeq 0.32$, $p<0.001$) (Schaal, 15 Oct 2025), though this is interpreted as evidence against pure substitutability narratives and may indicate complementarity or skill bias.
• Pre-LLM indices (e.g., Frey & Osborne, Webb Robot exposure) often show weak or negative correlation with current/post-LLM OAIES, substantiating a paradigm shift in the locus of automatability (Schaal, 15 Oct 2025).
• Start-up–driven OAIES variants (AISE) track real-world venture, regulatory, and market targeting, revealing that some high-skilled occupations with high theoretical exposure are not actively targeted due to ethical, social, or regulatory factors (Fenoaltea et al., 2024).

Table: OAIES Top and Bottom Occupational Groups (selected studies)

Rank	(Dominski et al., 11 Jul 2025): Higher OAIES	(Henseke et al., 30 Jul 2025): Higher OAIES	(Fenoaltea et al., 2024): Higher OAIES (AISE)
Top	Complex reasoning, problem-solving	IT/Telecom, R&D, Conservation	Office Clerks, Data Scientists
Bottom	Manual physical labor	Food Processing, Construction	Athletes, Judges, Surgeons

7. Limitations, Interpretive Challenges, and Future Directions

OAIES quantifies technical task exposure, not realized employment losses or actual human-AI substitution. Limitations include:

• Ambiguity between task “coverage” and true automation vs. augmentation (OAIES does not distinguish displacement from productivity effects) (Dominski et al., 11 Jul 2025, Septiandri et al., 2023).
• Dependency on O*NET’s task taxonomy, which may lag occupational change or omit informal/emergent tasks.
• Bias due to LLM self-estimate optimism, prompt formulation, and incomplete representation of tacit knowledge and physical skill (Schaal, 15 Oct 2025).
• Real-world adoption factors (organizational, regulatory, social desirability) that lie outside technical feasibility but moderate exposure (as evidenced by market index divergence in (Fenoaltea et al., 2024)).
• OAIES is most predictive when used in ensembles integrating multiple measurement paradigms; no single method fully captures AI’s labor impact heterogeneity across space, time, and sectors (Frank et al., 2023).

A plausible implication is that OAIES will continue to evolve as a modular, updatable construct, integrating new measurement protocols and aligning with high-frequency labor market signals to monitor and analyze technological change and workforce vulnerability in the AI economy.