Theoretical models evaluating artificial intelligence in the labor market have long leaned on capability ceilings, estimating which professional tasks an algorithm could hypothetically execute rather than tracking what practitioners actively automate. Traditional occupational exposure frameworks mapped vulnerability by flagging routine labor or measuring ad-hoc conversational queries, creating a massive analytical blind spot between theoretical capability and realized pipeline integration. As generative systems transition from passive conversational wrappers toward autonomous agentic architectures pursuing multi-step objectives, assessing labor transformation requires auditing where structured delegation actually takes place.
Quantifying Delegated Exposure Through Workflow Artifacts
To bridge the divide between theoretical automation risk and actual operational handover, KAIST researchers Hyeongjae Lee, Jihyang Cheon, and Lanu Kim introduced the concept of delegated exposure. Unlike raw chatbot telemetry that merely captures reactive prompting, agent configurations require developers and operators to explicitly define goals, toolsets, and execution logic up front. As the KAIST research team demonstrated, delegated exposure documents whether an engineer or knowledge worker has committed a task to autonomous code by formalizing it into a persistent workflow.
To measure this dynamic, the researchers constructed the Agentic Adoption Index (AAI), a metric benchmarking how closely standard occupational tasks align with real-world agentic routines. The methodology embeds roughly 53,000 agent skill specifications sourced from the Manus Skills Marketplace, evaluating their semantic similarity against approximately 18,000 occupational task statements in the O*NET database. By mapping these semantic matches across standard occupational classifications, the index cleanly separates deliberate systemic workflow codification from casual chat experiments.
Structural Deviations in Real-World Delegation
Empirical findings from the 53,000 agent specifications demonstrate that real-world delegated exposure diverges sharply from legacy automation forecasts. Earlier labor economics frameworks predicted displacement would concentrate primarily within routine mechanical tasks, but autonomous delegation clusters heavily in complex cognitive operations. Furthermore, the analysis reveals that the Agentic Adoption Index correlates with technical model ceiling limits far more tightly than it correlates with conversational prompt usage, exposing a profound split between what users type into a chat box and what operators build into autonomous pipelines.
"the AAI peaks below the top of the wage distribution and at the bachelor’s level, declining at both extremes."
The study uncovers a distinct distribution across compensation and education tiers: delegation intensity peaks across mid-to-high wage brackets and roles requiring a bachelor's degree, tapering off sharply among the most credentialed and highest-earning professions. While technical availability accounts for most variance across the broader labor market, it fails to explain this delegation deficit at the top of the knowledge stack.
The gap between theoretical capability and operational delegation proves that technical feasibility alone does not dictate enterprise automation. The adoption plateau among elite cognitive roles highlights structural operational barriers: senior workflows resist deterministic pre-specification, while experienced practitioners actively gatekeep task codification. Because the dataset relies on public configurations from the Manus Skills Marketplace, it mirrors proactive platform builders rather than closed enterprise stacks. Yet the strategic takeaway for technical leadership is unambiguous: closing the automation gap is not a matter of waiting for more capable foundation models, but of solving task ambiguity and formalizing pipeline logic where ad-hoc discretion currently rules.