Agentic Adoption Index from 53,000 configurations shows delegation peaks at bachelor's level
The study measures who actually delegates tasks to AI agents using 53,000 real configurations. It finds adoption concentrated at intermediate education and wage levels, not where technical exposure is highest. Distinguishing specification resistance from discretion requires longitudinal data.
The paper constructs delegated exposure by embedding agent routines and measuring cosine similarity to occupation tasks. This produces the AAI, which records actual workflow commitments rather than technical feasibility alone. Occupations identified differ from prior exposure indices focused on task replaceability.
AAI values track potential capability more closely than measured current usage rates. Technical availability accounts for most variance across the distribution, yet a persistent shortfall appears among the most educated occupations. This residual cannot be explained by feasibility metrics.
The pattern indicates that work resisting advance specification or protected by professional discretion limits codification. Repeated AAI measurement over time can separate these mechanisms. Operational systems can use the index to target integration where delegation already occurs rather than where models score highest on benchmarks.
Future releases of marketplace data will allow tracking whether the bachelor's peak shifts upward as specification tools improve.
Lanu Kim: AAI shortfall in doctoral occupations will remain above 25% of predicted level through 2028 unless specification tooling advances measurably.
Sources (2)
- [1]Primary Source(https://arxiv.org/abs/2608.20425)
- [2]Supporting Source(https://www.onetcenter.org/)