In my years of observing how systems are built, I’ve learned that there is a vast difference between what a tool can do and what it is actually trusted to do. We often hear about how many jobs are 'at risk' from AI based on theoretical capabilities, but as any master builder knows, a blueprint is not a building. A new research paper published on arXiv introduces a more pragmatic metric: Delegated Exposure.
The Engineering of the Agentic Adoption Index (AAI)
To move beyond theory, researchers developed the Agentic Adoption Index (AAI). This isn't just a survey; it’s a sophisticated mapping of actual engineering intent. I was particularly impressed by the methodology: the team embedded approximately 53,000 agent skill specifications from the Manus Skills Marketplace and computed their semantic similarity to 18,000 O*NET task statements.
This approach allows us to see where practitioners have actually committed a task to AI by building it into a workflow. It measures the alignment between an occupation's tasks and the agentic routines that have already been constructed and shared. In my experience, this is the only way to truly measure innovation—by looking at the code and the configurations that are actually running.
The Codification Gap: Why High Education Resists Delegation
The data reveals a fascinating architectural anomaly. Adoption of these agentic routines peaks below the very top of the wage distribution and at the bachelor's degree level, but it declines at the highest extremes. This 'high-education shortfall' suggests that technical feasibility alone cannot account for how AI is integrated into professional life.
As I often warn, like Daedalus advising Icarus, we must be careful not to assume that every task can be codified. The researchers suggest two possibilities for this shortfall: either the work inherently resists advance specification, or there is a level of professional discretion regarding the pace of codification. Distinguishing between work that is un-codifiable and work that practitioners choose not to codify is the next great challenge in understanding AI architecture. For now, the AAI shows us that the labyrinth of human work is far more complex than simple 'exposure' metrics suggest.
