Agentic AI in industry: 36% of tasks are “automatable” — nowhere near automated

Deloitte's latest aerospace and defense sector outlook, drawing on forecasts from IDC, puts forward a striking figure: 36% of tasks performed in industrial products manufacturing could benefit from augmentation by agentic AI — software agents able to chain diagnosis, planning and execution without constant operator supervision. The same outlook also quantifies the investment trajectory: US aerospace and defense AI spending is projected to grow 3.5x between 2025 and 2029, reaching $5.8 billion.

This kind of “automatable task” statistic has become a staple of consulting outlooks since the foundational work of Autor, Levy and Murnane (2003) on the skill content of technological change. Their key contribution was to show that a job is not a single task but a bundle of tasks, and that only routine, codifiable tasks yield quickly to automation — non-routine, relational or judgment-based tasks resist far longer. A 36% figure therefore says something about the technical task content of industrial work, not about the pace at which that potential converts into transformed jobs or measured productivity gains.

The gap between this technical ceiling and actual adoption has a well-documented explanation in the economics of innovation: Bresnahan, Brynjolfsson and Hitt (2002), using US firm-level data, showed that the value of information technologies only materializes through massive “co-investment” in organizational capital — process redesign, retraining, data infrastructure overhaul — whose rollout is measured in years, not quarters. The spending forecast (3.5x by 2029) measures tooling budgets; it does not measure that organizational co-investment, which has historically been the binding constraint.

For industrial operators in the sectors Cardan-AI tracks (aerospace, energy, oil & gas), the lesson is direct: a vendor or consultancy presenting an “automatable task” percentage as a roadmap is smuggling in an assumption of instant diffusion that four decades of technology-adoption research consistently reject. The metric worth tracking is not the technical ceiling but the budget and timeline committed to the process reorganization that has to accompany it.