AI doesn't just change what workers do —
it changes how many a manager can lead

What the model is doing — in plain English

Economists have long studied how technology changes which tasks workers do. Farach argues there is a second, equally important channel: technology also changes how work is organised. His model starts with a simple observation — every additional worker a manager oversees adds a slice of overhead: briefings, check-ins, resolving misunderstandings. That overhead is what the left chart captures as "coordination cost per worker."

Agent capital (K_A) is the paper's name for AI tools that specifically reduce this overhead — think shared dashboards, AI-drafted summaries, automated status updates, or AI agents that handle routine queries before they reach a human manager. As firms invest in more of this, the per-worker overhead falls (left chart), and the maximum team one manager can effectively run rises (right chart). When spans of control expand, firms need fewer management layers — hierarchies flatten.

The four curves represent how effective a particular AI deployment is at cutting friction. A firm using AI mainly for document drafting (low effectiveness) sees a gentle improvement; a firm using AI to orchestrate entire workflows (very high effectiveness) sees dramatic compression. The key insight: the same underlying AI technology can produce very different organisational outcomes depending on who captures its coordination benefits — broadly distributed gains, or concentration at the top.

Reference

Farach, A. (2026). AI as coordination-compressing capital: Task reallocation, organizational redesign, and the regime fork. arXiv. https://arxiv.org/abs/2602.16078