OpenAI Economics Paper Reveals Shift to Agentic Model in Knowledge Work
OpenAI paper "The Shift to Agentic AI: Evidence from Codex" tracks Codex data, showing that work units have shifted from single conversations to asynchronous delegation of complex tasks.
99.8% of internal employee tokens come from Codex, with engineers making up 99% and over 85% being non-technical staff; 28.6% of employees manage more than 5 agents simultaneously, with the most intensive users averaging over 71 hours per day.
The proportion of external users spending over 8 hours on tasks increased from 2.1% to 25.6% in six months, with a surge in non-developer users, and the usage rate of Skills jumped from 5.4% to 26.6% in three months.
In market mechanisms, AI agent users and developers have become the main adopters, with event-driven funding flowing towards agent platforms and workflow tools, benefiting the OpenAI Codex ecosystem while traditional chat interfaces are under pressure.
Source: Public Information
ABAB AI Insight
OpenAI has previously promoted Codex heavily, and this paper continues its strategy of validating the shift to agents through internal data. Earlier similar growth in skill reuse reflects the transition from dialogue to asynchronous coordination.
In terms of capital pathways, Codex is taking over daily tasks, with strategic motives aimed at enhancing internal efficiency and promoting externally, shifting resources from chat interfaces to agent coordination and Skills workflow delivery.
Similar to reports from other AI labs on agent adoption, OpenAI is currently in an expansion phase of fully deploying agent-based knowledge work, with the internal 99.8% token share highlighting an extreme state.
Essentially, this represents a technological replacement, where agent asynchronous delegation replaces single conversations, with the mechanism of workflow reuse reducing human coordination costs, leading to a shift in pricing power from chat tools to agent platforms, and driving the knowledge work industry chain towards AI agent reconstruction.
ABAB News · Cognitive Law
Work Model = Dialogue Unit × Agent Delegation × Skill Reuse
Human chat sells immediacy, AI agents sell asynchronously; those managing multiple agents control productivity leverage.
The deeper the adoption, the more thorough the transformation; the counterintuitive aspect is that agent takeover accelerates OpenAI's capital and user concentration.