OpenAI Codex Agent Adoption Surges to 17% at Organizations, Research Finds AI Work Delegation Accelerating
Summary
- • Joint OpenAI-university study shows Codex adoption jumped from near-zero to 17% at organizations in under a year
- • 99.8% of OpenAI employees' output tokens now produced via Codex agents, signaling near-total internal shift away from ChatGPT
- • 80.6% of individual Codex users delegated tasks representing 30+ minutes of human work; 25.6% delegated 8+ hour tasks
- • Non-developers are fastest-growing Codex user segment as agentic AI expands well beyond coding use cases
Details
OpenAI partnered with Columbia, Duke, and UPenn on Codex adoption study
Report measuring Codex vs. ChatGPT usage across OpenAI employees, outside organizations, and individual users
99.8% of OpenAI employee output tokens are routed through Codex agents
Compared to 63% for outside organizations and 16.5% for individual users — internal usage intended to model barrier-free adoption ceiling
Organizational Codex adoption rose from ~0% to 17% in under a year
Among active users of ChatGPT and Codex at outside organizations, Codex share grew from near-zero in August 2025 to ~17% now
80.6% of individual users delegated tasks worth 30+ minutes of human work
Sampled from individual Codex users; threshold is model-estimated based on experienced human equivalent
70.2% of individual Codex users saved at least one hour of human work per request
Demonstrates agents are handling substantial tasks, not just quick lookups
25.6% of users delegated tasks representing 8+ hours of human work
A quarter of individual Codex users have delegated genuinely complex, long-duration work to agents
Non-developers are the fastest-growing segment of Codex users
Despite software coding remaining the core use case, agents are expanding into non-technical workflows
Fewer than 1% of ChatGPT consumer plan users also actively use Codex
Covers Go, Free, Pro, and Plus plan users active in the last 28 days — agentic AI still early-adopter territory for consumers
Study drew from a 0.1% random sample of opt-in individual users
Task-time thresholds are model-estimated; data based on users who opted in to allow queries for training purposes
Agents reduce the 'psychological cost of action' for unfamiliar tasks
Workplace expert Jessica Kriegel: agents make unfamiliar work feel more approachable, lowering activation energy and increasing experimentation
Data from OpenAI/Columbia/Duke/UPenn joint report shared exclusively with Axios, June 2026
What This Means
AI agents are crossing from developer tools into mainstream workplace use, with documented organizational adoption jumping from near-zero to 17% in under a year. The data shows users who adopt agents quickly rely on them for substantial tasks — a behavioral shift signaling genuine utility rather than novelty. The stark gap between OpenAI internal usage (99.8%) and broader consumer usage (under 1%) suggests the bottleneck is familiarity and access rather than capability, meaning wider adoption could accelerate sharply as friction falls. Non-developer growth signals agents are beginning to reshape work patterns well beyond the software industry.