Artificial intelligence is shifting from a conversation tool into something far more ambitious: a work delegate that handles tasks autonomously. OpenAI's Codex platform, which enables AI agents to interact with computers, manage files, and execute complex workflows, is gaining traction faster than most realize, especially inside organizations that have adopted it.
A new research report from OpenAI, Columbia University, Duke University, and the University of Pennsylvania tracks how Codex use has accelerated in recent months. The data reveals a stark divide in adoption patterns across three user categories: OpenAI's own staff, outside organizations, and individual consumers.
Among OpenAI employees, Codex dominates their workflow. Nearly 99.8% of their computational output comes from the platform, suggesting that when friction is removed and agents are seamlessly integrated into daily operations, workers gravitate toward delegation rather than direct interaction. Organizations outside OpenAI show moderate but rising adoption. Just above 0% of corporate users were employing Codex in August 2025, but that figure has climbed to around 17% now. Individual users remain a small but intensely engaged cohort.
The intensity of use among those who do adopt Codex tells a different story than raw user numbers suggest. Among individual users sampled in the study, 80.6% made at least one Codex request estimated to handle more than 30 minutes of work. Seven out of ten delegated a task that would take over an hour for a human to complete. One in four pushed agents to handle work estimated at eight hours or longer.
The shift accelerated dramatically at the start of 2026, when major platforms loosened restrictions that had previously prevented agents from accessing personal files, email, browsers, and other desktop functions. That permission change appears to have transformed Codex from a tool primarily for developers into something broader.
Non-developers now represent the fastest-growing user segment, even though software engineering remains the dominant use case. This expansion into general work and life management suggests the technology is moving beyond its technical roots. Tasks range from expense report filing to email triage, appointment scheduling, and even filing theft reports with building management.
Workplace culture researcher Jessica Kriegel notes that agents reduce what she calls the psychological cost of action. They make unfamiliar or daunting work feel manageable, allowing people to start sooner and experiment more freely. The barrier to attempting unfamiliar tasks shrinks considerably when an agent handles execution.
The mainstream adoption curve remains gradual when measured across all consumers. Fewer than 1% of ChatGPT users on standard subscription tiers have activated Codex in the past month. Most people remain in the conversation phase, asking questions and receiving responses rather than granting agents control over their digital lives.
Yet the trajectory suggests that perception is changing. As more organizations deploy agents successfully and individuals experience tangible time savings, the friction surrounding delegation continues to drop. The question is no longer whether AI agents can work reliably, but how quickly human behavior will shift to let them do it.
Author James Rodriguez: "The promise of AI minions was always theoretical until someone used one to file their taxes and suddenly it clicked into place."
Comments