OpenAI study tracks the shift from chat to agents at work
Codex usage data shows longer tasks and growing adoption beyond engineering. The study also carries a warning: OpenAI's internal experience is not representative of a typical company.
Summary
OpenAI has published a study of Codex usage that examines the move from short AI conversations to work delegated to agents. The analysis compares individual accounts, organizational accounts, and OpenAI workers.
In June 2026, Codex accounted for 99.8% of output tokens generated by OpenAI workers across Codex and ChatGPT. The share was 63.3% for organizational accounts and 16.5% for individual accounts. This points to a genuine shift among adopters, but not broad adoption across every group.
In practice
The relevant pattern is not simply more AI use. People are handing off larger tasks that can involve files, tools, commands, and multiple steps. Among the individual users analyzed, 70.2% made at least one request that the study estimated would take a person more than one hour, while 25.6% made at least one request estimated at more than eight hours.
Use is also moving beyond engineering. At OpenAI, legal, finance, and recruiting teams use Codex for automation, data transformation, structured analysis, and technical execution. In business functions, more than a quarter of the work done with Codex was classified as engineering or coding.
What we still don't know
The figures need context. The paper acknowledges that OpenAI is an unusual workplace: employees have broad access to the models, low marginal costs, internal training, and strong organizational support. Its internal results show a possible frontier, not the current behavior of a typical organization.
Task duration is also estimated by a model from user requests and should be treated as directional rather than as proven hours of productivity. The study measures use and apparent complexity, but does not by itself establish equivalent gains in quality or final outcomes.
Why it matters
- AI work is moving from isolated questions to delegated workflows that require supervision and verification.
- Adoption beyond engineering can reduce technical bottlenecks, while placing more responsibility on the people who define and review the work.
- For businesses, the value of agents will depend on process redesign as much as model capability.