OpenAI has published new economic research arguing that AI agents are overtaking chatbots as the company’s main tool for workplace tasks, with its Codex system now used across technical and non-technical teams.
The research, released June 25, examines how agentic AI changes knowledge work by shifting users away from short chat exchanges and toward longer assignments that can run for minutes or hours. OpenAI says that pattern has become visible inside its own organization, where Codex has moved from an engineering-focused product to a primary work tool for departments including Legal, Finance and Recruiting.
According to the company, ChatGPT was still the default workplace AI at OpenAI during the first months after Codex was introduced publicly. Through August 2025, OpenAI said employees were spending less than 10% of their AI tokens on Codex. By now, the company says every department uses the system as its main AI tool for work.
OpenAI’s paper says the kinds of requests people make with agents have grown more complex over time. The company reports that by May 2026, 80.6% of sampled individual users had made at least one Codex request that the company estimated would take more than 30 minutes for a person to complete. More than 70% had made at least one request estimated to take over an hour, and about a quarter had used it for work the company estimated would take more than eight hours.
The company also said heavy users are now orchestrating substantial amounts of agent activity in a single day. By June 2026, users in the 99th percentile were regularly generating more than 60 hours of Codex agent turns per day, spread across multiple parallel agents.
OpenAI says that growth reflects improvements in Codex’s underlying models and product features, which have expanded the range of tasks the system can handle. The company describes agents as tools that can operate more independently than chatbots, calling on software tools, interacting with environments and iterating toward a goal over longer periods.
The research says engineers were the earliest adopters, but use later broadened quickly across the company. OpenAI said the average engineer now generates 99% of output tokens with Codex rather than ChatGPT. Legal, finance and recruiting moved to majority Codex use later, around April 2026, but now each of those departments generates more than 85% of output tokens on the platform.
OpenAI also said usage intensified over the last six months. It reported that research saw the largest increase in median token output, at 56 times its November 2025 level by June 2026. Customer support rose 32 times, engineering 27 times and legal 13 times.
The company highlighted strong growth among non-developers, saying adoption among those users has outpaced developer growth in individual, organizational and internal OpenAI groups. It said non-developer individual users rose 137-fold since August 2025, while non-developer organizational users increased 189-fold.
OpenAI says the findings matter beyond its own products because they suggest how agentic tools may reshape work design, skill needs and labor markets. The paper argues that when workers have broad access to capable agents, they use them for longer, more complicated and more cross-functional tasks.
The company says non-technical employees are already using Codex for work that goes beyond their formal roles, including automation, data transformation, debugging and structured analysis. In business functions, OpenAI says more than one-fourth of Codex work involved engineering or coding tasks.
OpenAI frames the results as an early look at how AI agents could expand the range of productive work available to employees outside traditional technical roles. The company says that trend is likely to continue as agent tools become more capable and easier to use.