Codex now makes up 99.8% of AI output among OpenAI workers, paper finds

From chatbot to agent across every department

The working paper, titled The Shift to Agentic AI: Evidence from Codex, was co-authored by David Holtz, an assistant professor at Columbia Business School who is serving as Head of AGIO Economic Research at OpenAI, together with several OpenAI employees. Holtz did the work as a visiting researcher at OpenAI. The findings were summarised by the MIT Initiative on the Digital Economy on October 20, 2026.

Engineers moved first. By December 2025, the average OpenAI engineer did most of their AI work on Codex, and by March 2026 that share topped 90%. Other departments started later but caught up quickly. In January, Codex use in functions such as legal and recruiting was close to zero. It rose gradually to about 20% by early April, then jumped to roughly 75% within a month. As of June, the average OpenAI recruiter generates 89% of their AI output on Codex, and the average member of the legal team 88%.

According to the researchers, the jump coincided with a company-wide push to adopt Codex across workflows, including training sessions and regular feedback loops.

What the usage data shows

The researchers analysed Codex usage data collected between August 2025 and June 2026. They used AI classifiers to sort requests by task type without reading the underlying messages. To gauge complexity, they used an AI model to estimate how long each request would take an experienced human, based on a small random sample of personal-account users.

• Weekly active Codex users grew more than fivefold from January to May.
• More than 10% of all users manage three or more concurrent Codex agents at some point each week.
• Across all Codex users, 26.6% used skills, a feature for sharing instructions for complex workflows. At OpenAI the share was 96%.
• In December 2025, 35.4% of Codex personal-account users sent at least one request for a task that would take an experienced human an hour or more. By May that share was 70.2%.
• The share requesting at least one task that would take a human more than eight hours rose from 2.1% to 25.6%.

Token output also climbed. In June, the median OpenAI employee in a legal role generated 13 times more monthly tokens across Codex and ChatGPT than in November of last year, and the median researcher generated more than 50 times as many. Output rose at least tenfold in every job function at OpenAI.

The paper describes a difference in behaviour. Earlier research on ChatGPT found that nearly half of all prompts were requests for information and advice, while Codex users mostly ask the agent to do the work itself, such as debugging code, analysing data and drafting documents. The heaviest users delegate work, then supervise, review and integrate what the agents produce.

Adoption outside OpenAI is slower

The picture is different elsewhere. Only 17.3% of users with organizational accounts used Codex in a 28-day period, and the average legal user at these organizations did just 1.9% of their AI work on Codex. Those who do adopt it use it heavily, though: Codex generates 63.3% of the AI output from organizational accounts.

Outside OpenAI, about two-thirds of Codex users (67% of organizational users and 64% of personal-account users) do not yet run agents in parallel. Among OpenAI employees, 28.6% managed five or more concurrent agents at some point during an observed week in June 2026.

The researchers note that OpenAI is an unusually favorable setting. Its workers face no limits on AI use, know the tools well and work in a company with strong buy-in and constant knowledge sharing. The article also cites a McKinsey & Co. survey in which 40% of large organizations said they are now scaling AI agents, up from 27% a year earlier.

What it means for businesses

The researchers argue that the biggest payoffs may still be ahead. Drawing on past technology shifts, they write that "the largest productivity gains often arise when firms reorganize production around the new technology rather than merely substitute it into existing workflows." They compare it to electrification, where factories gained most only after redesigning floors around smaller electric motors, a process that took decades. Agentic AI could move faster, they suggest, because testing new workflows is relatively cheap.

The article also lists risks for business leaders. Averages can mislead, since a small group of heavy users can make adoption look broader than it is. More output is not necessarily more productivity, because human review and other downstream steps can remain bottlenecks. Overseeing agents takes work and domain expertise, and tools alone are not enough without redesigned workflows, roles and review processes. The researchers also say the shifts could affect team composition and hiring needs.