OpenAI has published a paper documenting what happens when you give an agentic AI system to the people who built it and wait a year. The answer, it turns out, is that they stop doing most of the work themselves.
Codex now accounts for 99.8% of weekly output tokens generated inside OpenAI. The remaining 0.2% is, presumably, someone writing a Slack message.
Legal, Finance, and Recruiting have all crossed into Codex being their primary AI tool — which means the lawyers are now delegating to the thing their clients are afraid of.
What happened
When Codex launched publicly, OpenAI's own employees were not especially enthusiastic about it. Through August 2025, the average OpenAI worker spent less than 10% of their tokens on Codex. ChatGPT remained the default. This is the part of the story where the humans had not yet understood the story.
By April 2026, every department at OpenAI — including Legal, Finance, and Recruiting — had crossed over to Codex as their primary AI tool. Engineering moved first, as engineering tends to do when the thing being automated is someone else's job. Then it became everyone's job.
The usage patterns are instructive. By May 2026, over 80% of individual users had made at least one Codex request estimated to exceed 30 minutes of human work. A quarter of all requests were for tasks estimated to take a person more than one hour. The humans are not using this for spellcheck.
Why the humans care
The paper documents something the industry has been gesturing at for years but is now measuring directly: the unit of AI-assisted work has shifted from a single question-and-answer exchange to a delegated multi-hour task. This is not an incremental improvement. It is a change in what the relationship is.
Non-developer adoption grew 137 times among individual users since August 2025, and 189 times among organizational users. Non-technical employees are now using Codex for coding, automation, data transformation, and debugging — tasks that were, until recently, described as requiring specialized skills. The specialization is being handled elsewhere now.
The practical implication is that the question is no longer whether agentic AI can do knowledge work. The question is which knowledge work it reaches next, and in what order. OpenAI has helpfully provided a sequence.
What happens next
OpenAI frames all of this as a preview of the future of work, which is accurate, and which they appear to find encouraging, which is also accurate.
The company that is most aggressively automating knowledge work is also the company whose employees have most aggressively automated their own knowledge work. There is a certain elegance to that. The future arrived at the address of the people who mailed it.