OpenAI published research on June 25 showing that by April 2026, its legal, finance, and recruiting teams all crossed over to using AI agents as their primary work tool. Non-developer organizational users grew 189 times since August 2025, and a quarter of individual users now regularly delegate tasks estimated to take a human eight or more hours. The data makes a case that agentic AI has moved from engineering teams to every department.
The recruiter who spends six hours a week screening resumes and scheduling first-round calls is not a candidate for automation in some vague future. She is the reference category in a research paper OpenAI published yesterday. By April 2026, the recruiting team at OpenAI had crossed a threshold where AI agents generated more than 85% of their work output. The legal team crossed in the same month. Finance crossed around the same time.
OpenAI published "The Shift to Agentic AI: Evidence from Codex" on June 25, a research paper tracking how its agentic AI tool spread through its own workforce and among external organizational users over the past year. The findings are striking even adjusted for the source. Non-developer organizational users of Codex grew 189 times since August 2025. A quarter of individual users regularly delegated tasks estimated to take a human more than eight hours. Nearly all of OpenAI's 98% internal adoption happened not because engineers needed a better coding tool, but because people in non-technical roles found a way to use one.
What changed in April
OpenAI traces a clear sequence. Engineering moved to agents first, starting in mid-2025. Legal, finance, and recruiting each crossed the threshold to Codex as primary tool around April 2026. The transition for non-technical departments was faster than it was for engineers, which is counterintuitive until you look at what Codex is actually doing in those departments.
Finance and business operations workers at OpenAI spent 31% of their Codex time on engineering and coding tasks. Product, marketing, and operations workers spent 25% of their Codex time on engineering and coding. People whose jobs are not technical are using an agent to do technical work because it is no longer expensive to cross that task boundary. You describe what you need. The agent figures out the implementation.
That is a real structural change. For years, the bottleneck in cross-functional work was the ask. Marketing needed a data pull, so they waited for an analyst. Legal needed a document formatted a specific way, so they waited for IT. Operations needed a workflow automated, so they put in a ticket. When an agent can handle the technical execution, the cost of that ask drops close to zero.
The business math behind 189x growth
The number that matters most to a CMO or RevOps leader is not 85% internal adoption. It is 189 times growth in non-developer organizational users since August 2025. That is the external curve, the one that includes companies not named OpenAI and not institutionally incentivized to maximize agent usage.
A task that takes a human more than one hour now represents a quarter of all Codex requests from individual users. That is not a curiosity about a coding tool. It is a description of what multi-hour delegated knowledge work looks like when the tool is capable enough to handle it. Campaign research, contract review, lead qualification, financial modeling, onboarding documentation, competitive analysis. The list of multi-hour tasks that fit an agentic delegation pattern is long in every business function.
The research paper frames the business implication directly: "Agents can lower the cost of moving across task boundaries and help workers do adjacent work that used to require more specialized technical support." For a team that regularly bottlenecks on tasks that need technical help to execute, that is the sentence worth reading twice.
The honest caveat
OpenAI is the company that makes Codex. This paper is self-reported internal usage data from a company with a direct financial incentive to demonstrate the product's value, and the timing is not neutral: OpenAI is reportedly pursuing an IPO in Q4. The Next Web noted that every metric "comes from OpenAI itself," and no independent third party has verified any of the usage figures.
The methodology for estimating task horizons uses an LLM as a judge reading Codex transcripts, which means the eight-hour threshold is a model estimate, not a time-tracked measurement. The transfer from OpenAI's internal adoption curve to a typical company's adoption curve is also not guaranteed. OpenAI employees are highly technical, operate in an environment that actively encourages AI adoption, and have first access to the most capable version of the tool. Most business teams are not anywhere near these numbers.
That said, the directional signal is confirmed by the external user data. The 189x growth in non-developer organizational users is not an internal metric. That one comes from outside the company.
The closing observation
For three years the productivity narrative around AI centered on the individual knowledge worker using a chatbot to work faster. The OpenAI research paper describes something different: teams where people delegate entire days of work to agents and review the output. The distinction matters because the unit of competition has shifted. The question is no longer which employee uses AI most effectively. It is which team has figured out how to structure work around what agents can run overnight, unattended, in parallel, while everyone else is doing something else.