OpenAI published research showing 80% of knowledge workers are delegating tasks that take over 30 minutes to AI agents, with legal, finance, and recruiting teams crossing over to agents as their primary AI tool. Here is what the data signals for how businesses will be structured.
Your Friday reporting ritual, the 90 minutes of pulling numbers from three different systems and formatting a summary for leadership, is exactly the kind of task that new economic research from OpenAI shows is already being handed off to agents at measurable scale inside real organizations. That workflow does not disappear overnight. But the data from OpenAI's own teams suggests the window between "someone still does this manually" and "an agent does this automatically" is narrower than most business leaders are currently planning for.
Published June 25, the paper tracks how OpenAI's employees and a sample of external Codex users have been adopting AI agents over the past year. By May 2026, 80.6% of sampled individual users had made at least one agent request for a task estimated to take a person more than 30 minutes to complete. 70.2% had delegated something estimated to take over an hour. One in four users ran at least one agent task estimated to require more than eight hours of human work.
Those numbers are not projections. They describe what already happened.
The April crossover
For the first several months after Codex became publicly available, ChatGPT remained the default AI tool for work even at OpenAI. Engineers were the first to shift. By December 2025, the average engineer at OpenAI was generating the majority of their AI output through agents rather than conversational chat.
Legal, finance, and recruiting held on longer. Then in April 2026, they crossed over.
Today, the average lawyer or recruiter at OpenAI generates more than 85% of their AI output through Codex rather than chat. The crossover was not gradual once the tools were capable enough. It was fast.
For non-technical workers at the organizational level, the growth is striking: non-developer organizational users increased 189-fold since August 2025. The fastest-growing segment is not developers. It is everyone else.
The paper includes one specific data point that should land for anyone running a business operations or marketing function: over one-fourth of the work done by finance and business operations workers in Codex involved engineering or coding tasks. These are people who do not have engineering job titles, delegating work to agents that previously required technical support to execute. The line between "my job" and "the technical thing I need help with" has collapsed.
What this means for your team structure
The research documents something that rarely gets quantified: the point at which an organization decides that the agent, not the employee, is the right default for a given category of work.
That point has already arrived for legal and finance at OpenAI. The implication for CMOs, RevOps leaders, and agency operators is not that you need to immediately restructure. It is that the operational calculus around task ownership is already changing at peer organizations, and the gap between early and late adopters is compounding.
By June 2026, the heaviest Codex users were regularly generating more than 60 hours of agent runtime per day, coordinating multiple parallel agents. One person, directing the equivalent output of a team. The question for any business leader is not whether this capability exists. It is whether the workflows, the integrations, and the operational habits are in place to use it.
For a marketing ops team still doing manual attribution pulls on Fridays, this represents a concrete planning decision: build the reporting agent now, or continue paying in person-hours while competitors move their people to higher-value work.
The honest caveat
OpenAI's research is about OpenAI's employees and its own platform's users. The company has unusual advantages: early and deep access to frontier models, a culture that treats AI adoption as a performance metric, and a workforce already predisposed to experiment. The crossover timeline at OpenAI likely precedes what most companies will experience by somewhere between six and eighteen months, depending on industry and organizational readiness.
The paper also relies on LLM-based estimates to classify task duration. Those estimates are directional, not precise. What the model classifies as "over one hour" will vary in practice. The trend holds regardless of the precision of individual estimates.
And the research reflects a specific product, Codex, which skews toward technical and structured work. Adoption patterns for other agent platforms may differ in pace, though not in direction.
The bigger picture
Every productivity shift at scale has a leading indicator: the moment when a specific department stops treating a new tool as an experiment and starts treating it as the default. That moment happened at OpenAI for legal, finance, and recruiting in April 2026. It happened quietly, over a few months, and the paper captures it in percentages.
For most businesses, the same crossover is coming. It will not announce itself. One quarter, the Friday report will go out on time without anyone pulling the numbers. And nobody will be entirely sure when the last manual version was sent.