June 29, 2026. While the headlines focused on new models and government restrictions, OpenAI quietly published the more useful document for business owners: a research paper on how AI agents are actually changing work. Titled How agents are transforming work, it measures a year of real usage of Codex, OpenAI's agentic tool, and the numbers make one thing clear. The unit of AI work has shifted from short chats to delegated, multi-hour tasks, and the people adopting fastest are not developers.
What the data shows
- Agents do long work now. By May 2026, 80.6 percent of sampled users made at least one Codex request estimated to take a person more than 30 minutes, 70.2 percent made one over an hour, and a quarter made one over eight hours.
- Non-developers are the fastest-growing users. Since August 2025, non-developer usage rose 137 times among individuals and 189 times among organizations, outpacing developer growth.
- It became the default tool. Inside OpenAI, every department, including legal, finance, and recruiting, now uses Codex as its primary AI tool, accounting for about 99.8 percent of weekly output tokens company-wide.
- Heavy users run agents in parallel. By June 2026, the most active users generated more than 60 hours of agent work per day, spread across multiple agents running at once.
What it means for operators
OpenAI is describing its own staff, who have unusually good tools, but the direction is the signal. The competitive gap in 2026 is no longer between businesses that use AI and those that do not. It is between teams that delegate hours of real work to supervised agents and teams still treating AI as a chatbot that answers one question at a time. The encouraging part for small and mid-sized businesses is that the fastest adoption is coming from non-developers, which means the barrier is workflow design, not engineering headcount. The move is to find one or two repeatable tasks that currently take a person more than 30 minutes, such as lead research, proposal drafting, data cleanup, or reporting, and turn them into supervised agent runs you can measure and trust. That is the heart of our AI automation work, and pairing it with the discipline to scope and ship is what separates pilots that stall from agents that earn their keep.
This pairs with the cheaper model tiers OpenAI just introduced. Lower-cost models are the fuel, and proven agent workflows are the engine. If you want help turning a manual process into a reliable agent, our AI engineers do exactly this.
Frequently Asked Questions
The paper, How agents are transforming work, measures a year of Codex usage. It found that agents are increasingly used for long tasks, with 80.6 percent of sampled users making a request estimated to take a person over 30 minutes by May 2026, and that non-developer adoption grew fastest, up 137 times for individuals since August 2025.
Not necessarily. The fastest-growing group in the data is non-developers, which suggests the main barrier is designing the right workflow rather than writing code. Many high-value automations, such as research, drafting, and reporting, can be built and supervised without a large engineering team.
A chatbot answers one prompt at a time in a short exchange. An agent works independently for minutes or hours, calling tools, interacting with systems, and iterating toward a result. The OpenAI data shows usage shifting from the first pattern to the second, where more real work gets delegated.
Pick one or two repeatable tasks that take a person more than 30 minutes, define the steps and a way to verify the output, and run them as supervised agent workflows before expanding. Start with widely available, lower-cost models, measure the hours saved, and grow from there.