Agentic AI is changing not just what people ask AI to do, but how they work with it. Instead of treating AI as a chatbot that answers one question at a time, more users are delegating complex tasks that can run independently for minutes or even hours. That shift is becoming visible both inside OpenAI and among external users of Codex.
According to OpenAI’s report, nearly a quarter of all Codex requests now involve work estimated to take a human more than an hour to complete. By May 2026, more than 80% of sampled individual users had delegated at least one task equivalent to over 30 minutes of human work, while more than 70% had assigned work exceeding an hour. More than a quarter had trusted Codex with tasks estimated to require over eight hours of manual effort.
The trend reflects a broader change in how AI is used. Rather than replacing isolated searches or coding snippets, agentic systems are increasingly handling complete workflows. They can call tools, inspect results, revise their own work and continue operating toward a goal without constant user intervention.
The biggest surprise isn’t that software engineers embraced the technology first. It’s how quickly non-technical teams followed. Legal, finance and recruiting all reached the point where Codex became their primary AI tool during 2026. At OpenAI, engineers now generate almost all of their AI output through Codex, but lawyers and recruiters have also shifted more than 85% of their AI usage to the agent.
Adoption outside engineering has accelerated even faster than among developers. Since August 2025, the number of non-developer Codex users increased 137-fold among individual users and 189-fold among organizational customers. The report suggests this isn’t because everyone suddenly became a programmer. Instead, agents increasingly perform technical work on behalf of people whose primary expertise lies elsewhere.
That is evident in everyday tasks. Business teams use agents for automation, data transformation, debugging, structured analysis and building internal tools—activities that previously required dedicated engineering support. More than one quarter of Codex-generated work performed by employees in business functions was classified as engineering or coding.
Another interesting signal is how heavily advanced users rely on parallel execution. By June 2026, users at the 99th percentile inside OpenAI regularly accumulated more than 60 hours of daily agent runtime by running multiple independent agents simultaneously. The metric doesn’t imply people work 60-hour days; it highlights how delegated work scales when several agents operate in parallel.
Perhaps the most important takeaway is that AI adoption is no longer measured simply by the number of chat conversations. As agent capabilities improve, users appear willing to hand over increasingly complex, cross-functional tasks instead of using AI as an enhanced search engine or coding assistant.
If this pattern continues beyond early adopters, the next phase of workplace AI may be defined less by conversations with chatbots and more by managing teams of autonomous software agents.