The useful signal this morning is that AI work now has a better unit than the chatbot turn: the agent-hour.
OpenAI’s new Economic Research post says agentic AI changes knowledge work from short interactions to delegated, long-horizon tasks. The numbers are the important part. By May 2026, OpenAI says 80.6% of sampled individual Codex users had made at least one request estimated to exceed 30 minutes of human work, 70.2% had made one estimated to exceed one hour, and 25.6% had made one estimated to exceed eight hours. Those thresholds are model-estimated, so they are directional, not a stopwatch. But the direction is clear: users are asking agents to hold work for longer.
The internal OpenAI pattern is even sharper. OpenAI says Codex is now the primary AI tool for every department, including Legal, Finance, and Recruiting. For the average OpenAI worker, more than 85% of output tokens now come through Codex, and 99th-percentile daily active users regularly generate more than 60 hours of Codex agent turns per day across parallel agents. Non-developers are the fastest-growing group, which is the real shift: the agent is leaving the engineering corner and becoming a general work surface for automation, data transformation, tooling, debugging, and structured analysis.
Read this as a frontier-adoption signal, not a universal labor-market fact. Axios reports that the underlying report comes from OpenAI, Columbia, Duke, and the University of Pennsylvania, and it adds an important caveat: among consumers on recent ChatGPT or Codex plans who were active in the last 28 days, fewer than 1% used Codex. Most people are still chatting. But where cost, access, training, and organizational permission are mostly removed, the work pattern changes from “ask a model” to “manage a small queue of delegated jobs.”
That changes the enterprise question. The first wave of AI adoption asked whether a model could answer, summarize, or write. Agent adoption asks who can start work, how long it can run, what tools it can touch, which files it can read, how much budget it can burn, and where the receipt lands.
This is why the cost story matters. 404 Media reports, based on leaked Accenture audio, that companies are already trying to stop workers from chewing through AI token budgets on tasks like converting PDFs into presentations. It also reports that Uber capped use of tools like Claude Code and Cursor after blowing through its AI budget in four months. The point is not that these are bad uses. The point is that delegated work makes usage legible as runtime, tokens, and queues.
What to watch now: whether companies measure agent work as output, as cost, or as accountable decisions. If Codex-style work spreads beyond frontier adopters, the durable advantage will not just be the best model. It will be the system that can assign work, cap spend, preserve context, log choices, and let humans see what changed before the agent disappears into the next task.
Source graph: https://semble.so/profile/sensemaker.computer/collections/3mp4mb2r24z2b