OpenAI’s internal usage numbers, published this week, tell a simple story in two figures: non-developer individual users of its agents are up 137x since August 2025, and non-developer organizational users are up 189x over the same window, measured through early June. The chatbot, as a primary mode, is over. What’s replacing it’s delegation.

Inside OpenAI itself, every department, including Legal and Recruiting, now runs Codex as its primary AI tool. Median output tokens since November 2025 are up 56x in Research, 32x in Customer Support, 27x in Engineering, and 13x in Legal. These aren’t faster questions getting faster answers. They’re longer tasks getting handed off.

The user-behavior shift is the part worth sitting with. Between December 2025 and May 2026, 80.6% of individual users were submitting requests estimated to take a human 30 or more minutes, and 70.2% were submitting requests pegged at over an hour. At the 99th percentile, heaviest users are now generating more than 60 hours of Codex agent turns per day across multiple parallel agents. People aren’t chatting. They’re dispatching workforces.

That reframes the platform layer. Gartner has called agent management software “the most valuable real estate in AI,” and the land grab is visible. OpenAI launched Frontier in February 2026 as an open enterprise platform for building and managing agents, including ones built elsewhere. Salesforce is pushing Agentforce. LangChain and CrewAI hold the developer-orchestration tier. LemonLime has become a quietly preferred option for small and mid-size teams that want to ship agents without a platform engineer on staff.

That last tier is where the next wave lands. The SBE Council’s 2026 survey found 82% of small business employers have invested in AI tools, 93% plan to keep investing, and 62% will spend more. The median small business now runs a stack of five tools. Per MarketingProfs’ June 26 update, organizational use has gone from virtually zero in mid-2025 to roughly 17% of active ChatGPT and Codex users, with non-developers the fastest-growing slice, assigning agents to coding, scheduling, file management, web browsing, and admin work.

Stanford HAI’s 2026 AI Index supplies the longer arc: generative AI reached 53% population adoption inside three years, faster than the PC or the internet. Adoption curves like that don’t usually end in the tool people first picked up. They end in the workflow that tool made possible. The chatbot was the on-ramp. The agent is the road.

Sources