OpenAI's Own Data Proves Agentic AI Is Already Running the Business

The Number That Should Stop You Mid-Scroll

OpenAI has published internal economic research showing that Codex agents now account for 99.8% of weekly output tokens generated inside the company. That is not a projection. It is a measured internal figure, published on OpenAI's official blog on 25 June 2026, alongside a full economic research paper.

Legal, Finance, and Recruiting crossed into majority Codex use around April 2026. Engineering got there first, by December 2025. Every department has now followed.

What Happened, and When

Codex launched publicly in mid-2025. Through August 2025, the average OpenAI employee spent less than 10% of their tokens on Codex. ChatGPT was still the default. Then something shifted.

Between August 2025 and June 2026, non-developer individual users grew 137-fold. Non-developer users inside organisations grew 189-fold. The catalyst was not a single release; it was cumulative capability improvements that made longer, harder, more ambiguous tasks tractable for an agent.

The Timeline at a Glance

Source: OpenAI Economic Research Paper: The Shift to Agentic AI, published 25 June 2026.

How It Actually Works

The core mechanism is task duration. A chatbot interaction is self-contained: you ask, it answers, you move on. An agent takes a delegated objective and works independently, calling tools, iterating, and returning with a completed output. OpenAI's data shows that by May 2026, 25.6% of individual users had submitted at least one request estimated to represent more than eight hours of human work.

The heaviest internal users at OpenAI were running more than 60 hours of parallel agent turns in a single day by June 2026. That is not one agent doing one task. It is orchestrated, concurrent agent labour across multiple workstreams simultaneously.

Non-developers are not using Codex the way engineers do. Lawyers and recruiters are using it for automation, structured data transformation, and technical execution they previously had to outsource or deprioritise. The job description boundary is dissolving.

What This Means for Growth Teams in 2026

If you are a CMO, VP of Growth, or in-house search lead, the OpenAI data answers a question you have probably been asked in a planning meeting: at what point do agents become load-bearing infrastructure rather than a productivity experiment? The answer, from a company running at frontier, is that it happened between December 2025 and April 2026.

The practical implications are three things happening in parallel. First, task scope expands: agents are being assigned work that previously required a specialist or was simply not done. Second, team structure shifts: headcount planning based on task volume needs to account for tasks that can be fully delegated. Third, content and campaign operations change shape: if Research at OpenAI grew 56x in output tokens in six months, the throughput assumptions behind your editorial calendar are probably wrong.

The question is not whether to move to agent-first workflows. The question is which workflows you redesign first. OpenAI's own sequencing, engineering first, then operations functions, then legal and compliance, is a reasonable model to pressure-test against your own org.

For context on how OpenAI is building the infrastructure to support this scale of agentic work, see what OpenAI's custom inference chip means for cost and speed at scale. And if you want to understand how the broader product direction connects to this operational shift, the June 2026 ChatGPT product sprint analysis covers the signals worth tracking.

The Action to Take Now

Audit one high-volume, repeatable workflow this quarter: content briefing, campaign reporting, competitor analysis, or structured research. Map the task to a time estimate. If it exceeds 30 minutes of human work, it is a candidate for agent delegation today, not eventually. OpenAI's data shows that is where adoption accelerates fastest.