On June 25 OpenAI did a rare thing: it published its own internal numbers on how AI agents changed work inside the company. Not forecasts, not marketing, just data. And it's telling. Back in August 2025 the average OpenAI worker spent less than 10% of their tokens on the Codex agent, with ChatGPT still the default tool. Now Codex is the primary work tool in every department, including legal and recruiting, and it accounts for 99.8% of all the company's weekly output tokens.
This isn't a story about "AI getting smarter." It's a story about the unit of work changing: from a short question in a chat to a delegated task an agent runs for hours.
At Gless we put AI agents into client processes every week, so we read reports like this with one question: what here is a signal a normal company can use, and what's a frontier artifact you can't reproduce by buying a license. Let's go through the numbers. What OpenAI's data showed, what the real shift is, and where to actually start.
What OpenAI's data showed
In short, OpenAI measured how its own staff, plus individual and organizational users, moved from chatbots to agents over a year. The report itself is economic research on the Codex agent's potential at the frontier, published on June 25, 2026.
The headline numbers for OpenAI itself:
- Before August 2025 the agent ate less than 10% of the average worker's tokens. By the time of the report Codex had become the primary tool in every department, not just engineering but also Legal, Finance, and recruiting.
- The agent now accounts for more than 85% of the average worker's output tokens. And since active agent users burn more tokens than non-users, the company-wide share is even higher: 99.8% per week.
- Among 28-day active users, the share working through Codex is 97.9% inside OpenAI, 17.3% at client organizations, and just 0.7% among individual users.
That last line matters, and we'll come back to it. 97.9% versus 0.7% is the gap between a company that builds the tool and everyone else. But the direction of travel is the same for all of them: from chat to agent.
The real shift: from chat to an agent that works for hours
The most interesting part of the report isn't adoption percentages, it's what people started handing the agent. The unit of work grew. People used to ask AI for a short answer; now they delegate a multi-hour task.
The figures for individual users by May 2026:
- 80.6% gave Codex at least one task that would have taken a person more than 30 minutes.
- 70.2% gave it a task longer than an hour.
- 25.6% gave it a task longer than eight hours.
Nearly a quarter of all Codex requests are work that would take a person more than an hour. And among the heaviest users inside OpenAI the count runs into dozens of hours: by June 2026 the top 1% regularly generated more than 60 hours of agent turns per day, spread across several parallel agents.
Parallel agents are a signal of their own. One person can't physically run ten things at once, but orchestrating ten agents, each carrying its own task, is doable. What changes isn't the speed of a single task but a worker's throughput: one person covers a load that used to need a team.
That's the shift. Not "ask and get an answer" but "set a task and go do something else while the agent works." It's exactly the format we recently covered with an agent living inside Slack and agents running 24/7 in messengers. OpenAI simply showed, with its own data, that this is already the main way of working rather than a side experiment.
Not just for developers
For business, the most important takeaway in the report is this: non-developers are adopting agents fastest.
Since August 2025 the number of non-developers using Codex grew 137x among individual users, 189x among organizations, and 12x inside OpenAI. That outpaces growth among developers themselves. Lawyers, finance staff, and recruiters at OpenAI switched to the agent as their primary tool around April 2026, later than engineers but far faster.
And non-technical people don't turn into engineers. They just take on work that used to be bottlenecked by technical expertise: automation, data transformation, debugging, structured analysis. A telling detail from the report: more than a quarter of the tasks business-function staff did through Codex turned out to be engineering or coding. The agent lowers the cost of crossing a skill boundary and lets a person do adjacent work that previously meant pulling in a specialist.
In practice it looks mundane. A recruiter asks the agent to sort a hundred applications by their criteria and pull a short summary. A lawyer runs a batch of contracts through it and flags the discrepancies. A finance person turns a raw export into a finished report. Each of those used to mean queuing for an analyst or a developer.
And usage isn't just spreading, it's deepening. Over six months, median token consumption rose against November 2025: 56x in research teams, 32x in support, 27x in engineering, 13x in legal. That's why AI agents matter for your business, now backed by numbers from one of the most advanced companies at it.
What's true for a normal business, and what isn't
Now honestly, because the numbers are pretty but the context matters.
This is OpenAI's data: a company that builds the agent itself, operates at the frontier, and hands staff the tool with no friction and no limits. You don't buy 99.8% of tokens-through-an-agent with a license. You get there when people have a powerful tool, a low barrier to entry, and a culture where delegating to an agent is the norm. Most companies are a long way from those numbers, and that's fine.
What's a real signal here rather than hype:
- The unit of work moves from a chat question to a delegated task. That applies to any team with repeatable routine.
- Agents stopped being a developers-only story. Legal, finance, and operations teams gain just as much.
Strip away the excitement around the numbers and a simple idea remains. AI agents pay off most where the work is repeatable and there's a lot of it — not where the announcement is loudest, but where someone on your team does the same thing by hand every week. Those are the spots to look for when you size up AI agents against your own processes.
And where to be careful. "A task that would have taken a person eight hours" is a model estimate, and the report itself flags those thresholds as approximate. "The agent did the work" doesn't equal "you can send it on without review." Review, tests, and accountability for the result stay with people — we covered why AI doesn't replace engineers in detail. The higher the cost of a mistake, the more of a human has to stay in the loop.
Where to start adopting AI agents
The takeaway from the report is simple and needs no frontier budget. Take one team and one task they do by hand every week, give an agent access scoped exactly to it, and measure the result. Not "roll out AI across the whole company," but one clear routine where the effect is visible right away.
Good candidates to start with are what went first beyond engineering at OpenAI too: triaging inbound requests, drafting standard documents, assembling recurring reports from exports. Where the rules are clear and the volume is high, the effect shows up fastest.
OpenAI started with engineers and reached lawyers within a year. A normal business has a shorter path: the tools already exist, and the entry point is any team with a lot of repeatable work. If you want to figure out which of your processes you can genuinely hand to an agent, and which are better built as a reliable system, that's exactly our AI implementation services.
FAQ
What does OpenAI's research say about AI agents?
OpenAI published economic research on June 25, 2026 about how the Codex agent changed work inside the company. The core finding: staff moved from chatbots to agents, the unit of work grew from a short question to a multi-hour task, and non-developers are adopting agents fastest.
Do AI agents really replace employees?
No. Even at OpenAI, where the agent accounts for 99.8% of tokens, people stay on review, checking, and accountability for the result. The agent takes on routine and long-running tasks, but decisions and control remain with humans, especially where the cost of a mistake is high.
Are AI agents only for developers?
No, and OpenAI's data shows it most clearly. The fastest-growing use of agents is among non-developers: lawyers, finance staff, recruiters. The agent gives them automation and data work that used to require a technical specialist.
Can a normal company reproduce OpenAI's results?
Partly. OpenAI's numbers are frontier: the company builds the tool itself and gives staff unrestricted access. The direction, from chat to delegated tasks, is reproducible in any business; the 99.8% scale is not. The realistic goal is to cover a specific repeatable routine with agents.
How do you start adopting AI agents?
With one team and one repeatable task. Give the agent narrow access scoped to it, measure the time saved over a week or two, and expand if there's an effect. That's cheaper and safer than trying to "roll out AI" across the whole company at once.
If you'd like to discuss which processes in your company you can genuinely hand to AI agents, get in touch and we'll look at your case.