Most companies guess at where AI adoption is heading. In late June 2026, one company published the answer from inside its own walls. OpenAI released telemetry on how its own staff use AI — not a survey of opinions, but measured usage across every function — and the numbers describe a switch, not a ramp. By June 2026, the median employee in a research role was generating 56 times more output through agents than in November 2025. Customer Support rose 32-fold, Engineering 27-fold, Legal 13-fold. And Codex, the company’s coding agent, now accounts for 99.8% of the output tokens its workers produce across Codex and ChatGPT combined.

Read that last figure twice. Inside the company that built the chatbot, the chatbot has almost stopped being how people work. The interface won a few years ago; the agent is winning now. I’ve felt the same shift at a much smaller scale — I run an agent called Hermes on my home server, and somewhere this year I stopped opening a chat window and started handing it whole tasks. OpenAI just put a curve on the thing I’d only felt.

For the rest of us: what “agentic” actually means here

Two words in that telemetry do a lot of work, so it’s worth being precise.

A chatbot is conversational. You ask, it answers, you read the answer, you ask the next thing. Every exchange comes back to you. An agent is delegated. You hand it a goal — “find why this test fails and fix it,” “reconcile these two reports” — and it uses tools on its own: it reads files, runs commands, checks its work, and comes back when the task is done, not when the sentence is done. The paper behind the announcement draws the line exactly there: agentic systems “perform work on the user’s behalf rather than simply responding to queries.”

The second phrase is median output tokens. Tokens are the units of text a model reads and writes; output tokens measure how much the model produced. “Median” matters more than it looks — it’s not the average, which a handful of power users could inflate. It’s the middle employee. When the median researcher’s output rises 56-fold, that isn’t a few enthusiasts; it’s the typical person’s normal workday that has changed.

The shape of the switch

The four function multipliers are the headline, but the more telling number is the 99.8%. A year ago, work-related AI at OpenAI meant ChatGPT — a conversation. As of 11 June 2026, the conversation accounts for one-fifth of one percent of what its workers generate. Almost everything now flows through an agent that does the task rather than describing it.

The adoption breadth matches the depth: OpenAI reports that virtually all of its active workers use Codex every week. This isn’t a pilot with a champion and a long tail of holdouts. It’s the default tool, the way email or a code editor is the default tool — the thing you reach for without deciding to.

And it isn’t only the engineers, which is the part enterprise readers should sit with. The steepest curves are in Research and Support, not just Engineering. OpenAI also reports its fastest growth among non-developers — people whose jobs were never about writing code, now delegating multi-step work to agents as a matter of routine. The pattern isn’t “programmers got a better tool.” It’s “delegation became the default posture,” and it spread outward from the technical core.

Why this is a time machine, not a mirror

Here’s where the honesty has to come in, because it’s tempting to read those curves as a forecast for your own company next quarter. It isn’t, and the authors say so plainly.

OpenAI is, in their own words, “not representative of the typical organization today.” Of course its usage is off the charts — it employs the people who build the tools, it has zero procurement friction, no security review queue, no legacy system to integrate around, and a workforce that is unusually fluent by construction. Your organisation has all of those frictions, and they are real.

What the lab’s curve actually gives you is different and more useful: a picture of what the work looks like once the frictions are gone. The authors frame it as “a view of what agentic AI use may look like in the future, when adoption frictions are minimal.” The frictions are temporary — procurement gets sorted, security models mature, people learn the habit. The end-state behaviour is what’s on display. OpenAI isn’t a mirror of your company today. It’s a reasonably-lit window into it in eighteen months.

There’s one more caveat worth keeping honest about: measuring “agency” by tool use is imperfect. Some tool calls are trivial, and some genuinely agentic work uses few tools. The multipliers are directionally strong and internally consistent, but they’re a proxy, not a physical constant. I’d trust the shape — steep, broad, agent-dominant — more than any single decimal.

What this means for what you measure

If the steady state is delegated work, then most corporate “AI adoption” dashboards are quietly measuring the wrong thing. Counting chat sessions, seats activated, or prompts sent tracks the era that OpenAI’s own staff have already left. Those metrics will keep rising while telling you less and less about whether AI is actually doing your organisation’s work.

The metrics that survive the switch are the ones that describe delegation. How many tasks were completed by an agent end-to-end, not just started in a chat. What fraction of that agent work a human had to review or redo — the real signal of whether you can trust it unattended. How often it escalated, and whether it escalated the right things. Those numbers tell you what a chat-session count never will: not how much people are talking to the AI, but how much work they’ve handed it and gotten back.

That reframing is the whole governance problem in miniature, and it’s the one I keep circling back to in this research. The moment work is delegated rather than typed, “a human was in the loop” stops being automatically true. The safety case moves from watching each step to bounding what the agent may do — a budget it can’t exceed, an escalation path for anything ambiguous, an audit trail that can reconstruct what it saw and touched. OpenAI’s telemetry is a snapshot of the volume arriving. The unglamorous work is making sure your organisation can account for it when it does.

The lab measured its own future and published it. The generous reading is that they’ve handed everyone else a free preview. The question each organisation has to answer is whether it’s building the governance to meet that curve — or still counting chat sessions and calling it a strategy.

References