Inside OpenAI, the Chatbot Is Almost Gone. Here's What That Says About the Next AI Skill

New data from OpenAI suggests the ground under that advice already moved.

On June 25, 2026, OpenAI published a study of how its agent tool, Codex, is being used. The paper is called "The Shift to Agentic AI: Evidence from Codex," and it was co-written by OpenAI staff alongside academics from Columbia, the University of Pennsylvania, and Duke. The standout line is about OpenAI's own people. Inside the company, Codex usage is "nearly universal and has largely replaced business usage of ChatGPT." By June 2026, Codex accounted for 99.8% of the weekly output tokens generated inside OpenAI.

Read that again. The company that builds the most-used chatbot in the world barely uses the chatbot for work anymore.

If you run a small business, you do not need to care about Codex specifically; it is a tool for writing software. What you should care about is the pattern underneath it, because that pattern is coming for every kind of work, not just engineering.

The change is from asking to handing off

Here is how OpenAI describes what an agent does. It "changes the unit of knowledge work from single interactions to delegated, long-horizon tasks."

Plain version: a chatbot answers one question at a time. You type, it replies, you type again. An agent takes a whole job and goes off to do it. Minutes. Sometimes hours. You check the result instead of guiding every step.

Think about the difference in your own week.

A chatbot moment is asking for a subject line for an email. A delegated moment is handing over a whole job: go through last month's invoices, find the clients who paid late, and draft a polite follow-up to each one. One is a single question. The other is a task you would normally hand to a person.

OpenAI's data shows people are starting to hand over the bigger jobs. In May 2026, among individual Codex users, 80.6% submitted at least one request the study estimated would take an experienced human more than thirty minutes. 70.2% sent at least one request that would take more than an hour. And 25.6% sent at least one request estimated at more than eight hours of human work.

The eight-hour number is the one that moved fastest: since the start of 2026, the share of users sending at least one task that big has grown nearly tenfold.

The pattern holds across the data: the work people hand over keeps getting bigger.

This is not a developer story anymore

It would be easy to wave this off. Codex is a coding tool, you are not a coder, end of story.

The data says the opposite. The fastest growth is coming from people who are not engineers.

OpenAI reports that since August 2025, use of the tool by non-developers rose 137 times for individual users and 189 times for organizational users. Inside OpenAI, non-developer use grew 12 times. The steepest curve belongs to the people who were never the target audience.

Inside OpenAI, usage growth breaks down by department between November 2025 and June 2026. Research use grew 56 times. Customer support grew 32 times. Legal grew 13 times. Those roles exist in your business too, or you cover them yourself when you wear all the hats. They are nothing like software teams.

One more data point, attributed carefully. OpenAI reports that by June 2026 the median employee in a legal role was generating 13 times more monthly output tokens than in November 2025, and the median researcher more than 50 times as many. A quick honesty note here: output tokens measure how much the tool produced, not how much more got done. A 13 times jump in tokens is not a 13 times jump in productivity, and OpenAI does not claim it is. Treat it as a sign of how much work people are routing through agents, not a promise of results.

What this means for someone who feels behind

If you have been trying to get good at writing the perfect prompt, that effort was not wasted. But the skill that is growing in value is a different one.

It is the skill of taking a multi-step task, handing it to an agent, and checking the work that comes back.

That sounds small. It is not. Most owners have never written down how a task in their business gets done, step by step. The agent cannot run a process you have never defined. So the hard part comes before the handoff. You map the job first, then let the tool carry it.

A few signs from the data that this is where things are going:

• More than 10% of Codex users manage three or more agents running at the same time at some point each week. OpenAI employees at the 99th percentile recently ran about 71 hours of agent work in a single average day, several agents at once. One person, many tasks running in parallel.
• 26.6% of users use "skills," which let someone save and share instructions for a complicated workflow so it can be reused. That is the same idea as writing down how your business does a thing once, then running it again and again.

This is exactly the muscle The 7 Levels of AI Proficiency is built to teach. The early levels are about using a tool well in the moment. The higher levels are about designing the workflow first, then building or directing something to run it. A chatbot lives at the bottom of that climb. Delegating a real task and managing the result sits further up.

You do not jump to the top in a week. You climb one level at a time, on work you already do.

The honest caveat

One thing to keep in mind before you treat any of these numbers as gospel.

This is OpenAI measuring its own tool, mostly on its own staff, and reporting the results. The study used "an automated, privacy-protecting pipeline," and three outside academics co-authored it, which is good. But it is still a single company's internal data, not a neutral survey of the whole market. Take it as a strong signal of where work is heading. Do not take it as a census of where every business is today.

The direction, though, is hard to argue with. The people closest to this technology stopped using it the old way.

Related reading: Level 5: The Captain (Design Thinker).

Sources

1. How agents are transforming work (OpenAI)
2. The Shift to Agentic AI: Evidence from Codex (Johnston, Holtz, Richmond, Ong, Tambe, and Chatterji)

Harrison Painter

Executive AI Advisor. Founder, LaunchReady.ai and AI Law Tracker.

Harrison is an Indiana AI Advisor who helps business owners and executives get their time back by building AI systems that run the work for them. Nearly 20 years in business and author of You Have Already Been Replaced by AI. Creator of The 7 Levels of AI Proficiency.

Connect on LinkedIn