David Holtz, Assistant Professor at Columbia Business School who is currently serving as Head of AGIO Economic Research at OpenAI.
- Codex use at OpenAI grew from under 10% of workers’ AI output in August 2025 to 99.8% by June 2026, according to a new working paper.
- The shift went well beyond engineers. Legal and recruiting teams went from near-zero Codex use to nearly 90% of their AI output in about five months, helped by a company-wide push that included training and regular feedback.
- Outside OpenAI, adoption is lower and more uneven, and the researchers say the biggest payoffs will come to companies that redesign workflows, roles and review processes rather than just add agents to existing ones.
In August 2025, the average OpenAI worker generated less than 10% of their AI output on Codex, the company’s agentic AI tool. By June 2026, Codex accounted for nearly all (99.8%) of the output OpenAI workers generated across Codex and ChatGPT.
Holtz did the work as a visiting researcher at OpenAI, and co-authored findings with several OpenAI employees.
, which compared how OpenAI employees use AI agents with how business and personal-account users outside the company use them, revealed that agent use has accelerated. At OpenAI, it’s no longer just engineers taking advantage of the tool. Workers across the company are shifting from conversational chatbots to agentic tools to do their work.
In their working paper, The Shift to Agentic AI: Evidence from Codex, they say the change in AI tools could change not only the way people work, but also productivity, job design and the makeup of the workforce.
Could your company do the same? Maybe. But OpenAI had some big advantages, the researchers note.
How is agentic AI different from conversational GenAI?
Agentic AI tools are built on generative AI, the same technology behind ChatGPT. The difference is how you use them. With conversational tools like ChatGPT, you ask a question or give an instruction, and the model responds. With agentic tools like Codex, you hand off a task, and the agent carries it out on your behalf by using software, working with files and producing finished work. The line isn’t absolute. ChatGPT can take some actions on its own, and some Codex use is purely conversational.
The difference shows up in how people use each. Earlier research on ChatGPT found that nearly half of all prompts were requests for information and advice. Codex users, by contrast, mostly ask the agent to do the work itself: debugging code, analyzing data, drafting documents and more.
How was the research done?
The researchers analyzed usage data from OpenAI’s Codex collected between August 2025 and June 2026. Codex began as a coding agent, but people now also use it to draft documents, analyze data and coordinate communications.
- External personal-account users
- External business-account users
- Workers within OpenAI
The researchers used AI classifiers to sort requests by type of task, so they could see what people used Codex for without reading the underlying messages. To gauge how complex the tasks were, the team used an AI model to estimate how long each request would take an experienced human to complete, based on a small random sample of personal-account users.
What did they find?
Holtz and his fellow researchers discovered that far more people are using agentic tools than before. Across all Codex users, the number of weekly active users grew more than fivefold from January to May.
The researchers also observed measures of sophistication. Here they found that a growing number of Codex users have changed their workflows substantially. More than 10% of all users manage three or more concurrent Codex agents at some point each week. cross all Codex users, more than a quarter (26.6%) used skills, a feature that lets users share instructions for complex workflows. At OpenAI, that share was nearly universal at 96%.
Request complexity increased, too. In December 2025, about a third (35.4%) of Codex personal-account users sent at least one request for a task that would take an experienced human an hour or more. By May of this year, that share had doubled to 70.2%. Over the same period, the share requesting at least one task that would take a human more than eight hours rose from 2.1% to 25.6%.
Another measure of usage, token output, has also grown rapidly. In June of this year, the median OpenAI employee in a legal role generated 13 times more monthly —the basic units of text that AI models read and generate—across Codex and ChatGPT than they did in November of last year. The median researcher generated more than 50 times as many. In fact, output rose at least tenfold in every job function at OpenAI.
How do workers use Codex?
found that the most intensive Codex users work differently than occasional ones. They’re more likely to run longer, more complex tasks, use skills, and run several agents at once. Rather than just asking questions, they delegate work, then supervise, review and integrate what the agents produce.
To be sure, this is still new for most users. Outside OpenAI, about two-thirds of Codex users (67% of organizational users and 64% of personal-account users) don’t yet run agents in parallel. But among OpenAI employees, nearly three in 10 (28.6%) managed five or more concurrent agents at some point during an observed week in June 2026.
Did every department really make the switch?
At OpenAI, yes. Engineers moved first. By December 2025, the average OpenAI engineer did most of their AI work on Codex. By March of this year, that share topped 90%.
Other departments were slower to start, but they caught up fast. In January, Codex use in functions such as legal and recruiting was close to zero. It rose gradually to about 20% by early April, then jumped to roughly 75% within a month. As of June, the average OpenAI recruiter generates 89% of their AI output on Codex, and the average member of the legal team 88%.
What changed? The jump coincided with a company-wide push to adopt Codex across workflows, including training sessions and regular feedback loops.
Within OpenAI, Codex use also extends well beyond engineering, into research, planning, communication, data analysis, product work, recruiting and sales.
Could your company do the same?
Outside OpenAI, adoption remained lower and more uneven. For example, only 17.3% of users with organizational accounts used Codex in a 28-day period. And the average legal user at these organizations did just 1.9% of their AI work on Codex. Yet those who do adopt Codex use it heavily—enough that it now generates most (63.3%) of the AI output from organizational accounts.
OpenAI is an unusually favorable setting for agentic AI, the researchers note. Its workers face no limits on AI use, know the tools well and work in a company with strong buy-in and constant knowledge sharing.
However, a recent survey from McKinsey & Co. indicates that agentic AI is gaining traction at businesses. Four in 10 large organizations (40%) said they are now scaling AI agents, up from 27% a year earlier.
History suggests the biggest payoffs may still be ahead. Drawing on research into past technology shifts, the researchers write that “the largest productivity gains often arise when firms reorganize production around the new technology rather than merely substitute it into existing workflows.”
Consider electrification. When factories first switched from steam to electric power, many simply swapped one central engine for another. The big gains came only after they redesigned their factory floors around smaller electric motors—a process that took decades. Agentic AI could move faster, researchers suggest, because testing new workflows is relatively cheap.
What are the pitfalls?
There is a growing body of research on early AI agent adoption that indicates companies likely need to rethink how their work is allocated, reviewed and coordinated when adopting agentic AI tools.
Business leaders should also watch for these risks:
- Averages can mislead. A small group of heavy users can make adoption look broader than it is. Outside OpenAI, most Codex users still use the tool only intermittently.
- More output isn’t more productivity. Other research cited in the paper finds that big gains on individual tasks may not fully carry through to final output, because human review and other downstream steps remain bottlenecks.
- Oversight takes work. Managing agents means checking, correcting and coordinating what they produce—and that requires domain expertise.
- Tools alone aren’t enough. The value of agentic AI depends on whether organizations redesign workflows, roles and review processes around it.
- Jobs may change. The researchers say these shifts could affect team composition, hiring needs and career ladders.
Using agentic AI is not just asking a system for answers. It’s managing systems that act. The question for business leaders: Is your organization ready to manage them?
Read the working paper: The Shift to Agentic AI – Evidence from Codex, by Drew Johnston, David Holtz, Alex Martin Richmond, Christopher Ong, Prasanna Tambe and Aaron Chatterji