OpenAI just published internal usage data that should make every product team rethink their AI strategy. Codex, their agentic tool, went from less than 10% of employee token usage in August 2025 to 99.8% by mid-2026. Not among engineers—across every department, including Legal and Recruiting.
This isn't a press release about capabilities. It's a rare look at what happens when knowledge workers get low-friction access to capable agents. And the pattern is stark: once agents cross a capability threshold, adoption doesn't just grow—it explodes.
The shift from chatbots to agents is real
The core insight is simple but profound. Chatbots are for interactions. Agents are for delegation. The paper describes it as a change in "the unit of knowledge work"—from self-contained back-and-forth exchanges to "long-horizon tasks" where an agent runs independently for minutes or hours, orchestrating tool calls and iterating toward solutions.
By May 2026, 80.6% of sampled individual users made at least one Codex request estimated to exceed 30 minutes of human work. 70.2% made a request exceeding one hour. And 25.6% made at least one request estimated to exceed eight hours of human work.
Let that sink in. A quarter of users are delegating full-day projects to an agent.
The heaviest internal users are running more than 60 hours of agent work per day (99th percentile, June 2026), distributed across multiple parallel agents. This isn't experimentation—it's orchestration at scale.
Non-developers are adopting faster than developers
Here's where it gets interesting. Engineers moved first—predictable, given Codex started as a coding tool. The average OpenAI engineer shifted to majority-Codex usage by December 2025 and now generates 99% of their output tokens with Codex rather than ChatGPT.
But non-developers are moving faster. Since August 2025:
- Non-developer individual users rose 137x
- Non-developer organizational users rose 189x
- Non-developer OpenAI employees rose 12x (likely from a higher baseline)
Legal, Finance, and Recruiting crossed into majority-Codex usage around April 2026, and their transitions were faster than Engineering's gradual ramp. The average lawyer or recruiter at OpenAI now generates more than 85% of their output tokens on Codex.
This tracks with what we've been seeing in the wild—agents like Devin, Factory's Droids, and now OpenAI's own Codex are finding product-market fit outside traditional developer workflows.
The work expands to fill the agent's capacity
OpenAI shared a heat map comparing inferred employee occupations to the type of work done with Codex. Engineering and coding dominate for Data Science and Research roles—no surprise there. But here's the tell: over one-fourth of Codex work done by business functions was engineering or coding.
Agents lower the cost of crossing task boundaries. A finance analyst who used to wait on Engineering for a data transformation script can now delegate it to Codex. A recruiter can automate candidate outreach workflows without filing a ticket. A lawyer can write tooling for contract analysis.
This isn't about non-technical people becoming engineers. It's about adjacent work that used to require specialized technical support becoming accessible. The paper calls it "expanding the horizon of potential productive work."
Token growth is exponential, not linear
Median active users at OpenAI saw token output explode:
- Research: 56x higher by June 2026 vs. November 2025
- Customer Support: 32x
- Engineering: 27x
- Legal: 13x
This isn't marginal productivity improvement. It's a phase change in how work gets done. And it maps directly to capability improvements—as Codex leveraged stronger models and new product features, usage didn't plateau. It accelerated.
What this means for everyone else
OpenAI is a frontier user with low-friction internal access to frontier tools. But the pattern they're documenting is the future every knowledge-work org is heading toward.
Three takeaways:
First, agents will eat chatbots from the bottom up. ChatGPT is still dominant in consumer usage, but inside organizations with access to capable agents, the transition is fast. Once people can delegate instead of prompt, they don't go back.
Second, non-developer adoption is the unlock. The market for "coding assistants" is large but bounded. The market for "knowledge work agents" is the entire labor economy. The 137–189x growth in non-developer users is the signal.
Third, the bottleneck isn't demand—it's capability. Users didn't slowly ramp into 60-hour agent workdays. They jumped as soon as the tools could handle it. The implication: every capability improvement unlocks a new tier of delegation.
Open questions
The paper doesn't answer everything. How does Codex handle failure modes at scale? What's the distribution of task success vs. abandonment? How much human review are these "delegated" tasks actually getting?
And the big one: how much of this productivity translates to output vs. process? If Legal is generating 13x the tokens, are they closing deals faster, or just iterating more with an agent in the loop?
These are empirical questions, and OpenAI has the data. I hope they publish more.
The agent future is already here
Simon Willison likes to say the future is already here, it's just not evenly distributed. This paper is a receipt.
Inside OpenAI, the chatbot era is over. Agents account for 99.8% of weekly output tokens. Non-technical employees are running multi-hour delegated workflows. The unit of work has changed.
The rest of us are just catching up.