How OpenAI Got Its Lawyers and Recruiters Using AI Agents in Five Months

Inside OpenAI, it took engineers several months to move most of their AI work onto agents. Lawyers and recruiters started later, and once the company pushed, they caught up in a matter of weeks.

That is the most interesting finding in a new working paper on how people use OpenAI's Codex agent, summarized by MIT's Initiative on the Digital Economy on October 20, 2026. It is not mainly a story about coding. It is about what happened when non-technical teams were given an agent and a real push to use it, and what the rest of us can and cannot copy.

What actually happened inside OpenAI

The paper, The Shift to Agentic AI: Evidence from Codex by Drew Johnston, David Holtz, Alex Martin Richmond, Christopher Ong, Prasanna Tambe and Aaron Chatterji, analyzes Codex usage from August 2025 to June 2026. It compares three groups: personal-account users, organizational-account users and OpenAI's own staff. Several authors work at OpenAI, which is worth keeping in mind.

The headline inside the company is a near-total switch. In August 2025, the average OpenAI worker produced less than 10% of their AI output on Codex. By June 2026, Codex accounted for 99.8% of the output staff generated across Codex and ChatGPT.

The department-level timeline is the useful part:

• Engineers went first. By December 2025 the average OpenAI engineer did most of their AI work on Codex, and by March 2026 that share topped 90%.
• Legal and recruiting started near zero. In January 2026, Codex use in those functions was close to zero.
• Then the curve bent. Use rose gradually to about 20% by early April, then jumped to roughly 75% within a month.
• By June, the average recruiter generated 89% of their AI output on Codex and the average member of the legal team 88%.

According to the MIT summary, that jump coincided with a company-wide push to adopt Codex across workflows, including training sessions and regular feedback loops.

The volume of work changed too. In June 2026, the median OpenAI employee in a legal role generated 13 times more monthly output, measured in tokens, than in November 2025. The median researcher generated more than 50 times as much, and output rose at least tenfold in every job function.

The outside world looks very different

Outside OpenAI, the picture is much more uneven. Among users with organizational accounts, only 17.3% used Codex in a given 28-day period, and the average legal user at those organizations did just 1.9% of their AI work on Codex. About two-thirds of organizational users (67%) do not yet run agents in parallel, while nearly three in ten OpenAI employees (28.6%) managed five or more agents at once during an observed week in June.

People who do adopt it use it heavily, though. Codex now generates most (63.3%) of the AI output from organizational accounts.

A separate survey helps explain the gap. Workday's September 2026 survey of nearly 6,000 workers found that only 38% of employees understand the term "AI agent" well enough to explain it confidently. 88% of senior leaders have used agentic AI to complete tasks, compared with 54% of individual contributors. And only 32% of individual contributors trust their employer's approach to AI, versus 66% of senior leaders.

So the typical company is not short of agents. It is short of the understanding, permission and habits that turned OpenAI's legal team from near zero to 88%.

Why OpenAI is not a fair comparison

The researchers are careful about this. OpenAI is an unusually favorable setting: staff face no limits on AI use, know the tools well, and work in a company with strong buy-in and constant knowledge sharing. Most organizations have none of those advantages by default.

The paper also lists cautions that apply to everyone:

• Averages can mislead. A small group of heavy users can make adoption look broader than it is.
• More output is not more productivity. Gains on individual tasks may not carry through if human review and other later steps remain the bottleneck.
• Oversight takes work. Managing agents means checking, correcting and coordinating what they produce, and that requires domain expertise.
• Tools alone are not enough. The value depends on redesigning workflows, roles and review processes around the agents.

That last point echoes the history the authors cite: when factories first electrified, many simply swapped one central engine for another. The big gains came only after they redesigned the factory floor around the new motors.

Sign up for our newsletter

Get the latest updates, product launches, and helpful tips delivered straight to your inbox.

Email addressSubscribe

What does transfer: four habits from the legal and recruiting jump

Strip away OpenAI's special advantages and four habits remain that any company can copy.

1. A deliberate push, not a license. The legal and recruiting curve moved when the company made adoption a priority across workflows. Buying seats and waiting is not the same thing.
2. Training sessions tied to real work. Generic "intro to AI" sessions do little for a recruiter. Sessions built around that team's own tasks, such as drafting outreach, summarizing candidate packets or preparing interview loops, give people something to use the next morning.
3. Regular feedback loops. A standing weekly slot where people share what worked, what broke and what the agent got wrong turns individual experiments into team knowledge.
4. Shared, written workflows. At OpenAI, 96% of Codex users used skills, a feature for sharing instructions for complex workflows, compared with 26.6% of all Codex users. Writing down how the team does a task, once, so everyone's agent does it the same way, is one of the clearest differences between OpenAI and everyone else.

A starting playbook for a non-technical team

For a legal, recruiting, finance or operations team that wants to try this:

• Pick two or three recurring tasks that take real time and are easy to check, such as a first-pass review of a standard contract against a checklist, or turning interview notes into a structured candidate summary.
• Write the instructions down as a shared workflow, including what a good result looks like and what the agent should never do on its own.
• Name a reviewer. Every output gets checked by someone who knows the domain, at least at first.
• Meet weekly for a month. Review a few real examples together, fix the instructions and decide what to add next.
• Measure the work, not the tool. Track turnaround time, rework and reviewer corrections rather than how often people open the AI tool.

What this means going forward

The OpenAI data is a ceiling, not a forecast. Most companies will not hit 88% of legal work running through an agent in five months, and they should not try to. But the shape of the curve is encouraging: non-technical teams were not the problem. Once they had training, feedback and shared ways of working, they adopted faster than the engineers had.

For most organizations, the gap to close is not technology. It is the plain, unglamorous work of teaching teams what agents can do, writing down how the work should be done, and giving people permission and time to change how they work.

Frequently asked questions

Is Codex only for programmers? It began as a coding agent, but the paper finds people also use it to draft documents, analyze data and coordinate communications. Inside OpenAI, use spread into research, planning, product, recruiting, legal and sales.

Did OpenAI's legal team really do 88% of its AI work through an agent? That is the paper's figure for the average legal team member as of June 2026, measured as a share of AI output across Codex and ChatGPT. It describes how they used AI, not what share of all legal work was done by AI.

What is the single most transferable lesson? Shared, written workflows plus regular feedback. Those habits are what separated OpenAI's staff from typical users, and neither requires special technology.