Summary: Two OpenAI studies describe enterprise AI moving from assistance to execution: frontier firms produced 8.3 times the output tokens per active user of typical firms by June, up from 2.6 times in January. Codex accounted for 64% of enterprise output tokens across Codex and ChatGPT, while Plugins and skills were used more heavily at frontier firms. OpenAI recommends wiring agents into company context and tools with governance and human review.
OpenAI has just published two studies that put a figure on a shift enterprises are living through: AI moving on from answering questions to actually carrying out work. The headline comparison is blunt. Measured per active user, frontier firms—the top 10% by usage—now turn out 8.3× the output tokens of typical firms; in January that same ratio stood at 2.6×. Depth of use is what the multiple captures, and it holds across industries and company sizes alike. What follows for anyone building products or running AI programs is a fairly concrete playbook: hook agents into company context and tools, put governance and human review in place, and make one team’s working method everybody’s standard.
The Shift: From Asking to Doing
Assistance and execution are different phases of the same job, and enterprise AI is crossing from the first into the second. An assistant is something you think alongside; an agent is something you hand work to. Tools such as ChatGPT Work and Codex can pick up other tools, write files, and return finished material for a person to check. The request changes shape with it: rather than asking AI how a presentation should be put together, a worker can hand over the job—find the relevant material across sources, then draft the deck—and let the agent do it.
Usage figures tell the same story. Among enterprise customers, 64% of all output tokens across Codex and ChatGPT came from Codex by June. Longer, multi-step tasks simply emit more text, which means that 64% carries two signals at once: how often Codex gets used, and how much each of those tasks gives back.
Read this as a shift in what buyers expect. Execution changes the ask from an answer to a finished deliverable. A product that wants to serve that has to give agents a way to drive tools, reach company data, and leave behind something a human can sign off on.
The Frontier Gap Is Widening
Every month, OpenAI sorts its enterprise customers by output tokens per active user. The top 10% for that month count as frontier firms; typical firms are those sitting between the 45th and 55th percentiles. By June, the frontier group’s per-user output was 8.3× the typical group’s—three times the 2.6× gap recorded in January. And because the same pattern repeats across industries and company sizes, heavy AI use clearly isn’t a technology-sector habit.
A second paper, the working paper How Organizations Use AI: Evidence from ChatGPT, looks at U.S. public companies and reports stronger financial measures among adopters than non-adopters: bigger asset bases, larger workforces, heavier R&D spending. Taken as a pair, the reports argue that licences and access, by themselves, will not scale AI. What carries adoption wider and deeper is spending on things adjacent to the model—ongoing employee learning, workflows people share, data infrastructure, governance.
For a product team this separates two things people tend to conflate: having every feature, and being deeply adopted. Tools alone won’t get you there; customers have to form habits. Ask what your product does to make use routine, to make workflows shareable, and to fit the data and governance the customer already runs.
Plugins and Skills: Where Frontier Firms Lead
An agent is only as good as the context and tools it can reach. Plugins package what a particular workflow requires, wrapping skills—instructions that can be reused—around apps that open the door to company data, tools, and actions. Take a sales Plugin: pair the team’s playbook with CRM access, and the agent can pull current customer details and earlier proposals together into a response the rep reviews before it goes out.
Frontier firms are well ahead on this front. Among weekly active users, Plugins are used by 21% and skills by 19% at frontier firms, versus 9% and 3% at typical ones. Yet even the leaders are nowhere near the ceiling—OpenAI’s own internal numbers show what headroom looks like, with 95% of active users reaching for Plugins each week.
There is a lesson in how Plugins are put together. Combining context with tools is precisely what turns an agent from a talker into a doer. When you design, ask how a user can package their reusable instructions alongside access to the systems and data that are theirs.
Agents Move Beyond Engineering Into Knowledge Work
Engineering came first, but the steepest growth in Codex use has moved on to other knowledge-work functions. Weekly active enterprise Codex users are up 108× in legal since February, 41× in sales, 41× in recruiting, and 26× in marketing—next to just 5× in engineering.
Virgin Atlantic is one place where the pattern is visible end to end. Legacy refactoring that used to eat two weeks of an engineering team’s time now takes 30 minutes with Codex, while its product teams compress weeks of competitive research into hours, work that feeds the airline’s five-year digital strategy.
So don’t market agents solely as an “engineer’s assistants” tool. Work in legal, sales, recruiting, and marketing tends to lean hard on internal knowledge and connected systems, which is exactly the ground where an agent earns its keep. Design for those functions as deliberately as you design for developers.
Early-Career Employees Are the Heaviest Users
Surveys keep telling us that senior people are the heaviest AI users. Administrative data covering millions of conversations says otherwise: once adoption is six months old, early-career employees are sending 13 more messages a week than executives are.
Leaders can read that as an opening rather than a curiosity. Find whoever has the strongest AI habits, give their workflows visibility, and let the practice travel to every level. For product builders it is a corrective: your heaviest users may sit nowhere near where you assumed. Build for early-career workers who adopt fast, and give them a way to share and scale what they invent.
How Leaders Can Narrow the Gap
The models are available to everyone; what separates frontier firms is how quickly and how widely they put those models to work. For a leader, the opening is to shrink the frontier gap—push agentic workflows past engineering, and turn single-team wins into something the whole organization repeats.
In practice, that has two halves. First, wire agents into company context and tools behind clear permissions, governance, and human review. Second, take the workflows that already work for an individual and promote them into team practice. Slogans narrow nothing; this path does.
The takeaway for product builders: AI’s value lives less in the model than in how deeply an organization folds it into daily execution. Adding features isn’t the next move—getting customers from occasional use to a weekly habit is, and so is making agents finish real work. OpenAI’s full report, Enterprise Signals, goes deeper on industries and functions.