For the past few years, most companies have treated AI like a very quick research assistant. You ask a question, read the answer, and return to your work. Enterprise AI agents work differently. Recent research from OpenAI suggests the most advanced organizations now hand over entire tasks and simply review what comes back. The distance between those companies and everyone else is growing fast.
From Giving Answers to Getting Work Done
Ask an assistant how to build a presentation and it will explain the steps. Ask an agent and it will pull material from several sources, assemble the deck, and send it back for your approval. That difference is now visible in the numbers.
By June 2026, Codex, OpenAI’s agentic tool, accounted for 64% of the output tokens its enterprise customers generated across Codex and ChatGPT combined. Some of that share is simply how agents behave, since they run longer, multi-step jobs that produce more output. So the figure reflects both how often agents are used and how large the tasks are.
The Frontier Gap Is Widening
OpenAI ranks its enterprise customers every month by output tokens per active user. The top 10% are called “frontier” firms, while typical firms sit around the middle.
In January, frontier firms produced about 2.6 times more output per user than typical ones. By June, that multiple had jumped to 8.3. It isn’t just a tech-sector story either, because the pattern shows up across industries and company sizes.
A companion working paper looked at U.S. public companies and found that enterprise adopters tended to be bigger, employ more people, and spend more on R&D than non-adopters. The lesson is that buying access to a model isn’t enough. The companies pulling ahead also invest in training, shared workflows, data infrastructure, and governance.
What Frontier Firms Do Differently
Agents are only as useful as the context and tools they can reach, and frontier firms are more likely to use features that provide both.
Plugins are the clearest example. They combine skills, meaning reusable instructions, with connections to company apps and data. Picture a sales plugin that pairs a team’s playbook with its CRM. An agent could then draft a tailored proposal using up-to-date customer details and past deals.
Weekly usage shows how big the divide is. About 21% of active users at frontier firms use plugins, compared with 9% at typical firms. For skills, the gap is wider still, at 19% versus just 3%. Even so, frontier firms have room to grow. At OpenAI itself, 95% of active employees use plugins every week.
Agents Are Moving Beyond Engineering
Software development was the first home for agentic AI, but it’s no longer the fastest-growing one. Since February, weekly enterprise Codex users have grown 108-fold in legal, 41-fold in both sales and recruiting, and 26-fold in marketing. Engineering grew 5-fold. These jumps start from small bases, but the direction is unmistakable.
Virgin Atlantic offers a useful snapshot. Its engineers used Codex to refactor legacy code in 30 minutes instead of two weeks. Its product teams squeezed weeks of competitive research into a matter of hours, which helped inform a five-year digital strategy.
Your Most Skilled AI Users May Be Junior
You might assume executives are the heaviest AI users. OpenAI’s administrative data, drawn from millions of conversations, suggests otherwise. Six months after adoption, early-career employees were sending 13 more messages per week than executives.
That’s good news for leaders. The people with the strongest AI habits may sit several levels below them. Finding those employees and making their workflows visible is a practical way to spread good practice across the company.
How Leaders Can Close the Gap
Most companies have access to the same models, so the difference comes down to execution. The research points to three priorities.
First, connect agents to real context. Give them access to the data, tools, and playbooks that valuable work depends on. Second, set clear guardrails. Define permissions, review steps, and governance so people can trust what agents produce. Third, turn individual wins into shared practice by packaging effective workflows so whole teams can use them.
It also pays to look beyond engineering. Legal, sales, recruiting, and marketing teams are adopting agents quickly, and their repetitive, document-heavy work is well suited to delegation.
A Word of Caution
This research comes from OpenAI and describes OpenAI’s own customers, so it reflects one vendor’s view of the market. Output tokens are also a rough proxy for how deeply people use AI, not a measure of business value. A company can generate enormous output without improving results. Pair usage numbers with outcomes such as time saved, quality, and cost.
Conclusion
Enterprise AI is shifting from helping people think to helping them finish. The firms moving fastest are connecting agents to their data, putting governance in place, and sharing what works across teams. For everyone else, the gap can still be closed, but the data suggests it is getting wider every month.
