OpenAI: frontier firms now use AI agents 8.3x more than typical firms
OpenAI has published two reports examining how enterprise AI use is changing. One, Enterprise Signals, looks at agentic AI across OpenAI's enterprise customer base and draws on a sample of more than 10 million messages. The companion working paper, How Organizations Use AI: Evidence from ChatGPT, studies how adoption spreads across companies, roles and seniority levels, including a look at U.S. public companies.
Both reports describe a shift from assistance, where AI answers questions, to execution, where agents carry out the work itself and produce a result for review. The central finding is a widening gap between 'frontier' firms, defined as the top 10% of firms by monthly AI usage, and 'typical' firms, those between the 45th and 55th percentiles. Frontier firms now generate 8.3x as many output tokens per active user as typical firms, up from 2.6x in January, which OpenAI calls a threefold increase in five months.
Agentic tools are a major driver of that gap. As of June, Codex generated 64% of combined Codex and ChatGPT output tokens among enterprise customers, a figure OpenAI reads as a sign that agents are enabling more substantive, delegated work rather than simple question answering. Codex use is also spreading well beyond engineering: since February, weekly active enterprise Codex users grew 108x in legal, 41x in sales, 41x in recruiting and 26x in marketing, compared with 5x growth in engineering itself.
Frontier firms also lean harder on OpenAI's connected-tool features. Weekly, 21% of active users at frontier firms use Plugins versus 9% at typical firms, and 19% use skills versus 3% at typical firms. Internally, 95% of OpenAI's own employees use Plugins weekly, a figure the company presents as evidence that even frontier firms have adopted only a fraction of what is possible.
As a case study, the reports point to Virgin Atlantic: its engineering teams use Codex to refactor legacy code in 30 minutes instead of two weeks, while its product teams use ChatGPT Work to compress weeks of competitive research into hours, feeding directly into the airline's five-year digital strategy.
The working paper adds a financial angle: among the U.S. public companies it studied, enterprise AI adopters had stronger financial measures than non-adopters, holding more assets, employing more workers and investing more in R&D, though no dollar figures are given for these comparisons.
On seniority, administrative data drawn from millions of real conversations found the opposite of what many surveys report: early-career employees use AI more than executives, not less. Six months after adoption, early-career employees sent 13 more messages per week than executives, which OpenAI reads as a potential comparative advantage for junior staff and an opportunity for leaders to find employees with strong AI habits and spread their workflows more widely.
OpenAI frames the overall message as a call to action for enterprise leaders: connect agents to company context and tools, set clear permissions and governance, and turn effective individual workflows into shared organizational practice, since having access to the same models is not, by itself, enough to close the frontier gap.
Key facts
- Frontier firms (the top 10% by monthly AI usage) now generate 8.3x as many output tokens per active user as typical firms, up from 2.6x in January.
- As of June, Codex generated 64% of combined Codex and ChatGPT output tokens among OpenAI's enterprise customers.
- Since February, weekly active enterprise Codex users grew 108x in legal, 41x in sales, 41x in recruiting and 26x in marketing, versus 5x growth in engineering.
- Weekly Plugin use is 21% at frontier firms versus 9% at typical firms, and weekly skills use is 19% versus 3%; 95% of OpenAI's own employees use Plugins weekly.
- Six months after adoption, early-career employees send 13 more messages per week than executives, the opposite of what many surveys report.
Why it matters
Enterprise AI adoption is not converging as access spreads, it is bifurcating. The frontier-to-typical usage gap tripled in five months (2.6x to 8.3x), and OpenAI ties that widening directly to agentic tools: firms that connect agents to company context and let them execute tasks, not just answer questions, pull further ahead. The companion working paper adds that adopting firms already show stronger financial measures than non-adopters, which raises the stakes for firms still stuck at shallow, assistant-style usage.
Who it affects
Enterprise leaders and IT or AI-governance teams at large organizations, particularly outside engineering: legal, sales, recruiting and marketing teams are the functions where Codex adoption is growing fastest off a low base. It also affects OpenAI's own enterprise customer base directly, and by extension any company weighing whether individual employees' AI habits should become organization-wide workflows, since early-career staff are shown using AI more heavily than executives.
How to use it
OpenAI's stated agenda for leaders is to connect agents to company context and tools (through Plugins, which bundle reusable instructions with access to systems like a CRM), set clear permissions, review and governance, and convert individual employees' effective workflows into shared team practice. OpenAI Enterprise customers can request a customized benchmark comparing their organization against frontier firms, and the full Enterprise Signals report and the How Organizations Use AI working paper are both published for further detail.
How solid is it
Enterprise Signals draws on a sample of more than 10 million messages from OpenAI's own enterprise customer base; the companion working paper separately studies a set of U.S. public companies. Both are first-party OpenAI reports measuring usage of OpenAI's own products (Codex, ChatGPT, Plugins), based on OpenAI's internal usage data rather than independent audit. The source article does not give a methodology for how 'output tokens per active user' is computed, how firm size is controlled for, or how many firms or industries the samples cover.
Risks and caveats
OpenAI is reporting on adoption of its own products while also using the same data to promote an enterprise benchmarking offer, so the framing has an obvious commercial angle. The 'stronger financial measures' claim for AI adopters is stated only qualitatively, with no dollar figures, and the reports do not establish that AI adoption causes stronger financial performance rather than simply correlating with it. No individual researchers or authors are named anywhere in the source article.