OpenAI reports enterprise AI use is moving from assistance toward agentic execution

OpenAI has published two enterprise AI studies arguing that business use of AI is shifting from assistance toward execution. The August 12 release combines an Enterprise Signals report based on OpenAI’s enterprise customer base with a working paper on how organizations use ChatGPT. Together, the studies describe a widening gap between companies that merely provide access to AI tools and companies that connect agents to context, permissions, tools and repeatable workflows that can complete useful work for review.

The clearest metric is output depth. OpenAI defines frontier firms as enterprise customers in the top 10 percent of AI usage each month, measured by output tokens per active user. As of June, those frontier firms generated 8.3 times as many output tokens per active user as typical firms, compared with a 2.6 times gap in January. OpenAI treats output tokens as an imperfect but useful signal: deeper, longer-running agentic workflows tend to produce more output than simple question-and-answer use.

Codex is central to the shift. OpenAI says that, as of June, Codex generated 64 percent of combined Codex and ChatGPT output tokens among enterprise customers. That does not mean every enterprise task is software engineering. Instead, OpenAI argues that agents are being asked to do longer, delegated work: using tools, creating files, producing drafts and preparing artifacts for human review. The company frames this as a move from asking an assistant how to do a task toward asking an agent to perform part of the task itself.

Advanced capabilities appear to separate heavier adopters from typical firms. Among weekly active users, OpenAI says 21 percent at frontier firms use Plugins, compared with 9 percent at typical firms. Skills show an even larger difference, with 19 percent usage at frontier firms versus 3 percent at typical firms. Internally, OpenAI says 95 percent of active employees use Plugins weekly, a figure it presents as evidence that deeper usage depends on making tools, data access and reusable instructions part of everyday work rather than treating AI as a standalone chat box.

The data also suggests agentic workflows are spreading beyond engineering. Since February, OpenAI says weekly active enterprise Codex users grew 108 times in legal, 41 times in sales, 41 times in recruiting and 26 times in marketing, compared with 5 times in engineering. The Virgin Atlantic example in the report shows that contrast: engineers used Codex to refactor legacy code in 30 minutes instead of two weeks, while product teams used ChatGPT Work to complete weeks of competitive research in hours for a five-year digital strategy process.

OpenAI also highlights a difference by seniority. Administrative data from millions of conversations showed that six months after adoption, early-career employees sent 13 more messages per week than executives. That runs against many survey-based narratives in which leaders report higher AI use. For managers, the implication is practical: the most effective workflows may already be emerging among less senior employees, and organizations may need to identify, refine and share those habits rather than assuming adoption will only cascade from the top.

The reports are self-published by a major AI vendor, so they should be read with that context in mind. The evidence describes OpenAI’s customer base and ChatGPT usage rather than the entire economy. Still, the pattern is important for buyers and builders. Model access alone is becoming a weak indicator of AI maturity. The more meaningful question is whether a company has the data infrastructure, governance, permissions, review processes and training needed to turn individual prompts into repeatable agentic workflows that can safely produce business output. That makes adoption a management problem as much as a tooling problem, because successful patterns have to travel from early power users into ordinary teams.