On August 12, 2026, OpenAI released two complementary studies: Enterprise Signals, which examines usage patterns among enterprise customers, and the academic paper How Organizations Use AI: Evidence from ChatGPT, which analyzes how AI adoption spreads within organizations. The data shows that enterprises in the top 10% by usage generate 8.3 times more tokens per active user than the average enterprise. Just six months earlier, the gap was only 2.6 times, indicating that enterprises investing deeply in AI are building an exponential performance advantage.
Enterprises have moved from asking AI questions to having AI perform work
OpenAI distinguishes between two levels of AI use. Conversational AI helps users clarify their thinking through dialogue, while an AI Agent (AI Agent) can independently plan steps, call tools, execute multi-step tasks, and produce auditable results.
As of June 2026, Codex, OpenAI’s AI Agent for software development, accounted for 64% of all tokens generated by enterprise customers. Even when powered by the same models, AI Agent tasks are growing much faster than simple Q&A because they involve longer and more complex workflows.
Why is the gap widening between leading enterprises and the rest?
The report defines leading enterprises as those in the top 10% by monthly token output per active user, while average enterprises fall between the 45th and 55th percentiles. The gap spans industries and company sizes. It is largest in information technology, at 11.7 times, and smallest in manufacturing, at 5.3 times. Usage among average enterprises has grown only 1.9 to 2.8 times over the past year, with most organizations still using AI primarily for simple Q&A.
Another source of the gap is whether enterprises connect AI Agents to real business environments. Among weekly active users at leading enterprises, 21% use Plugins, which integrate enterprise systems and data, while 19% use skills, or reusable work instructions. At average enterprises, the corresponding figures are only 9% and 3%.
Which departments are adopting AI Agents?
AI Agent adoption is no longer limited to engineering. Since February, weekly active Codex users within enterprises have increased 108-fold in legal, 41-fold in both sales and recruiting, and 26-fold in marketing. Engineering saw a comparatively modest fivefold increase. Virgin Atlantic’s engineering team used Codex to reduce the time required to refactor legacy code from two weeks to 30 minutes.
The report also identifies a trend that runs counter to most enterprise surveys: six months after adoption, frontline employees send 13 more messages per week than senior executives. This suggests that identifying employees who already use AI extensively and scaling their practices across the organization may be more effective than relying on executives to lead by example.
Among the U.S. public companies analyzed in How Organizations Use AI, those that had adopted AI tended to have more assets, larger workforces, and higher R&D investment.
What does this mean for traditional enterprises adopting AI Agents?
The report repeatedly stresses that leading and average enterprises use the same models. The difference is whether they connect AI Agents to their own data, tools, and workflows while establishing clear permission and review mechanisms. For most enterprises, the main barriers are data integration, workflow mapping, and governance.
Traditional industries in Taiwan most often encounter these same three obstacles when adopting AI Agents: they lack internal engineering teams capable of handling systems integration, and they have yet to establish clear rules defining who may use AI and who must review its output.
EgentHub is a Taiwan-based enterprise AI Agent management platform serving more than 200 manufacturers and 300 retail brands. Through hybrid cloud and on-premises deployment, four knowledge base modes, and the MCP protocol, EgentHub integrates with existing enterprise ERP and CRM systems. Its FDE consultants also guide enterprises through implementation, from workflow mapping to prompt engineering. Partner enterprises save an average of more than 2,000 labor hours per month.
The widening gap between leading and average enterprises comes down to who establishes robust workflows and governance first.
FAQ
How does OpenAI define a leading enterprise?
In its Enterprise Signals report, OpenAI classifies enterprises in the top 10% by monthly token output per active user as leading enterprises. Those between the 45th and 55th percentiles are classified as average enterprises. These groups span industries and company sizes, so the metric reflects depth of use rather than company size.
Manufacturing has the smallest gap at 5.3 times. Does this mean traditional industries are less suited to adopting AI Agents?
Quite the opposite. Because 5.3 times is the smallest gap across all industries, average manufacturing enterprises have even more room to catch up. The report attributes the gap to workflow integration and governance. These are foundational capabilities that enterprises can build regardless of industry.
What should an enterprise do first when adopting AI Agents?
The report points to three priorities: data integration, workflow mapping, and access governance. Enterprises should begin with one highly repetitive workflow, establish the necessary integrations and review mechanisms, and then expand into other departments.
Why do frontline employees use AI more than executives?
The report found that six months after adoption, frontline employees send 13 more messages per week than senior executives. One possible explanation is that frontline employees handle more repetitive tasks, allowing AI Agents to reduce their workload directly. This also suggests that organizations may achieve better results by identifying employees who already use AI extensively and developing them into internal champions rather than relying solely on top-down demonstrations.
References
- Enterprise Signals — Enterprise data — OpenAI’s official report on enterprise usage data
- From assistance to execution: How enterprises put AI to work — OpenAI’s analysis of the report
- How Organizations Use AI: Evidence from ChatGPT — Academic paper (arXiv preprint co-authored by Aaron Chatterji, David Holtz, and others)