Most corporate executives currently view generative AI as a sophisticated version of a search engine or a helpful intern that drafts emails. In boardrooms across the globe, the conversation centers on seat licenses and basic productivity gains. However, a quiet divergence is happening beneath the surface of the enterprise dashboard. While the average company is using AI to summarize meetings, a small cohort of power users is fundamentally rewriting the nature of professional labor. This is not a gap in access to the technology, but a gap in how the technology is integrated into the actual machinery of work.
The Token Divide and the Shift to Execution
OpenAI recently analyzed its enterprise customer base using a specific proxy for AI depth: the volume of output tokens generated per monthly active user. The findings reveal a staggering disparity. The top 10% of leading companies are generating 8.3 times more output tokens than companies in the median range (the 45th to 55th percentiles). This metric is critical because output tokens represent the actual work produced by the model; the more tokens a model generates, the more complex the task it is likely performing.
This gap is accelerating at an alarming rate. In January 2024, the difference between leading firms and average firms was a relatively modest 2.6 times. By June 2024, that gap had widened to 8.3 times, meaning the divide tripled in just six months. This trend is not confined to a specific sector or company size. Whether in manufacturing or finance, the leading firms are utilizing the same underlying models as their peers but are achieving vastly different results. The difference lies in the transition from assistance to execution.
Average companies treat AI as an assistant. They ask a question, receive an answer, and the interaction ends. Leading companies treat AI as an execution engine. They build environments where the AI does not just suggest a path but walks it. This involves connecting the model to proprietary corporate context, external software tools, and repeatable workflows. When an AI is tasked with completing a multi-step project—reading internal data, selecting the right tool, and iterating on a draft—the volume of generated tokens naturally spikes. OpenAI now allows enterprise customers to request custom benchmarks to see exactly where their organization sits relative to these high-performing leaders.
The Architecture of the Agentic Workflow
To understand how these leaders generate 8.3 times more output, one must look at the adoption of plugins and skills. Among the weekly active users of leading firms, the plugin usage rate is 21%, more than double the 9% seen in average firms. Plugins act as the connective tissue, allowing the model to move beyond text generation and interact directly with internal databases or third-party software. When combined with skills—sets of reusable instructions that define standard operating procedures for the AI—the model transforms into an agent.
Consider a sales workflow. An average firm might ask AI to write a follow-up email. A leading firm uses a plugin that bundles the team's sales playbook with CRM access. The agent first calls a skill to understand the objective, uses the plugin to extract the latest customer data and previous proposal history from the CRM, drafts a hyper-personalized response, and delivers it to the human employee for final review. The AI has not just assisted; it has executed the bulk of the operational work.
The gap in skill adoption is even more pronounced. Only 3% of users in average firms utilize skills, compared to 19% in leading firms. This suggests that the real competitive advantage is not the model's intelligence, but the engineering capacity to design and deploy reusable instruction sets across an organization. While plugins provide the tools, skills provide the methodology. The ceiling for this potential is high; within OpenAI's own internal usage data, the weekly plugin usage rate among employees is 95%. This indicates that the 21% usage seen in top firms is still only a fraction of what is possible.
This shift is most evident in the rise of Codex. Since February, the growth rate of weekly active enterprise users has exploded in non-technical fields: legal grew by 108 times, sales by 41 times, and marketing by 26 times. In contrast, engineering grew by only 5 times. This proves that AI is moving from a developer's tool to a general knowledge worker's utility. By June, Codex accounted for 64% of all output tokens generated by enterprise customers, far surpassing standard ChatGPT interactions. Because Codex can call tools, create files, and produce complex deliverables, it generates far more text than a simple chat interface, cementing the move toward agentic workflows.
The Paradox of the Power User
The real-world impact of this execution-centric approach is best seen in the collapse of project timelines. The engineering team at Virgin Atlantic utilized Codex to refactor legacy code, reducing a process that typically took two weeks of manual dependency analysis and modification down to just 30 minutes. Similarly, product teams using ChatGPT Work automated the collection and analysis of fragmented market data, shrinking the timeline for five-year digital strategy development from several weeks to a few hours. By removing the manual labor of information gathering, these teams shifted their focus toward high-level decision-making and strategic review.
However, the most surprising finding in OpenAI's data is the demographic reversal of AI usage. While corporate surveys often suggest that executives and leaders are the primary drivers of AI adoption, the actual API logs tell a different story. Six months after implementation, early-career employees were sending an average of 13 more messages per week than executives.
This discrepancy reveals a fundamental difference in how different tiers of the hierarchy perceive AI. Executives tend to use AI as a top-down assistance tool for summarizing reports or brainstorming ideas. Junior employees, however, use it as a bottom-up execution tool. They break complex tasks into smaller units, iterate with the agent through a feedback loop, and use the AI to actually produce the work. The 13-message gap is a signature of the agentic process: the more a user treats AI as a collaborator in execution, the more they interact with it.
For organizations looking to close the gap, the path forward requires more than just buying more licenses. The transition to execution requires three pillars: deep context connection, strict governance and permissioning, and a mandatory human-in-the-loop review stage. Without a robust governance framework, connecting agents to internal data risks security breaches or the execution of erroneous commands.
Furthermore, the most successful organizations are those that identify their internal AI power users and formalize their individual 'hacks' into organizational standards. When a junior employee finds a way to reduce a two-week task to 30 minutes, that workflow must be documented as a reusable skill and distributed across the team. The ultimate differentiator in the AI era is not the model itself, but the ability to convert individual productivity gains into a standardized corporate operating system.
