OpenAI’s latest Enterprise Signals data tells a story that should worry anyone still treating AI as a productivity side dish. Frontier firms—the top 10% of enterprise AI users—now generate 8.3 times as many output tokens per active user as typical companies, a gap that’s more than tripled since January when it stood at 2.6x. That’s not a marginal edge. That’s a fundamentally different operating model.
The widening divide comes down to how leading teams connect agents to real work. Instead of bolting ChatGPT onto isolated tasks, they’re wiring agents into company context, giving them persistent access to tools and data, and delegating substantive workflows. Then they’re making the successful ones repeatable. The startups OpenAI highlights—Basis, Clay, and Exa Labs—each took a stable process, taught it to an agent, and kept human judgment at the point of action.
Basis cut employee onboarding from two hours to 30 minutes by turning the process into a reusable Codex skill that handles integration setup automatically. HR updates the skill when exceptions surface, and the next cohort benefits. Clay built a persistent subagent for every sales account that reviews primary sources overnight, then hands the account manager a morning list of priority moves—saving roughly an hour of inbox triage daily. Exa Labs automated its developer integration pipeline: Codex monitors repos for high-priority opportunities, creates pull requests, runs tests, and prepares weekly updates, while humans still decide what ships.
The throughline isn’t automation theater. It’s trust infrastructure. Each company defined clear triggers, known steps, access to the right tools, and a definition of “done”—then gave people the evidence to inspect before acting. The question for enterprise leaders isn’t whether to adopt agents. It’s whether their workflows are stable enough to teach, and whether their teams are willing to let go of tasks that no longer deserve human attention.
