The AI Gap Between Companies Tripled in Five Months

OpenAI published new enterprise research on August 12, and one ratio in it belongs on every marketing leader’s wall. The top tenth of its enterprise customers, the group OpenAI calls frontier firms, now generate 8.3 times as much AI output per active user as typical firms. In January, that multiple was 2.6.

Call it the delegation gap, because that’s what actually separates the two groups. The leaders stopped asking AI questions and started handing it finished work. As of June, Codex, OpenAI’s agent, the kind of AI you give a whole task rather than a prompt, produced 64% of all output tokens among enterprise customers. Output tokens are the units AI work gets billed in, so read that as: most of the work OpenAI’s business customers get from it no longer happens in a chat window.

The number that matters for this readership sits further down the report. Since February, weekly active Codex users grew 108x in legal teams, 41x in sales, and 26x in marketing and communications. Engineering, where agents started, grew just 5x, because engineers were already there. OpenAI doesn’t disclose the starting bases, and a 26x jump off a tiny base is still a small number. The direction is the news: the agent wave has left the engineering department, and marketing is in its path.

Richard Masters, VP of data and AI at Virgin Atlantic, put it plainly in VKTR’s coverage of the report: “The trajectory of Codex is thinking beyond pure engineers. It’s moving into a real tool for everyone.”

What do the frontier firms do differently? Not more chatting. When OpenAI’s chief economist Aaron Chatterji launched this research program in May, he flagged the same pattern in that quarter’s data: “Message volume explains only 36% of the frontier advantage, and most of the gap comes from deeper AI use.” The August report shows where the depth lives. At frontier firms, 21% of weekly active users work with plugins, connections into company tools and data, against 9% at typical firms. For skills, saved workflows an agent can repeat, it’s 19% against 3%.

The mechanism is compounding, and it’s worth stating in budget terms. A team that connects AI to its CRM, its analytics, and its content system gives every AI hour more context, which makes delegation work, which justifies connecting more. A team still pasting text into a chat window gets linear returns and concludes AI is overhyped. Both conclusions are locally correct. That’s why the multiple moved from 2.6 to 8.3 in five months instead of narrowing, and why your AI bill behaves the way it does once delegation starts.

The honest caveats: this is OpenAI measuring its own customers on its own platform, and output tokens measure work produced, not results achieved. A company can generate 8x the tokens and 1x the revenue. Nothing in the report ties the gap to margin, though per-employee output is where it would show up first.

For a 5-to-30 person B2B SaaS company, the report is still useful, because it’s a leaked playbook. The frontier firms’ moves are copyable at any size. Connect the AI to the systems where your context lives. Turn the tasks you repeat into saved workflows. Move one real deliverable per week from “drafted with AI” to “delegated to AI, reviewed by a human.” The teams doing this are simply further along a curve that compounds.

The gap tripled in five months while most marketing teams were still debating tool choice. It won’t wait for the debate to finish.