OpenAI published a report analyzing how enterprises actually use AI — more than 1,500 teams, more than 17 million messages — and the headline finding is counterintuitive. Now that everyone can access the same models, the gap between companies isn't closing. It's widening — threefold in six months. Korean tech outlet Mobiinside's breakdown surfaces the mechanics.
OpenAI ranks enterprise-plan customers monthly by output tokens — a measure of how much work AI actually produces — defining the top 10% as frontier companies and the median as typical ones. In January, frontier companies out-produced typical ones by 2.6x. By June, the multiple hit 8.3x. Notably, frontier usage varies by industry while typical-company usage looks similar everywhere and grew only 1.9–2.8x over the past year — which OpenAI reads as a sign many teams still use AI as a simple chatbot. (Tokens aren't a perfect metric, but as a proxy for how much work employees hand to AI, they're serviceable.)
The gap's engine is a shift in usage mode. By June, 64% of enterprise output came from agents like Codex, with conversational ChatGPT down to 36% — a transition from AI that answers to AI that does. The transition speed differed sharply by company: 21% of active users at frontier companies use plugins versus 9% at typical ones; for skills, it's 19% versus 3%. Plugins bundle reusable work instructions (skills) with company data, tools, and execution permissions — a sales plugin might attach CRM access to a sales playbook, letting an agent pull current customer data and past proposals to draft a reviewable response.
It's no accident agents debuted in software development: a codebase concentrates the background and rules of the work in one place, and tests provide a built-in standard for judging output. Where work lives scattered across meeting notes, messengers, spreadsheets, and people's heads — with no verification standard — agents underperformed. This year that wall started falling. Models now read any format and carry multi-step work to completion, while plugins and skills pipe the scattered material in. The result: active agent users grew 108x in legal teams, 41x in sales, 41x in HR, and 26x in marketing during the first half.
The report's recommendations compress into three questions. Which of your team's work can be systematized into documents? Do you have criteria for judging whether AI output passes? And is the workflow of your single best AI user being transferred to the whole team? That last one comes with evidence: junior employees send 13 more messages per week than executives. Your best AI user already works for you — find their method and make it the team standard.
If companies holding the same model diverge, the cause lives outside the model — in context. AI doesn't know which document is current or what "approved" looks like. Agents earn their keep only in teams where brand tone, creative approval criteria, and the reasons past campaigns won or lost are written down. That's the global-data version of the conclusion Korean companies reached in What Socar and Samjjeomsam Learned the Hard Way — the AI Gap Is Accumulation, Not Tools, and the individual-level companion is 12 Career Rules for the AI Era, From a Meta Engineer — "The Gap Is Smallest Today". The starting move is changing the adoption question — from "which model should we use?" to "can we explain to an AI what standards we work by?"
Between January and June, the output gap between OpenAI's frontier companies (top 10% by output tokens) and typical companies (the median) grew from 2.6x to 8.3x — more than tripling in six months.
Usage mode, not model access. Frontier companies connect agents to company data, tools, and playbooks via plugins (21% vs. 9%) and skills (19% vs. 3%), moving from asking AI to delegating to it — while typical companies largely remain in chatbot mode.
Codebases concentrate work context in one place and tests provide a verification standard. Where context is scattered and no pass/fail criteria exist, agents struggle — though non-engineering agent use exploded this year: 108x in legal, 41x in sales and HR, 26x in marketing.
Three things: which work can be systematized into documents, whether criteria exist for judging AI output, and whether the best AI user's workflow is being transferred to the whole team.
To apply what you just read to your own site, start with a free audit of where things are now.
A strategist replies within 24 hours on business days.
Buzzvil's BD strategy team has zero developers but runs 32 AI workflow automations — from 7-minute revenue drop analysis to 90% fewer manual steps. How they scaled from personal hacks to team systems.
Meta’s Muse Wonder and Muse Wonder Full add Active Memory and Active Focus so its AI agent can handle tasks that take days. But a Toronto user’s Marketplace mishap shows the risks of agentic AI.
Pew cloned thousands of panelists as AI digital twins and found a 12.4-point average error. Five ways synthetic respondents fail — and where marketers can still use them in research.