The video is narrated in Japanese. This page is the English write-up: the same argument, key points, full transcript in translation, FAQ and primary sources.
The short answer
OpenAI's enterprise usage data suggests that the companies using AI most deeply connect it to their own context and tools, widening from answering questions to executing multi-step work. But 8.3x is output tokens per person — not a multiple of productivity or profit. What matters is turning an individual task that worked into a shared workflow with defined inputs, tools, completion criteria and a human check.
Key points
- The data is aggregated from within OpenAI's own enterprise customers; it is not a survey representing all companies.
- 8.3x is a difference in output token volume, not a difference in ROI or productivity.
- Within enterprise usage, Codex is reported to account for 64% of total output tokens.
- The most advanced users make heavier use of features such as Plugins and skills that connect the model to company context and tools.
- The implementation problem is not handing out accounts but turning work that succeeded into a repeatable shared workflow.
Transcript
Read the full transcript
OpenAI's "From assistance to execution" and Enterprise Signals, published on 12 August 2026, cover the transition of enterprise AI from answering questions to executing work.
The publisher is OpenAI, which sells ChatGPT Work and Codex. It has the advantage of observing its own customers' usage, but this is the seller's own analysis, not third-party verification of effectiveness.
Enterprise Signals analyses aggregated, anonymised enterprise usage data from a sample of more than 10 million messages. It is behavioural data showing how usage deepens within OpenAI's customer base.
Monthly output tokens are divided by active users; the top 10% are called frontier companies and the 45th-55th percentile typical companies. This is not a classification chosen by revenue or productivity.
In June, frontier companies generated 8.3 times the output tokens per person of typical companies. But token volume is not the value of the work, and it is not a multiple of productivity or profit.
Among enterprise customers, Codex was reported to account for 64% of combined Codex and ChatGPT output tokens. That is one signal of growth in delegating multi-step work.
Advanced users made heavier use of Plugins and skills, connecting the model to company context and business tools rather than only asking it questions.
High growth rates are shown for legal, sales, recruiting and marketing, but a small base makes multiples look large, so read them separately from the absolute headcount and results today.
Even where usage correlates with company performance, causation is not proven. Output tokens do not directly measure quality, cost-effectiveness, or the burden on the recipient.
In a Japanese company, pick one AI task an individual finished well, then turn the required inputs, the data and tools, the completion criteria and the human check into a template, and make it a shared departmental workflow.
AI readiness is not measured by how many people attended training. It is measured by whether you can explain which AI acts for whom, on what data and tools, what it can complete, and how permissions, evaluation and auditing work.
FAQ
What is the 8.3x figure in OpenAI's data?
The difference in output tokens per active user between companies in the top 10% by usage depth and typical companies. It is not a multiple of productivity or profit.
How do you move enterprise AI from assistance to execution?
Take one task that succeeded, define its inputs, the data and tools it uses, its completion criteria and the human check, and turn it into a workflow others can share.
Does this data represent all companies?
No. It is usage data from within OpenAI's enterprise customers, and does not cover companies using other providers' models or companies that have not adopted AI.
Sources
- OpenAI — From assistance to execution — Official write-up of enterprise usage trends and the headline metrics
- OpenAI — Enterprise Signals — The Enterprise Signals data and an explanation of the method