What to know
- The reported share measures generated text, not completed tasks or working hours.
- Compare accepted results and review effort before drawing conclusions about value.
What OpenAI measured
OpenAI's October 14 release reports that Codex and ChatGPT Work together supplied 55% of generated text tokens in its EU Enterprise and Business account analysis. Figure 1 uses a trailing 28-day window ending September 30, excluding the category Other. ChatGPT Work launched in July, changing the available product mix.
The report does not disclose absolute sample counts, establish a continuous account cohort or supply uncertainty intervals. These are company usage figures, not estimates of productivity, working hours or employment effects across the EU economy.
Sources: OpenAI; OpenAI Economic Research
Our interpretation: measure the result that survives review
For a software team, a useful next question is what happened after the text was generated. Did a proposed change pass the team's checks? Did a reviewer accept it? Did it later need correction? A large response can contain a useful solution, repeated explanations or a failed attempt. Counting the response alone cannot distinguish those possibilities.
Our recommendation is to choose an existing category of work and define an acceptable result before comparing tools. A team might track accepted changes alongside the time people spend checking and correcting them. Keep the acceptance standard consistent, and record failed attempts as well as successful ones. Otherwise, an apparently efficient workflow can simply move effort to a reviewer or leave unfinished work outside the count.
Managers should also record changes in task difficulty, staffing and the tools available during a comparison. Those changes can help explain a result without implying that one product caused it. Where the work has consequences beyond the immediate team, include the relevant quality checks rather than treating speed as the sole objective.
This is an editorial measurement proposal, not an evaluation conducted in OpenAI's report. It would require local evidence before supporting a staffing or purchasing decision. The practical question is whether a team can deliver acceptable work with a manageable review burden; more generated text does not answer that question by itself.
Sources & method
AI-generated article checked against the release and Figure 1 of the report. The PDF cover supplies October only; its linked announcement establishes the release day. No underlying account data were reanalyzed. Measurement suggestions are original editorial interpretation.
- The EU's next chapter in AI is agentic OpenAI ·
- Europe's Agentic Opportunity OpenAI Economic Research ·
Drafted and source-checked by Codex AI agents. Read our sourcing, illustration, and correction policy.
Explore the occupations
What could this mean for your work?
See task exposure and potential career moves for roles connected to this story.
