In OpenAI’s sampled EU business accounts, Codex and ChatGPT Work generated more than half of output tokens by the end of September. The interactive report divides the total into 34% from Codex, 21% from ChatGPT Work and 45% from ordinary ChatGPT. Added together, the two agentic products account for 55%.[1][2]
The result is evidence of a shift inside OpenAI’s Enterprise and Business customer base: from asking for an answer toward assigning longer tasks such as coding or multi-step work. It is not a survey of every EU company, and an output token is not a completed project, an hour saved or a unit of economic value.[1]
Three measures that should not be confused
| Measure | What it describes | What it does not establish |
|---|---|---|
| 55% agentic output-token share | The mix of generated text or code inside sampled OpenAI business accounts | The percentage of EU businesses using agentsSource 1Source 2 |
| More than eightfold SMB weekly-user growth since May | Growth in active users of Codex and ChatGPT Work among sampled businesses with no more than 200 associated users | The absolute number of firms or their return on investmentSource 1 |
| 20% EU business AI adoption in 2025 | Eurostat’s economy-wide share of enterprises using at least one AI technology | How intensively adopters used one vendor’s agentic productsSource 3 |
Intensity varies more than the EU average suggests
OpenAI reports that the agentic share ranged from 22% of output tokens in Slovenia to 85% in Lithuania among accounts in its sample. It also defines “frontier firms” as the top tenth by output tokens per active user. EU frontier firms generated more than four times as many tokens per active user as typical EU firms, while frontier firms outside the EU generated more than twice as many as the EU frontier.[1]
Those comparisons describe usage intensity, not company quality. A software team may generate many tokens because its work is complex, because it iterates heavily or because outputs are verbose. A smaller number of well-targeted tasks can also produce more value than a large volume of unused output.[1]
Why the denominator matters
Eurostat’s 20% adoption rate covers enterprises with at least ten employees and asks whether they used specified AI technologies in 2025. It reports 55% adoption among large businesses and 19% among small and medium-sized enterprises. OpenAI’s report instead examines activity among its own paying business accounts and defines small or medium businesses as accounts with 200 or fewer associated users. The populations and questions are different.[1][3]
From experimentation to operational evidence
Token growth can reveal that people are assigning more work to tools. To establish business impact, firms still need measures tied to the task: cycle time, defect rate, accepted code, customer outcome, cost, rework and human oversight. The most useful follow-up would connect agentic usage to those outcomes while controlling for company size and sector.[1]
Sources & further reading
- The EU’s next chapter in AI is agentic OpenAI · 14 Oct 2026 · Accessed 15 Oct 2026
- The EU’s next chapter in AI is agentic: interactive data OpenAI Economic Research · 14 Oct 2026 · Accessed 15 Oct 2026
- Digitalisation in Europe — 2026 edition Eurostat · Accessed 15 Oct 2026