OpenAI reported on October 14, 2026 that weekly active users of its Codex and ChatGPT Work agents had grown more than eightfold since May among EU small and medium-sized businesses in its customer sample. For owners and managers weighing a wider AI rollout, the figures show rising use—not how much more work those businesses completed.
The distinction matters: how many people use a tool, how much they use it and whether it improves a result are different questions. OpenAI itself acknowledges that token volume is an imperfect indicator of business value. Unmeasured benefits are not absent benefits, but usage alone cannot establish them.
Read the denominator before the headline
OpenAI defines small and medium businesses as those with 200 or fewer associated users. That is not the same as saying they have 200 or fewer employees.
Reported figure | What OpenAI counted | What it does not establish |
|---|---|---|
Weekly active Codex or ChatGPT Work users in the EU small-business sample: May 1–7 versus September 29–October 5, 2026. | An eightfold productivity gain, or the share of all EU businesses using agents. | |
Share of generated text tokens attributable to Codex and ChatGPT Work in EU Enterprise and Business accounts, excluding “Other”; trailing 28 days to September 30. | The percentage of tasks automated, employees using AI or businesses adopting it. |
Tokens are units of text processed or generated by a model, not finished jobs. OpenAI notes that longer agent tasks can require more computation. More generated text can therefore accompany more demanding work without telling a reader whether the result was useful.
There is also a changing product mix: the report dates ChatGPT Work’s launch to July, during the comparison period. Our interpretation is that aggregate growth cannot, on its own, separate new customers and newly available tools from deeper use by the same workers. The published materials do not establish a fixed customer group or representative coverage of all EU businesses.
For contrast, Eurostat reported that 20% of covered EU enterprises used AI in 2025. Its survey covers specified economic sectors and enterprises with at least ten employees or self-employed people, across multiple AI technologies. That counts businesses, not tokens. The different population, year and unit mean it cannot be compared directly with OpenAI’s 55% figure.
A usage ranking is not an efficiency ranking
OpenAI says its EU frontier firms generated over four times as many output tokens per active user as typical EU firms in September. Its Enterprise Signals methodology ranks customers monthly on that metric: “frontier” means the top 10%, while “typical” means the 45th–55th percentiles.
The analytical consequence is straightforward: higher consumption is part of the definition of a frontier firm. The ranking identifies intensive users; it does not independently demonstrate that those firms are more productive or profitable.
A hypothetical example shows why the denominator matters. Suppose ten active workers generate 100,000 output tokens in one week: 10,000 each on average. In another week, twenty generate 120,000: 6,000 each. Active users have doubled and total output has risen 20%, yet output per active user has fallen 40%.
Those are illustrative assumptions and arithmetic, not OpenAI observations. They show how wider adoption and lower usage intensity can coexist. Neither direction reveals how many acceptable tasks were completed.
What would justify a wider rollout?
For a rollout decision, track accepted work, review effort and total cost alongside adoption. The following is a proposed editorial framework, not a finding of the EU report. It follows the general principle in OpenAI’s evaluation guidance of assessing specific tasks against explicit criteria and human judgment.
Define a usable result before starting. Choose a recurring task and specify the accuracy, completeness and format it must meet. Count rejected and unfinished attempts as well as successes.
Compare similar work. Assess the existing and AI-assisted processes on comparable tasks. Record differences in staff experience, training and tool versions, plus time spent preparing, checking and correcting results.
Count the full burden. Include tool charges, allocated subscriptions, infrastructure and human effort across every attempt. Divide those costs by accepted results, while reporting quality separately. If nothing is accepted, cost per accepted result is undefined—not zero.
Task-level evidence can show benefits. In the 2024 revision of Generative AI at Work, researchers studying 5,172 customer-support agents reported a 15% average increase in issues resolved per hour after introducing an AI assistant. That is an outcome measure from a particular setting, not an estimate for today’s EU agents or a promise for another business.
For a related example of why usable output matters to an automation budget, see RohitAI’s analysis of Asana’s browser-agent cost study.
Set the rollout decision around a reviewed pilot’s quality and practical value, including workers’ experience of supervision. A successful workflow might justify more AI use—or achieve the same result with less. A token-consumption target cannot decide between them.
Methodology: AI-assisted reporting and analysis of published sources, checked on October 14, 2026. RohitAI did not access private usage data, test these products or independently replicate the studies. The arithmetic example and rollout framework are editorial analysis, not measured findings.