What OpenAI reports
The report examines usage data from OpenAI's EU sample of ChatGPT Enterprise and Business customers. Key statements: weekly active users of Codex and ChatGPT Work at SMEs grew from early May to 5 October to 8.1 times, with more than fivefold growth in every member state. In individual countries growth was markedly higher, such as 12.1-fold in Belgium, 10.8-fold in Italy, and tenfold in Luxembourg. Together, both tools generated more than half of output tokens in the sample by the end of September.
The share varies widely. Per the reports, it ranges from 22 percent in Slovenia to 85 percent in Lithuania, with Cyprus at 81 and Hungary at 79 percent. We found no figure for Germany in any of the checked sources.
What the figure measures - and what it doesn't
Three caveats belong with the figure, even though they're missing from the headline. First, the base: these are OpenAI customers using ChatGPT Enterprise or Business, so companies that have already chosen the product. That's a different circle from 'European SMEs'. The reports themselves stress the figures show activity in this sample, not adoption across companies overall.
Second, the starting value: a multiple says little without a baseline. If a small number grows eightfold, it's still a small number. The checked reports name no absolute user numbers. Third, timing: ChatGPT Work launched in July, after the May baseline. Per reports, OpenAI attributes the SME growth to that launch. Part of the multiple is therefore a product-launch effect and not a sign of gradual spread of existing tools. How large that part is can't be derived from the reports.
In addition: weekly active users and output tokens measure that a tool is used and how much it produces. They don't measure whether work gets better, faster, or cheaper. None of the checked reports contains productivity gains; the productivity statements are forward-looking.
The more telling figure: the intensity gap
The report contains a second observation more interesting to companies. The top 10 percent of companies by output tokens per active user, which OpenAI calls 'frontier firms', generated more than four times a typical EU company's output in the EU in September. Outside the EU, frontier firms were again more than twice as high as their European counterparts. EU companies make up nearly 10 percent of frontier firms worldwide, North America around 65 percent.
That describes a spread in depth of use, not of impact. It does fit a finding we described in late August: in the NBER survey of roughly 6,000 executives, in which Germany's Bundesbank participated, differing speeds of adoption were, per a companion study, closely tied to organizational support - encouragement, tools, and training together. Our link is a hypothesis: that frontier firms probably stand out mainly through deeper use says nothing about whether their results are better. The report doesn't show that.
What OpenAI draws from it
Per the reports, OpenAI recommends moving from experimenting to tools that carry out tasks 'from planning to completion', improving access to advanced AI, building workers' skills and confidence, and making the tools more accessible and affordable. That's the position of a vendor whose products would be used more as a result. We think it's an understandable wish, but it isn't a neutral recommendation.
What this means in practice
- With vendor usage figures, check base, starting value, and timing: whose customers, what baseline, which product launch falls in between?
- Measure your own usage and your own impact separately. Active users and tokens say nothing about time saved or quality in your own operations.
- Look at depth of use rather than breadth: who uses agents for completed tasks, who only for individual questions?
- Plan for training and support, not just licenses, as the NBER study suggests.
- Read a vendor's figures about its own customers as a market signal, not as market statistics.
The core: the figure shows that OpenAI's European customers are trying out a new product very quickly. It doesn't show whether impact results. For a company the question isn't whether others use agents, but whether use in your own operations changes anything, and that can only be measured there.