OpenAI published a new analysis on October 14, 2026 arguing that the EU's next phase of AI adoption is agentic: tools that carry out multi-step tasks rather than answer one-off questions. The report, from OpenAI Economic Research and summarized in the post The EU's next chapter in AI is agentic, makes three headline claims. In OpenAI's EU sample of ChatGPT Enterprise customers, Codex and ChatGPT Work account for more than half of output tokens. Among EU small and medium businesses, weekly active users of those two tools grew more than eightfold since May 2026. And the spread between countries is wide, from 22% in Slovenia to 85% in Lithuania.
These are OpenAI's own figures about its own customers, so they say a lot about how paying users behave and little about the EU economy at large. This post lays out what the report says, how to put it next to official statistics, and what a builder or business owner should and should not conclude.
TL;DR: the report in one table
| Question | What OpenAI reports |
|---|---|
| Who is in the data? | EU ChatGPT Enterprise (and Business, for the frontier-firm cut) accounts; SMB means 200 or fewer associated users |
| Share of agentic use | Codex and ChatGPT Work are more than half of output tokens in the EU enterprise sample |
| SMB growth | Weekly active users of Codex and ChatGPT Work up more than 8x since May 2026, in every member state |
| Country spread | Agentic share of output tokens from 22% (Slovenia) to 85% (Lithuania) |
| Intensity gap | EU frontier firms use over 4x the output tokens per active user of typical EU firms, but trail non-EU frontier firms by more than 2x |
| Global standing | EU-based firms are only 10% of frontier firms worldwide |
| Who is quoted? | Sarah Friar, OpenAI's chief financial officer |
| What is not public in the post? | Absolute user counts, token volumes, sample size and per-country firm counts |
What OpenAI actually said
The post is short, and most of the substance sits in an interactive data page and a downloadable PDF that OpenAI links. The text itself makes these points:
- Businesses are "moving beyond asking questions to using agents to carry out tasks from planning through completion."
- Among ChatGPT Enterprise customers, agentic use "already represents a substantial share of activity," and in the EU sample Codex and ChatGPT Work "now account for more than half of output tokens."
- The shift reaches smaller firms: since May 2026, weekly active users of Codex and ChatGPT Work among EU SMBs grew more than eightfold, "with growth across every member state."
- Adoption is "far from uniform," ranging from 22% in Slovenia to 85% in Lithuania, and the type of work differs by country.
- On depth of use, the takeaway is that EU frontier firms use AI much more intensively than typical EU firms but "still trail firms at the global frontier."
OpenAI's recommendation is generic but clear: help businesses move beyond experimentation, give workers skills and confidence, and make advanced capabilities more accessible and affordable. Sarah Friar, OpenAI's CFO, is quoted saying more than half of AI use among its EU enterprise and business customers now involves these tools, and that "there is still plenty of room for AI use to grow across the EU."
Codex and ChatGPT Work: why they count as "agentic"
The report uses the two products as its proxy for agentic use. That is a definitional choice worth understanding. Codex is OpenAI's coding agent, which reads a repository, runs commands and edits files. ChatGPT Work brings the same Codex tools and skills to non-coding office tasks, and has gained features such as a cloud browser that stays signed in to websites.
Two consequences follow. First, "output tokens" is a volume measure, and agent runs are token-hungry: a single multi-step task can generate far more output than a chat answer. A majority share of tokens therefore does not mean a majority of users or tasks are agentic. Second, tokens flowing through Codex include developer work, so a country with a large software-services sector can show a high agentic share without that reflecting wider business adoption. Lithuania at 85% and Slovenia at 22% may reflect who OpenAI's enterprise customers are in each country as much as how firms in those countries work.
OpenAI does not publish, in the post text, how many accounts sit behind each national figure. Small samples in small member states can swing percentages sharply.
How this compares with official EU numbers
The best public benchmark for company-level AI use in Europe is Eurostat. Its latest release (data extracted in December 2025, covering 2025) reports that 19.95% of EU enterprises used AI technologies, and 55.03% of large EU enterprises did. Eurostat counts firms with at least ten employees across a broad definition of AI, including text analysis, speech recognition and generative models, and it measures whether a firm uses AI at all, not how intensively.
Put side by side, the two sources answer different questions:
| Eurostat | OpenAI report | |
|---|---|---|
| Measures | Share of firms using any AI technology | Share of tokens through agentic products, among OpenAI customers |
| Population | Representative sample of EU enterprises with 10+ employees | Paying OpenAI Enterprise and Business accounts |
| Period | 2025 survey | May to September 2026 usage |
| Reveals | Breadth of adoption | Depth of use among adopters |
OpenAI's report is useful precisely because it measures intensity, which surveys miss, but it cannot tell you how many European firms have started. Its own most reliable claim is relative: within its customer base, heavy users pull away from typical users.
The "frontier firm" gap
The most policy-relevant finding is the stack of gaps. In September, EU frontier firms (the top 10% of firms by output tokens per active user) generated more than four times as many output tokens per active user as typical EU firms. Yet frontier firms outside the EU generated more than twice as many as the EU's frontier firms. And EU-based firms were only 10% of frontier firms globally, "low considering that the EU accounts for a higher share of registered companies globally."
Read cautiously, this suggests the long tail of heavy users is thinner in Europe. Again the definition matters: it ranks firms by token intensity within OpenAI's customer set, so a European firm that uses a different vendor, a self-hosted open-weight model or a sovereign provider does not appear. For the policy backdrop, see our guides to Europe's AI landscape and what changes after EU AI Act enforcement.
Why a lab is publishing country-level adoption data
Labs increasingly publish economic research as part of their public policy work in Brussels and national capitals, and the timing is not accidental. The EU is deciding how to apply the AI Act to agents, and officials have already voiced concern about autonomous systems; our coverage of the EU technology commissioner's comments on rogue AI agents shows where regulators' attention is. A report framed around adoption gaps and "making advanced capabilities more accessible and affordable" supports an argument for lighter friction on deployment, which fits OpenAI's commercial interest. That is not a reason to dismiss the numbers, but it is a reason to read them as advocacy backed by real data rather than neutral statistics.
What to take from it as a builder or business owner
- Use the 22% to 85% range as a prompt, not a benchmark. If you run a European firm, the useful question is what share of your own AI spend goes to agentic workflows versus chat, not where you sit among OpenAI's national averages.
- Expect agent costs to dominate. If agentic tools produce most output tokens, they will likely dominate your bill too. Our notes on Codex usage limits explain why quota drains faster with agent loops.
- Check your governance before scale. Agents act on systems and data. Under the EU's rules, deploying them raises questions about oversight and logging that chat does not.
- Compare vendors on the same axis. None of this data covers Anthropic, Google, Mistral or open-weight deployments. A fair read of European adoption needs all of them.
If you deploy agents with access to company systems, a runtime control layer helps: AgentBeam, the agent security platform from the explainx.ai team, stops AI agents before they take dangerous actions.
What is missing
The post and its summary leave several things unstated: the number of accounts in the sample, absolute weekly user counts, the per-country sample sizes, how "weekly active" is counted for ChatGPT Work versus Codex, and how much of the 8x growth reflects a low May baseline (the products were newer then) rather than a mature base. A growth multiple from a small starting point is easy to achieve and hard to interpret. The interactive data page and PDF OpenAI links may answer some of these; we could not load media from the post itself, so check the original for tables and methodology.
Related reading
- Europe AI landscape: sovereign compute and the EU AI Act
- What changes after EU AI Act enforcement
- EU commissioner Virkkunen on rogue AI agents
- Codex passes 20 million users
- ChatGPT Work tools and skills reference
- Why Codex quota drains fast
- Official sources: OpenAI's report post, Eurostat: Use of AI in enterprises
Figures are as published by OpenAI and Eurostat on the dates cited and may be revised; OpenAI's data covers its own customers only.