By Thorsten Meyer | October 15, 2026 | ThorstenMeyerAI.com
Get business pricing on tech for your team
- Business-only prices and quantity discounts
- Tax-exempt purchasing
- Multiple users, one account, clear invoices
OpenAI’s new Europe report says agents now produce more than half of the AI output in its EU business accounts, small-business agent use grew 8.1× since May, and Europe’s top firms trail the global frontier by a factor of two. The direction is believable. The headline number measures something narrower than it sounds.
On October 14, 2026, OpenAI’s Global Affairs team published The EU’s next chapter in AI is agentic, together with a ten-page report from OpenAI Economic Research, Europe’s Agentic Opportunity: Moving from breadth to depth with AI agents. OpenAI’s global Enterprise Signals page adds the worldwide view. Together they are the most detailed public picture of how European companies use AI agents at work.
They are also a vendor describing its own customers with a metric that favors its newest products. Both statements are true, and this article takes both seriously.
What OpenAI reports
The report covers ChatGPT Enterprise and ChatGPT Business accounts in the 27 EU member states. “Agentic” means Codex, OpenAI’s coding agent, and ChatGPT Work, the agent mode that launched in July 2026 and can work with files, browsers and company tools.
- More than half of output is agentic. By September 30, 2026, Codex produced 34% and ChatGPT Work 21% of the output tokens these EU accounts generated, 55% together. Plain ChatGPT made up the other 45%.
- Small businesses are adopting fast. Among businesses with 200 or fewer users, weekly active users of Codex or ChatGPT Work rose 8.1× between the first week of May and the week ending October 5. Every member state grew at least fivefold. Belgium (12.1×), Italy (10.8×) and Luxembourg (10.0×) led.
- The member states differ widely. The agentic share of output ranges from 85% in Lithuania and 81% in Cyprus to 22% in Slovenia and 26% in Austria. Germany sits at 51%, 17th of 27.
- Europe’s frontier trails the global one. “Frontier firms” are the top 10% by output tokens per active user. In September, EU frontier firms produced about four times as many tokens per user as typical EU firms. Frontier firms outside the EU produced more than twice as many as the EU frontier. EU companies make up about 10% of frontier firms worldwide, North America about 65%.
The report’s conclusion is a policy message. Europe’s opportunity is to “turn broader adoption into deeper use”, which means moving businesses beyond experiments, building workers’ skills and confidence, and “making advanced capabilities more accessible and affordable.” OpenAI’s CFO Sarah Friar is quoted saying that more than half of AI use among its EU enterprise and business customers “now involves these tools”.
OpenAI’s Europe report · in one view
From 22% to 85%: how agentic the EU’s OpenAI business accounts are
Share of output tokens produced by Codex and ChatGPT Work in September 2026, per member state, next to how fast small businesses adopted the agents since May.
Agentic share of output tokens, September 2026
Dark tick = EU median (56%). Right column: small-business agent users vs. May; green = at or above the median (8.1×).
- 1Lithuania85%6.8×
- 2Cyprus81%7.5×
- 3Hungary79%8.1×
- 4Poland76%6.5×
- 5Sweden73%8.0×
- 6Finland72%5.1×
- 7Estonia71%7.8×
- 8Netherlands68%8.5×
- 9Croatia68%6.8×
- 10Italy65%10.8×
- 11Slovakia58%7.8×
- 12Belgium57%12.1×
- 13Denmark56%9.9×
- 14Malta56%9.5×
- 15France55%9.2×
- 16Czechia52%8.1×
- 17Germany51%7.5×
- 18Romania49%5.9×
- 19Luxembourg47%10.0×
- 20Bulgaria42%5.9×
- 21Ireland39%9.5×
- 22Greece37%9.8×
- 23Latvia35%5.8×
- 24Spain30%8.8×
- 25Portugal29%6.5×
- 26Austria26%8.2×
- 27Slovenia22%9.2×
Read as a customer map, not a league table: no account or user counts per country. Share and growth barely relate (rank correlation -0.19): Slovenia is last on share but grows 9.2×, Lithuania first on share but grows 6.8×.
Source: OpenAI Economic Research, “Europe’s Agentic Opportunity: Moving from breadth to depth with AI agents” (October 2026); ChatGPT Enterprise and Business accounts. “Agentic” = Codex + ChatGPT Work (launched July 2026). Small and medium businesses = 200 or fewer associated users. OpenAI publishes no sample sizes per country. Rank correlation: own calculation (Spearman) from Figures 2 and 4.
Powered by Thorsten Meyer AI
As an affiliate, we earn on qualifying purchases.
Read with care: five things the numbers don’t say
1. Tokens are not tasks
The headline 55% is a share of output tokens, the text the model generates. That is not a share of users, tasks or business value, and OpenAI says so itself. Its Enterprise Signals page calls tokens “an imperfect measure of business value”. It also explains that agentic tasks “can require substantially more computation”, which helps explain why agentic token use has grown so fast.
An agent run reads files, plans, writes code or documents, checks its work and tries again. That produces far more tokens than a chat answer. A simple example shows how much this matters, using an assumption of mine rather than a figure from the report. If an agent run produced 10 to 30 times the output of a chat reply, a 55% token share would mean that only about 4 to 11% of interactions were agentic.
So the honest reading is narrower: in these accounts, agents now consume more of OpenAI’s output than chat does. That says something about where the computation goes. It doesn’t show that most European AI use is now agentic.
I see this in my own small operation. I lean heavily on coding agents (my current stack is here), and by output tokens per user I would look like a frontier firm many times over. That reflects how agent-heavy my work is. It doesn’t prove I get more value from AI than a team that uses chat well.
2. A large part of the jump is one product launch
Reading off OpenAI’s own chart (Figure 1), Codex accounted for roughly a fifth of EU output tokens in May and 34% by September 30. ChatGPT Work didn’t exist until July and added 21 points within about ten weeks. So of the roughly 35-point rise, ChatGPT Work alone contributed more than half.
The 8.1× small-business growth works the same way. It is measured against the week of May 1–7, before ChatGPT Work launched, and the report says “a major part of this growth followed the introduction of ChatGPT Work.” It gives no absolute user numbers. Growing eightfold from a small base after a product launch shows that the product found users. It doesn’t yet show a structural shift in Europe’s small-business economy.
3. The country ranking has no sample sizes
The 27-country chart is the most quotable part of the report, and the easiest to over-read. OpenAI gives no number of accounts or users per country. It notes that smaller countries such as Cyprus, Estonia and Lithuania “have relatively high agentic use within their larger firms”. In a small country, a handful of large Codex-heavy customers can set the national number.
The two country charts also barely agree. I ranked the 27 countries by agentic share and by small-business growth. The rank correlation is −0.19, weak and slightly negative. Slovenia has the lowest share (22%) but above-median growth (9.2×). Lithuania has the highest share (85%) but below-median growth (6.8×). Low-share countries are not obviously falling behind. They may just be starting from a different customer mix.
Treat the chart as a picture of OpenAI’s customer base in each country, not a league table of national AI readiness. Germany, for what it’s worth, sits below the EU median on both measures: 51% share (median 56%) and 7.5× growth (median 8.1×).
4. This is the inside of a selected group
Every number covers OpenAI’s paying business customers. The report itself cites the broader context: in 2025, 20% of EU companies with at least 10 employees used any AI at all, up from 13.5% in 2024 (Eurostat). Four out of five such companies used none. The agentic shift OpenAI describes is happening inside the AI-using minority, and within that, inside the part that pays OpenAI.
The “EU is only 10% of frontier firms” comparison has the same problem. OpenAI contrasts it with the EU’s larger share of companies registered worldwide. The fair comparison would be with the EU’s share of OpenAI’s business customers, and the report doesn’t give that number. If European firms make up roughly a tenth of OpenAI’s enterprise base, they are represented at the frontier about in proportion. If they make up a quarter, they are badly under-represented. We can’t tell which.
5. OpenAI’s EU and global numbers are not comparable
It is tempting to set the EU figures next to the global Enterprise Signals page. The definitions differ:
| Measure | Global page | EU report |
|---|---|---|
| Agentic share | 64% of Codex + ChatGPT output tokens | 55% of output tokens |
| What counts as agentic | Codex only | Codex + ChatGPT Work |
| Month | June 2026 | September 30, 2026 |
| Accounts | Enterprise customers | Enterprise and Business |
| Frontier vs. typical firm | 8.3× (typical = 45th–55th percentile, June) | “over four times” (September) |
The direction matches. Agents take a growing share of the work, and leading firms pull away from the rest. The numbers can’t be compared one-to-one, and anyone who puts “64% globally vs. 55% in the EU” in a slide is comparing different things.
The gap · and the fine print
Europe’s leaders pull ahead at home, but trail the global frontier by 2×
Output tokens per active user, relative to a typical firm’s level in May. “Frontier” = the top 10% of firms by that measure.
May → September 2026
Hollow dot: May (read from OpenAI’s chart, approximate) · filled dot: September
- Global frontier (ex-EU)12.9×May ≈5.4×
- EU frontier6.1×May ≈3.9×
- Global typical (ex-EU)1.6×May ≈1.0×
- EU typical1.5×May ≈1.0×
What holds up: typical firms barely moved (1.5–1.6×) while both frontiers climbed. EU frontier ≈ 4× typical EU firms; global frontier ≈ 2.1× the EU frontier (May: ≈ 1.4×).
Where the world’s frontier firms are
Share of frontier firms by geography, September 2026
Missing denominator: the EU’s share of OpenAI’s business customers is not published, so “only 10%” can’t be judged.
55% of tokens ≠ 55% of work
Share of interactions that would be agentic if an agent run produced … (illustration, not reported data)
| Agent run output | Agentic share of interactions |
|---|---|
| 1× a chat reply | 55% |
| 3× a chat reply | 29% |
| 10× a chat reply | 11% |
| 30× a chat reply | 4% |
Before you quote the report
- Tokens, not tasks. OpenAI itself calls tokens an imperfect measure of value.
- One launch. ChatGPT Work (July) added 21 of ~35 points; SMB growth is indexed to the week before it existed.
- No sample sizes per country; a few large accounts can move a small country.
- Selected group: only paying OpenAI business accounts. Eurostat: 20% of EU firms (10+ staff) used any AI in 2025.
- Not comparable to the global page (64%, Codex only, June, Enterprise only).
Source: OpenAI Economic Research, “Europe’s Agentic Opportunity: Moving from breadth to depth with AI agents” (October 2026); ChatGPT Enterprise and Business accounts. “Agentic” = Codex + ChatGPT Work (launched July 2026). Small and medium businesses = 200 or fewer associated users. OpenAI publishes no sample sizes per country. Figures 6 and 7: frontier = top 10% of firms by output tokens per active user; global lines exclude EU firms. Interaction table: own illustration assuming a chat reply = 1 unit of output and the reported 55/45 token split.
Powered by Thorsten Meyer AI
ChatGPT Enterprise business accounts
As an affiliate, we earn on qualifying purchases.
As an affiliate, we earn on qualifying purchases.
What is real
None of this makes the report worthless. No one outside OpenAI can see this data, and several findings hold up even with the caveats:
- The gap between leading and ordinary firms is growing. Measured against a typical firm’s level in May, the global frontier rose from about 5.4× to 12.9× by September, and the EU frontier from about 3.9× to 6.1×. Typical firms in both regions barely moved (1.5–1.6×). Tokens overstate the gap, but even heavily discounted, the trend is clear.
- Agents are moving beyond software teams. Globally, weekly active Codex users grew 108× in legal, 41× in sales and recruiting and 26× in marketing since February, compared with 5× in engineering. The base was small, but the direction matches what I see: once an agent can work with documents and tools, it is no longer just a developer tool.
- Leading firms use the deeper features. Globally, 21% of weekly active users at frontier firms use plugins and 19% use skills, compared with 9% and 3% at typical firms. The difference isn’t access to a better model, since everyone uses the same models. It is how much context and how many tools the agent is given.
- Leading firms set rules for agents. OpenAI’s description of frontier firms includes clear rules for where agents may operate, what data they may see, when they may act and how people review higher-risk decisions. That part of the report deserves more attention than the token shares.
The report also explains why coding went first: codebases give agents clear context, and tests make results easy to check. Much other knowledge work offers neither. That is the useful lens for any European business deciding where to put agents next.
AI agent software for small business
As an affiliate, we earn on qualifying purchases.
As an affiliate, we earn on qualifying purchases.
What I would take from it as a European business
- Measure outcomes per workflow, not tokens. A token-intensity leaderboard rewards whoever lets agents generate the most text. Count hours saved, errors avoided and turnaround time on a specific process.
- Start where results can be checked. Coding worked first because tests exist. Find the equivalent in your business: reconciliations, document checks against a template, report drafts against last month’s numbers.
- Write the agent rules before giving access. Decide where agents may act, what they may read, and which decisions a person approves. That is also the part of the work that European data protection rules will ask about anyway.
- Invest in depth with a few people. The gap OpenAI describes is between companies that went deep and those still using AI as a chat box. One team that uses agents well teaches the rest more than a company-wide license does.
AI productivity tools for enterprises
As an affiliate, we earn on qualifying purchases.
As an affiliate, we earn on qualifying purchases.
Bottom line
OpenAI’s Europe report gets the direction right: AI work in businesses is moving from asking to delegating, and the leaders are pulling away. But “more than half of EU business AI is agentic” is a statement about tokens, driven largely by a product OpenAI launched in July, measured inside OpenAI’s own customer base, with no sample sizes per country.
Read it as what it is: a useful view inside a vendor’s customer data, published by its policy team with a policy message attached. The conclusion I’d keep isn’t the 55%. It is the gap between firms that changed how they work and firms that only bought access.
Sources: OpenAI, “The EU’s next chapter in AI is agentic” (October 14, 2026); OpenAI Economic Research, “Europe’s Agentic Opportunity: Moving from breadth to depth with AI agents” (October 2026, PDF); OpenAI Enterprise Signals (as of October 15, 2026). Values for May in Figures 1 and 7 are read from the published charts and are approximate. The rank correlation between agentic share and small-business growth is my own calculation from the report’s two country charts. The 10–30× ratio in section 1 is an illustrative assumption, not a reported figure.
Fall Picks
fall essentials
As an affiliate, we earn on qualifying purchases.
