OpenAI’s EU Agentic Report Turns Token Share Into a Country Ranking Uncorrelated With Eurostat’s AI Data

On October 14, OpenAI published “The EU’s next chapter in AI is agentic”, a blog post that summarizes an Economic Research report, “Europe’s Agentic Opportunity”. The headlines are easy to repeat. In OpenAI’s EU sample, “Codex and ChatGPT Work now account for more than half of output tokens.” Weekly active users of the two among EU small and medium businesses reached 8.1 times their May level. The agentic share of tokens runs from 22 percent in Slovenia to 85 percent in Lithuania. And EU frontier firms, the heaviest-using tenth, generate less than half as many output tokens per user as frontier firms outside the EU.

I rechecked the report three ways: by recomputing its ratios from the chart data embedded in the blog page, by comparing its country figures with Eurostat’s firm-level AI statistics, and by reading the OpenAI pages and public datasets it leans on. The arithmetic holds. The trouble is what the numbers measure. Tokens stand in for use, a product name stands in for “agentic”, seats stand in for company size, and an unreported sample stands in for the denominator that would show whether the EU’s 10 percent of frontier firms is low. Most definitions are printed in the figure notes, but the blog headlines and the closing executive quote read them more generously than the notes allow.

What each headline counts

The table pairs each headline with the unit behind it and with what the report, the blog post and OpenAI’s Enterprise Signals guide do not print.

Three populations appear in the material. The blog says the country range holds “among the ChatGPT Enterprise accounts in our sample”, while the figure note says “ChatGPT Enterprise and Business accounts”. Figure 5, a chart of each country’s most distinctive work task, draws on a third population: messages from “users with Free, Go, Plus, or Pro plans”, which OpenAI’s individual data page describes as accounts “generally managed by individuals rather than institutions”. Six countries (Cyprus, Estonia, Latvia, Luxembourg, Malta and Slovenia) have no published data in that figure.

Token share is not the share of use

Figure 1 reports that 55 percent of EU enterprise output tokens in the 28 days to September 30 came through Codex (34 percent) or ChatGPT Work (21 percent), against 45 percent through ChatGPT. The report’s own sentence is precise: “agents via Codex and ChatGPT Work accounted for more than 50% of output tokens generated by firms in the EU.” The executive quote that closes the report is looser. Sarah Friar, OpenAI’s chief financial officer, says: “More than half of the AI use among our EU enterprise and business customers now involves these tools.” Tokens are a volume of generated text. Use could mean users, sessions, tasks or hours.

The gap matters because agent runs are long. OpenAI’s own Enterprise Signals guide says that “Completing complex, long-running tasks can require substantially more computation, helping explain why agentic token usage has grown rapidly”, and that “Tokens are an imperfect measure of business value”. Ten short questions in ChatGPT and one long Codex task can differ enormously in output tokens, whatever each is worth. A token share therefore says how much text the agent products emit, not what fraction of people or tasks they handle.

“Agentic” is also a product label here, not a behavior. In the Signals guide’s June figure, “agentic AI use (defined as Codex tokens)” accounted for 64 percent “of combined Codex and ChatGPT output tokens” among enterprise customers. The EU report adds ChatGPT Work, so a one-line question typed into Work counts as agentic, and a long multi-step chain in plain ChatGPT does not. Work alone is 21 of the 55 points, which means 38 percent of what the report counts as agentic comes from a product that launched in July. The EU’s 55 percent for Codex plus Work is not comparable with that 64 percent, because the sample, the window and the denominator all differ. But the report does not say how its EU number compares with the global one, which is the comparison its title invites.

The eightfold rise spans a product launch

Figure 3 plots weekly active users of Codex and ChatGPT Work in EU small and medium businesses as a multiple of the May 1 to 7 level. The blog page embeds the data behind the chart, which makes the timing visible. OpenAI’s own launch and release-note dates are set beside it.

By July 9, the day ChatGPT Work launched, the multiple stood at 1.3, which is 30 percent growth in nine weeks. From July 9 to October 5 it rose 6.2-fold. Work did not exist on May 1 to 7, so the baseline can only hold Codex users, while the October figure counts users of either product. The report concedes the point: “A major part of this growth followed the introduction of ChatGPT Work.” The blog sentence does not separate the two: users “grew more than eightfold, with growth across every member state.” With a second product entering the count, growth in every state is unsurprising. OpenAI’s small-business program also began in the same window, offering training on ChatGPT Work and “exclusive promotions” from partners, so marketing aimed at exactly this group overlaps the rise. The report does not try to separate organic adoption from launch and promotion effects.

A multiple also hides scale. The report gives no count of weekly users in May or October, so 8.1 times could mean hundreds of users or hundreds of thousands. The country multiples, from Belgium at 12.1x and Italy at 10.8x down to Finland at 5.1x, are ratios to unreported bases. A country that already had many Codex users in May can show a smaller multiple than one that started near zero, so the order says as much about the starting point as about the growth.

The country ranking is uncorrelated with Eurostat’s AI data

Figure 2 sorts all 27 member states by the agentic share of September output tokens, under the heading “Agent use differs across EU member states”, and the blog introduces the same range with the sentence “the EU’s adoption of agentic AI is far from uniform”. It is natural to read that order as which countries’ firms are further along. Eurostat has an official measure of how far firms are along: the share of enterprises with 10 or more employees that use at least one AI technology, from its ICT usage survey. OpenAI’s report quotes it, “In 2025, 20% of EU companies with at least 10 people employed used AI, up from 13.5% in 2024”, and the isoc_eb_ai dataset reproduces both numbers: 19.95 and 13.48 percent.

A check you can rerun

Across the 27 countries, the two orderings show no detectable relationship. The Spearman rank correlation between Figure 2 and Eurostat’s 2025 figure for enterprises with 10 or more employees is +0.06, with a permutation p-value of 0.79 from 100,000 shuffles. By size class it is +0.06 for 10 to 49 employees, +0.08 for 50 to 249 and -0.09 for 250 or more. The script below needs only the Python standard library. It reads the Eurostat figures from the API and compares them with the bar labels of Figure 2.

# rho_check.py
import json
import urllib.request

URL = (
    "https://ec.europa.eu/eurostat/api/dissemination/statistics/1.0/data/isoc_eb_ai"
    "?format=JSON&lang=EN&indic_is=E_AI_TANY&unit=PC_ENT&size_emp=GE10&time=2025"
)
data = json.load(urllib.request.urlopen(URL))
position = data["dimension"]["geo"]["category"]["index"]  # geo code -> index into "value"
eurostat = {code: data["value"][str(i)] for code, i in position.items() if str(i) in data["value"]}

# Figure 2 of "Europe's Agentic Opportunity": agentic share of output tokens, September 2026 (percent)
openai = {
    "LT": 85, "CY": 81, "HU": 79, "PL": 76, "SE": 73, "FI": 72, "EE": 71, "NL": 68, "HR": 68,
    "IT": 65, "SK": 58, "BE": 57, "DK": 56, "MT": 56, "FR": 55, "CZ": 52, "DE": 51, "RO": 49,
    "LU": 47, "BG": 42, "IE": 39, "EL": 37, "LV": 35, "ES": 30, "PT": 29, "AT": 26, "SI": 22,
}


def ranks(values):
    order = sorted(range(len(values)), key=values.__getitem__)
    out = [0.0] * len(values)
    i = 0
    while i < len(order):
        j = i
        while j + 1 < len(order) and values[order[j + 1]] == values[order[i]]:
            j += 1
        for k in range(i, j + 1):
            out[order[k]] = (i + j) / 2 + 1  # average rank for ties
        i = j + 1
    return out


def pearson(a, b):
    n = len(a)
    ma, mb = sum(a) / n, sum(b) / n
    cov = sum((x - ma) * (y - mb) for x, y in zip(a, b))
    return cov / (sum((x - ma) ** 2 for x in a) ** 0.5 * sum((y - mb) ** 2 for y in b) ** 0.5)


countries = sorted(openai)
xs = [openai[c] for c in countries]
ys = [eurostat[c] for c in countries]
print(f"countries: {len(countries)}   EU27 in the same table: {eurostat['EU27_2020']}")
print(f"Spearman rho: {pearson(ranks(xs), ranks(ys)):+.3f}")

Running it on October 14 printed:

countries: 27   EU27 in the same table: 19.95
Spearman rho: +0.055

Where the two measures disagree

Slovenia and Lithuania show it best. Eurostat puts them 0.3 points apart, while Figure 2 puts them 63 points apart. Slovenia has the lowest agentic share and, among enterprises with 250 or more employees, the fifth highest AI use in the EU (71.7 percent). Cyprus and Poland are near the bottom of Eurostat’s order and near the top of OpenAI’s.

No relationship is to be expected, and that is the point. Eurostat counts firms using any AI technology, from analysing written language to generating pictures, video and audio, across all enterprises with 10 or more employees. Figure 2 counts the product mix inside the tokens generated by one vendor’s Enterprise and Business accounts. The first describes an economy. The second describes OpenAI’s customers in it. The report prints no account or token counts per country, and a token-weighted share in a small market can be dominated by a handful of heavy accounts. The report itself remarks that “countries with smaller populations like Cyprus, Estonia, and Lithuania have relatively high agentic use within their larger firms.” Read Figure 2 as customer mix, not as a ranking of national adoption.

“Small and medium” is counted in seats

The report defines small and medium businesses as “businesses with 200 or fewer associated users”. It does not define “associated users”. If the phrase means accounts or seats, as it suggests, then a 20,000-person company with 150 ChatGPT seats is a small business in this analysis, and a 240-person company with 300 seats is not. The European Commission’s SME definition uses staff headcount below 250 plus either turnover up to €50 million or a balance sheet up to €43 million, and warns that “A firm that is part of a larger group may need to include staff headcount/turnover/balance sheet data from that group too.”

The report sets the growth figure beside SME statistics, “63% of its private sector employment” and “over 5 trillion Euros”, that it links to Eurostat’s September 30 release. They check out. Micro and small enterprises employed 48 percent and medium-sized ones 15 percent of the business economy workforce, and generated 35 and 16 percent of the €10.9 trillion of value added, which is about €5.6 trillion by my sum. But Eurostat defines those groups “solely by the number of people employed”, so the statistics describe a different set of firms from the one measured in Figure 3. The difference matters because AI use climbs steeply with size. Eurostat’s 2025 figures are 17.0 percent of enterprises with 10 to 49 employees, 30.4 percent with 50 to 249 and 55.0 percent with 250 or more.

The frontier gap and the wrong yardstick

Figures 6 and 7 compare “frontier firms”, defined as “the top 10% of firms by output tokens per active user”. The Enterprise Signals guide adds that OpenAI ranks customers “each month”, so the frontier is a different set of firms every month, and that “typical firms fall between the 45th and 55th percentiles”. The EU report does not define typical firms. A top-decile average rises whenever usage gets more uneven, for example when a new product is taken up heavily by a few firms, and it is the figure most sensitive to a small number of firms. The report prints no firm counts and no intervals for any of these series.

Read from the chart data, the EU gap sits mostly in the top decile. In September, typical EU firms were at 90 percent of the typical level elsewhere, while EU frontier firms were at 47 percent of the frontier level elsewhere. In May the same ratios were 110 and 72 percent. The blog’s takeaway, that EU frontier firms “still trail firms at the global frontier”, is true. The text does not say that the typical EU firm trails by about a tenth, or that the frontier gap widened between May and September, a window that includes the ChatGPT Work launch.

The blog also calls the EU’s 10 percent share of frontier firms low, “considering that the EU accounts for a higher share of registered companies globally”, and links the World Bank’s Entrepreneurship Database. That page counts “the total number of registered companies with limited liability”. In its Excel file, the 26 EU members that appear (Malta has no row) hold 18.6 million of the 73.1 million companies in the 135 economies covered, or 25.5 percent, using each economy’s latest year (2022 to 2024 for the EU members). The sentence is literally true of that file. But the file has no row for the United States, includes “only the data for Shanghai” for China and part of Canada, so it is not a global share. More importantly, the right yardstick is the EU’s share of firms in OpenAI’s own sample of Enterprise and Business accounts, which the report does not print. If the EU supplies 10 percent of the sample and 10 percent of the top decile, its representation is exactly proportional, whatever its share of registered companies.

What holds up

The report’s arithmetic checks out. From the chart data, EU frontier firms produced 4.15 times the output per user of typical EU firms in September, and global frontier firms 2.12 times the EU frontier, matching “over four times” and “more than two times”. All 27 country values in Figure 4 exceed 5 times, as stated. The Eurostat figures it quotes match the dataset, and its SME statistics match Eurostat’s release. The figure notes print the definitions of small and medium businesses and of frontier firms, and OpenAI’s own guide concedes that tokens are an imperfect measure of value. Nothing here suggests a calculation error. The issue is the distance between what the numbers count and what the headlines and the executive quote say.

Five questions for any vendor adoption number

1. What is the unit: tokens, seats, users, firms or tasks?
2. What is the denominator, and who is outside it? Here that means an undefined “Other”, everyone without an Enterprise or Business account, and every firm with no OpenAI contract.
3. What is the baseline of a growth multiple, and did the product set change inside the window?
4. How many accounts sit behind each bar, and how heavy is the tail?
5. Does a named group (small business, frontier, typical) match the definition that policy documents and official statistics use?

Independent counters exist, and each uses yet another unit. Ramp’s card-spend data, covered in OpenAI Narrows Anthropic’s Business AI Lead, New Ramp Data Shows, tracks which vendors U.S. businesses pay for. Firm-level studies can go further, as in A Harvard Study of 718 Firms Turns AI Coding Agents Into a Code-Review Queue Problem, where more code output per worker did not show up as more finished tickets. Tokens measure activity. Neither adoption nor outcomes follow from them automatically.

Five additions would settle most of the open questions: account counts per country behind Figure 2, the weekly-user series split into Codex-only and ChatGPT Work, an agentic share measured in users rather than tokens, the EU’s share of the sample next to its 10 percent of the frontier, and intervals on the frontier series. Each should be computable from the data the report is already built on.

The bottom line

OpenAI’s EU report is a reasonable description of how OpenAI’s own enterprise customers in the EU use its products: the agent products now generate more than half of their output tokens, and user counts rose sharply after a product launch. It is not evidence of how far EU firms have adopted agents, of how small companies in the Commission’s sense are doing, or of which countries lead. For those questions, the official statistics the report itself links are the better instrument. Eurostat shows one in five EU enterprises with 10 or more employees using AI in 2025, 55 percent of the largest ones, and a country order with no visible resemblance to Figure 2.

Sources and method

I read the blog post and all pages of the report, including the links embedded in its PDF, which point to two Eurostat news items, a Bureau of Labor Statistics page and OpenAI’s Signals pages. I also read OpenAI’s Enterprise Signals guide (updated August 12, 2026), its ChatGPT Work launch post (July 9), the ChatGPT release notes, the small-business program post (July 21) and the individual data page. The external sources are Eurostat’s December 11, 2025 news item on AI use (20.0 percent in 2025 and 13.5 percent in 2024; Denmark 42.0, Finland 37.8, Sweden 35.0, Romania 5.2, Poland 8.4), its September 30, 2026 news item on business value added, the isoc_eb_ai dataset (indicator E_AI_TANY, unit PC_ENT, size classes GE10, 10-49, 50-249 and GE250, last updated June 15, 2026), the Commission’s SME definition page, and the World Bank’s Entrepreneurship Database page and Excel file.

The Figure 2 values were read from the report’s text layer and typed into the script above. Ties in the Spearman calculation use average ranks, and the p-value comes from 100,000 random permutations with a fixed seed. Figure 1’s split of 45, 21 and 34 percent was read from the chart labels on the PDF’s second page. The frontier and small-business series come from the chart data embedded in the blog page, and the ratios, growth multiples and the World Bank share are my own arithmetic from those published numbers. For the World Bank share I took each economy’s latest year with data and summed the registered companies. I did not have access to OpenAI’s account-level data, so every statement about what the report leaves out means what is missing from the report, the blog post and the Enterprise Signals guide. The correlation compares two different measures and is a descriptive check, not a test of any claim OpenAI makes. I did not examine how Figure 5’s work-activity labels were assigned.