OpenAI’s ‘Most EU Business AI Is Agentic’ Claim, Put To The Denominator Test

🔍 Read the full analysis: OpenAI’s ‘Most EU Business AI Is Agentic’ Claim, Put To The Denominator Test on ThorstenMeyerAI.com

Before you orderOffer from Amazon

Get the latest gadgets delivered free with Prime

• Fast, free delivery on millions of items
• Prime Video, Amazon Music and more included
• Member-only deals all year

Start your free Prime trial →Free trial for eligible customers · Cancel anytime

As an affiliate, we earn on qualifying purchases.Hide

OpenAI said Codex and ChatGPT Work generated 55% of output tokens in its ChatGPT Enterprise and Business accounts across the 27 EU member states by September 30, 2026. The figure offers a snapshot of activity on OpenAI’s own services, but it does not mean that agents handled most European business AI tasks: the measure counts generated tokens, not users, interactions or business value.

The figures appear in OpenAI Economic Research’s report Europe’s Agentic Opportunity: Moving from breadth to depth with AI agents, published October 14 alongside a post from the company’s Global Affairs team. OpenAI defines agentic use in this report as activity from Codex, its coding agent, and ChatGPT Work, an agent mode launched in July 2026 that can work with files, browsers and company tools. By September 30, Codex accounted for 34% of output tokens and ChatGPT Work for 21%; ordinary ChatGPT use made up the remaining 45%.

OpenAI also reported that weekly active users of Codex or ChatGPT Work at businesses with 200 or fewer users increased 8.1 times between the first week of May and the week ending October 5. The company said every EU member state recorded at least fivefold growth. Belgium had the highest reported increase, at 12.1 times, followed by Italy at 10.8 times and Luxembourg at 10.0 times. OpenAI did not provide absolute user counts alongside the growth figures.

Reported agentic output shares varied sharply by country, from 85% in Lithuania and 81% in Cyprus to 22% in Slovenia and 26% in Austria. Germany’s share was 51%, which placed it 17th among the 27 member states. OpenAI also compared its highest-output users: EU “frontier firms,” defined as the top 10% by output tokens per active user, produced about four times the tokens of typical EU firms, while frontier firms outside the EU produced more than twice the amount of the EU frontier group.

What the Token Share Can Show

The report points to growing use of OpenAI’s agent products among its business customers, and gives companies and policymakers a view into how that activity is changing on the platform. But token volume is not a direct measure of adoption across Europe. An agent may read documents, plan steps, generate code or text, check its work and repeat parts of the process. That can produce far more output tokens than a single chat response, even when it is one task.

OpenAI itself describes tokens as an imperfect measure of business value and says agentic tasks can require substantially more computation. The 55% figure therefore describes where output is being generated within the accounts studied; it cannot establish that agents account for most AI interactions or deliver most of the value. The distinction matters when interpreting headlines about workplace adoption, comparing countries or deciding whether firms have moved from pilots to routine use.

The small-business growth figure also needs its time window and baseline. It compares activity with the first week of May, before ChatGPT Work launched in July. OpenAI says a major part of the increase followed that launch. The numbers show increased use of the products over that period, but they do not establish how many businesses adopted them, how much work they perform, or whether the increase will persist.

How OpenAI Defines Agentic Use

The report covers ChatGPT Enterprise and ChatGPT Business accounts in the EU, not every AI provider or all European companies. Its global comparison draws on OpenAI’s Enterprise Signals material. The source describes the data as a picture of activity among OpenAI customers; it should not be treated as a census of the region’s business AI use.

Product timing is relevant to the reported growth. OpenAI’s chart shows Codex accounting for roughly a fifth of EU output tokens in May and 34% by September 30. ChatGPT Work, which was not available at the start of that comparison, contributed 21 percentage points by September. The launch thus coincided with a substantial part of the rise in the combined agent share. That supports a product-use interpretation, but does not by itself show a broad economic shift.

OpenAI’s policy framing is that Europe should move from broad AI adoption toward deeper use, supported by worker skills, confidence and access to advanced tools. That is the company’s interpretation of its findings, rather than a conclusion independently established by the account-level metrics.

What the EU Data Leaves Out

OpenAI does not disclose account or user sample sizes by country, absolute small-business user counts, or a task-based measure that would allow readers to compare agent runs with ordinary chat interactions. Without those details, the country rankings could reflect differences in the customers represented in each market, as well as differences in broader business adoption.

The report does not establish whether token growth corresponds to more completed work, improved productivity, lower costs or durable changes in company practice. It also does not show how OpenAI’s customers compare with businesses using other AI tools or no such tools. The source material’s token-based figures cannot resolve those questions.

OpenAI’s comparison of EU and non-EU frontier firms is likewise based on output tokens per active user, not independently verified productivity or economic performance. The extent to which product availability, customer mix and the timing of launches account for the reported gaps remains unclear.

What to Watch in Follow-Up Data

The next useful evidence would show whether agent use continues to rise after ChatGPT Work’s launch effect is no longer the main feature of the comparison. Further reporting could also provide absolute user counts, country-level sample sizes and task-based measures, while distinguishing between trials and regular workplace use.

OpenAI’s report frames its policy case around skills, confidence and access to advanced AI tools. Whether those recommendations lead to measurable changes in European business adoption, and whether the gap it reports between EU and global frontier firms narrows, remains to be seen. For now, the figures describe activity in OpenAI accounts through late September and early October 2026; they do not settle how widely agents are used across Europe’s economy.

Key Questions

What does OpenAI mean by “55% of output tokens”?

It means Codex and ChatGPT Work produced 55% of the text tokens generated in the EU ChatGPT Enterprise and Business accounts covered by the report by September 30, 2026. It does not mean agents handled 55% of tasks or users’ AI interactions.

Does the report show that most European businesses use AI agents?

No. The data covers OpenAI business accounts in the EU, not all European businesses or users of other AI providers. The report does not provide a region-wide adoption rate.

How did small-business agent use change?

OpenAI reported that weekly active users of Codex or ChatGPT Work at businesses with 200 or fewer users rose 8.1 times from the first week of May to the week ending October 5. It did not publish the absolute user count in the supplied report summary, and it said a major part of the growth followed ChatGPT Work’s July launch.

Why might token share overstate how common agent tasks are?

Agent workflows can generate many tokens while reading material, planning, drafting and checking work. As a result, a smaller number of agent interactions could account for a large share of tokens. The report provides no conversion from tokens to tasks or interactions.

What remains unknown about the country comparisons?

The report does not provide country-level account or user sample sizes in the supplied material. That makes it difficult to tell how representative each national figure is or to interpret the rankings as a measure of overall national AI readiness.

Source: ThorstenMeyerAI.com