Does more than half of output tokens going to agents mean companies own what those agents do? No. A usage metric counts activity inside a vendor’s customer base. It doesn’t say who approved an agent’s output, who sent it, or who answers for it when it’s wrong.
I’m Senior Trading and Operations Associate at Stillman Digital, first hire on the trading desk, with 10+ years in crypto and digital asset markets and $30B+ personally traded across institutional digital-asset markets. I also founded Cryudine Group, an AI agency for U.S. SMBs. I read every adoption number through one question: who owns the button?
What did OpenAI publish about EU agent use on Wednesday?
On Wednesday, Oct. 14, OpenAI published “The EU’s next chapter in AI is agentic,” alongside an OpenAI Economic Research report, “Europe’s Agentic Opportunity.” The headline finding, in OpenAI’s words: “In our EU sample, Codex and ChatGPT Work now account for more than half of output tokens.” The report’s Figure 1 puts agentic work at about 55% as of Sept. 30, measured on a “28-day trailing window” and “excluding Other.”
Scope matters. OpenAI’s page says “Sample includes only ChatGPT Enterprise and Business accounts,” and the report’s Figure 1 note says the same. So this is OpenAI’s EU sample of ChatGPT Enterprise and Business accounts, not EU companies in general. The agentic bucket is defined by product: Codex and ChatGPT Work. Regular ChatGPT is the comparison.
On small businesses, OpenAI says “Since May 2026, weekly active users of Codex and ChatGPT Work among EU small and medium businesses grew more than eightfold, with growth across every member state.” OpenAI defines small and medium businesses as those with 200 or fewer associated users. The report’s Figure 3 gives 8.1× the May 1–7 level, and adds that “A major part of this growth followed the introduction of ChatGPT Work,” which it says launched in July 2026. OpenAI publishes no absolute user counts.
Country figures vary widely. OpenAI says “the share of output tokens generated through Codex and ChatGPT Work ranged from 22% in Slovenia to 85% in Lithuania.” That’s the share of output tokens within OpenAI’s sampled business accounts in each country, not how many companies there use agents.
What do “output tokens” and “weekly active users” actually measure?
Output tokens are text the model generates. Weekly active users count people who used a tool in a given week. Both are activity measures, and both are useful for seeing where usage is moving.
The report measures output tokens and weekly active users inside OpenAI’s own customer sample; it doesn’t measure completed work, productivity, or revenue. OpenAI doesn’t disclose a sample size or account count, and much of the small-business growth followed a product launch, by OpenAI’s own account.
OpenAI also measures intensity. It defines frontier firms as “the top 10% of firms by output tokens per active user,” and says that in September “frontier firms in the EU generated more than four times as many output tokens per active user as typical EU firms.” That’s a measure of how much text a firm’s users generate, per user. It isn’t a measure of what got finished or who checked it.
Why isn’t usage the same as ownership?
I didn’t run OpenAI’s EU sample and I’m not grading Codex or ChatGPT Work as products — this is a reading filter for what a usage metric actually measures.
A token count records that an agent produced output. It can’t say who reviewed that output before it left the building. Ownership lives in four moments: who approves, who sends, who settles, and who can roll back. None of them appears on a usage dashboard.
That’s not a criticism of the data. Usage numbers answer an adoption question, and OpenAI presents them as adoption and intensity measures. The mistake is reading an adoption number as an accountability number. More agent output means more decisions waiting for an owner, not fewer. When a dashboard shows usage climbing, the useful follow-up is who signs off on each agent action, and where that sign-off is written down.
On a desk, volume on the screen is not the same as knowing who signed the ticket. That’s framing, not a story. A busy screen tells you activity is high. It doesn’t tell you whose name is on what crossed it.
Is “8× SMB agent growth” an accountability transfer?
No. OpenAI’s Sarah Friar put it this way: “More than half of the AI use among our EU enterprise and business customers now involves these tools, and the number of small business users has grown 8x across the region.” That’s a statement about use and users.
Building AI systems for U.S. small businesses through Cryudine, I read an ‘8× SMB agent growth’ headline as a seat-and-token count, not as a transfer of who owns the button when the agent acts. Growth in seats and tokens tells you more people are handing work to agents. It doesn’t say who on those teams approves what the agent sends or settles.
That matters more at small companies, not less. In a business with 200 or fewer users, the person who approves, sends, and rolls back is often the same person. When agent use grows eightfold, that person’s review load grows with it, unless someone writes down who owns which action.
How does this fit the ownership series?
It’s the same filter on a new axis: adoption dashboards versus accountability. Google launched a universal Gemini agent. Finished work still isn’t ownership. covered finished agent work. Anthropic disclosed unintended Claude actions on live websites. Evaluation access still isn’t ownership. covered eval access. OpenAI is notifying third parties. A notice still isn’t ownership. covered disclosure.
OpenAI shipped always-on Dots. Always-on still isn’t ownership., Shopify opened checkout to agents. Checkout tools still aren’t ownership., OpenAI paused tool-use after rogue agents. Autonomy still isn’t ownership., and JetBrains Air unifies agents. It still doesn’t own the merge. made it for persistence, commerce, autonomy, and code. The test goes back to Execute or vanity and Jev grades decisions. It still doesn’t get the button.: analysis near the button, never the button.
Key takeaways
- On Wednesday, Oct. 14, OpenAI said Codex and ChatGPT Work account for more than half of output tokens in its EU sample of ChatGPT Enterprise and Business accounts, about 55% as of Sept. 30 per its Figure 1.
- OpenAI says weekly active users of those tools at EU small and medium businesses, which it defines as 200 or fewer associated users, grew more than eightfold since May.
- Country figures (22% to 85%) are token shares within OpenAI’s sample, not national adoption.
- The report measures activity, not completed work.
- Usage is an adoption number. Ownership of the approve, the send, the settle, and the rollback is still a person’s job.