For two years, the question about AI agents in Europe was whether they would stick. OpenAI’s own numbers, published October 14, suggest that question is settled, at least inside its own customer base. What is still open is how deep that use actually goes, and under what rules.

OpenAI Economic Research published a short report on agentic AI in the EU. Treat everything in it as vendor data: OpenAI measured its own products, in its own customers, and has every reason to like the result.

What the report says

Among ChatGPT Enterprise customers in OpenAI’s EU sample, Codex and ChatGPT Work, the tools that carry out tasks rather than just answer questions, now account for more than half of all output tokens. Plain question-and-answer chat is now the minority use.

Small businesses are moving too. OpenAI counts a small or medium business as one with 200 or fewer associated users. Since May 2026, weekly active users of Codex and ChatGPT Work at EU SMEs grew more than eightfold, in every member state.

Then there is the spread. The agentic share of output tokens ran from 22% in Slovenia to 85% in Lithuania. Same tools, same vendor, same continent, a gap of almost four to one.

OpenAI’s last finding is the one worth sitting with. It calls the top 10% of firms by output tokens per active user “frontier firms.” In September, EU frontier firms generated more than four times the output tokens per active user of a typical EU firm. Frontier firms outside the EU generated more than twice what the EU frontier firms did. The EU made up only 10% of frontier firms worldwide.

Adoption is no longer the question

Eightfold growth in five months, at small firms, in every member state, is past the novelty stage. A company can roll out Codex to 300 seats, watch weekly active users climb, and still have nothing running in production that a finance lead would sign off on. The dashboard looks good. The numbers on a P&L do not move. That gap between activity and real use is likely where a lot of Europe sits, given how wide the Slovenia-to-Lithuania spread is.

Token volume on its own does not settle the question either. More output per user can mean more real work got done. It can also mean longer prompts and agents retrying in circles. OpenAI’s report measures intensity. It does not measure whether that intensity produced anything useful.

The actual work is depth, and governance belongs in it from day one

Moving from a typical firm to a frontier firm means building the plumbing around the tool: deciding which workflow the agent owns end to end, wiring it into the systems of record, setting its permissions, logging what it did, and defining what happens when it gets something wrong.

In the EU specifically, governance cannot be an afterthought. With the AI Act in force and GDPR sitting over every process, an agent that touches invoices or customer records without an audit trail will not survive its first compliance review.

What this means for a buyer

Nobody needs convincing that agents are useful. OpenAI’s own report shows where most companies are stuck: wide use, shallow results. If your team has seats and no shipped workflow, you are in the typical-firm column. Getting out of it starts with one governed workflow in production, not a bigger pilot.

Sources

Kush AI Automation builds in accounts you own, with permissions, logging and an owner on each workflow from day one. Bring the one process you most want in production and book a free call.