In Lithuania, agents produce 85% of ChatGPT Enterprise output; in Slovenia, 22%. The gap is depth

Two companies can buy the same AI licenses on the same day and end the year in very different places. One uses AI to answer questions. The other hands it tasks to finish. OpenAI’s new research on European business customers puts numbers on that difference, and the numbers say more about organizations than about technology.

What OpenAI found

The report, published on October 14 by OpenAI’s economic research team, looks at agentic use among ChatGPT Enterprise and Business customers in the EU. Its main findings:

• In the EU sample, Codex and ChatGPT Work, OpenAI’s tools that carry out tasks rather than answer questions, now account for more than half of output tokens among ChatGPT Enterprise customers.
• Since May 2026, weekly active users of those tools among EU small and medium businesses, defined as businesses with 200 or fewer associated users, grew more than eightfold, with growth in every member state.
• The agentic share of output tokens ranged from 22% in Slovenia to 85% in Lithuania.
• “Frontier firms,” the top 10% by output tokens per active user, generated more than four times as many output tokens per active user as typical EU firms in September. Frontier firms outside the EU generated more than twice as many as frontier EU firms, and EU firms were only 10% of frontier firms globally.

Share of ChatGPT Enterprise output tokens from agentic tools

0%100%

1. Slovenia22%
2. EU sampleMore than half
3. Lithuania85%

What the numbers measure

Before these figures reach a board slide, three limits need to travel with them:

1. The sample is OpenAI’s customers. It is not the European economy, and it says nothing about companies using other vendors or none.
2. Output tokens are not outcomes. A coding agent produces many tokens per task; a short answer produces few. A high agentic share means agent-style tools generate most of the text, not that they deliver most of the value, or any measured productivity.
3. The vendor has an interest. The report is credible as usage data and closes with a policy argument for broader, deeper use. Read the data and the advocacy separately.

Why intensity is still the right lens

With those limits, the central idea holds and is useful: the gap between organizations is no longer access to AI. Nearly every large company in Europe, the U.S. or Brazil can buy the same tools. The difference is how deeply work has been reorganized around them. A fourfold gap in use per person between frontier and typical firms is not explained by licenses. It is explained by which tasks were handed over, who was trained, which processes were changed and which data the tools were allowed to reach.

That matches what we see in the evidence elsewhere. Agents reach production fastest when they live inside a process people already run, and value shows up only when someone decides what the released time is for.

How to read your own intensity

An enterprise can run OpenAI’s analysis on itself, with any vendor’s usage data, and learn more than any country ranking will tell it:

• Measure depth by team, not adoption by company. Active users per license says who logged in. Tasks delegated per active user says who changed how they work.
• Find your internal frontier. The top 10% of teams by intensity are the ones to study: which tasks they delegate, what they had to fix to do it, and what they stopped doing.
• Pair intensity with outcomes. For each high-intensity team, name one business metric that should move. If none does, the intensity is activity, not progress. Workday’s approach of tracking change at the task level is one way to make that link.
• Watch for the governance lag. Teams that delegate the most are also the ones whose agents touch the most data and systems. Intensity should trigger a review, not a celebration.

For European companies, there is a regulatory layer too: deeper agentic use means more systems that may fall under the EU AI Act’s deployer duties, and more processing to map under GDPR. For everyone else, the lesson is the same without the statute. The race is not to buy AI first. It is to change the work underneath it, and to know whether the change paid.