Evidence at a glance
From Asking Questions to Delegating Tasks
On October 14, 2026, OpenAI Economic Research published an analysis of agentic AI use among EU businesses, examining Codex and ChatGPT Work activity in ChatGPT Enterprise and Business accounts. The shift it describes is from asking AI questions to handing tools tasks that can proceed from planning toward completion.
The report is worth a technology leader’s attention not because it proves that “half of Europe has adopted AI agents,” but because it offers a concrete signal of changing behavior inside business accounts: Codex and ChatGPT Work together account for more than 55% of output tokens in the sampled businesses. Meanwhile, weekly active users of these tools in the report’s SMB accounts grew to 8.1 times their May 2026 level by September. Both figures point to deeper product usage, but neither measures adoption across the EU as a whole.
What the Metrics Actually Measure
The report uses output tokens generated by Codex and ChatGPT Work as a proxy for agentic use. This is intended to capture whether users are handing work to tools rather than simply requesting an answer. It can help track usage intensity within the products and compare businesses and countries in the sample.
But tokens measure usage, not outcomes. More output may indicate more complex tasks or deeper interaction; by itself, it does not show whether work was completed correctly, how many hours were saved, or whether the business earned a return on its investment. It is particularly important to distinguish “55%” from “8.1 times”: the former is the share of output tokens in sampled business accounts, while the latter is the change over time in weekly active users of the relevant tools in the SMB sample. They are different measures and cannot stand in for one another.
Fast Growth, Uneven Distribution
OpenAI defines a small or medium-sized business as one with no more than 200 associated users. Under that product-account definition, weekly active users of Codex or ChatGPT Work in the EU SMB sample grew to 8.1 times their May 2026 level by September. The report says every member state was above five times its May baseline. This suggests growth was not confined to a handful of countries, but it cannot be used to estimate how many EU small businesses have deployed agents.
Usage intensity also varied sharply by country: agentic tokens accounted for 85% of output in sampled Lithuanian business accounts and 22% in Slovenian ones. These figures describe the composition of output in specific OpenAI accounts in September 2026, not the adoption rate or capability of all businesses in either country. For a multinational technology team, they are better treated as prompts to investigate differences in deployment, training, and workflows than as a national productivity ranking.
The Frontier Gap Cannot Be Explained by Tokens Alone
The report defines frontier firms as the top 10% of businesses by output tokens per active user. In September 2026, EU frontier firms generated more than four times as many output tokens per active user as typical EU firms, indicating a substantial gap in usage depth within the region. Yet frontier firms outside the EU generated more than twice as many tokens per active user as EU frontier firms, and EU businesses made up only 10% of global frontier firms.
These comparisons suggest that the EU’s challenge may not be limited to getting more businesses to “try AI.” It also involves extending usage into deeper workflows and enabling more firms to reach the level of the most intensive users. But “frontier” here is defined by token intensity, not verified business performance. The report does not establish that higher token use causes higher productivity, or identify which organizational, skills, or cost factors account for the gap.
Connect Usage Signals to Business Outcomes
EU-wide AI adoption figures provide context, but they cannot be directly compared with this study of product accounts. Eurostat measures businesses with at least 10 employees: 20.0% used at least one AI technology in 2025, up from 13.5% in 2024. That broad category includes many kinds of AI, not specifically agentic AI. The population, business definitions, and measurement methods all differ from OpenAI’s Enterprise and Business account sample.
For technology leaders, a more actionable approach is to track usage intensity alongside task outcomes: which tasks are delegated, how often they are completed successfully, how much rework they require, how much time is actually saved, and whether they demand new human review. Multinational teams can use country differences as a diagnostic starting point, then examine access to tools, employee skills, and workflows rather than drawing conclusions from token rankings alone. The report shows that product usage is changing; whether that change delivers value is a question each business must answer with its own outcome measures.