Do agentic tools now make up most business AI use in the EU?
Within OpenAI's Enterprise and Business customer sample, yes: Codex and ChatGPT Work produced 55% of included output tokens at the end of September. That is not a measure of all EU firms, completed work or productivity.
The short version
3 key points
- 1.At 30 September, Codex and ChatGPT Work generated 55% of included output tokens among sampled EU ChatGPT Enterprise and Business accounts, up from 21% in May.
- 2.Weekly active users of those products in OpenAI-defined EU small and medium businesses were 8.1 times the 1–7 May baseline by 29 September–5 October, after ChatGPT Work launched in July.
- 3.The report does not disclose the number of firms, users, messages or tokens behind the headline figures and measures neither task success nor productivity, wages, job quality or business returns.
The answer: most tokens in one customer sample, not most business AI use
Agentic tools produced most of the output tokens in OpenAI's sample of EU ChatGPT Enterprise and Business customers at the end of September. The company's new Economic Research report puts the share from Codex and ChatGPT Work at 55%, compared with 21% in May. That is a meaningful product-usage shift inside the measured accounts. It does not show that agents now perform most AI-assisted work across the EU economy.
The denominator is output text tokens generated through selected OpenAI products. A token is a fragment of text used for model processing and billing; it is not a completed task, an hour saved or a useful business result. Agentic workflows can also produce long intermediate outputs while searching, coding or operating tools, so their token share can rise faster than their share of successful work. OpenAI's related Enterprise Signals page explicitly calls token volume an imperfect measure of business value.
The practical conclusion is therefore narrower than the headline. Among organisations already paying for OpenAI's business products, use is shifting from conversational assistance towards tools that can carry out multi-step actions. Employers should prepare for that workflow change, but they still need evidence that the actions are accurate, secure and worth the time people spend specifying and reviewing them.[1][2][3]
What OpenAI measured
Figures on agentic token share, country differences and frontier firms use ChatGPT Enterprise and Business accounts. The report defines agentic activity as output from Codex and ChatGPT Work. It shows a 28-day trailing series from May to 30 September and says ChatGPT Work launched in July. In September, the reported agentic share ranged from 22% in Slovenia to 85% in Lithuania.
A separate analysis covers small and medium businesses, defined by OpenAI as accounts with 200 or fewer associated users. It tracks weekly active users of Codex or ChatGPT Work relative to the week of 1–7 May. By 29 September–5 October, the EU index was 8.1 times its May level; country multiples ranged from 5.1 in Finland to 12.1 in Belgium. Those are relative growth measures, not adoption rates. The report does not publish the starting count, so an eightfold rise could begin from a small base.
The work-task map uses another population: a random sample of messages from individual Free, Go, Plus and Pro accounts during the third quarter of 2026. Automated classification maps work-related messages to US Bureau of Labor Statistics intermediate work activities and highlights the category unusually common in each country relative to the EU overall. That analysis excludes enterprise and Codex usage, meaning the report combines distinct samples that answer different questions.[2][3][4]
The missing denominators limit confidence
The report does not state how many EU firms, active users, messages or output tokens underpin the main figures. It does not provide country-level sample sizes, confidence intervals, suppression thresholds or rules for handling accounts whose location could not be mapped. Without those denominators, readers cannot judge the precision of an 85% country share or whether rankings among smaller customer bases are stable.
OpenAI says its broader Enterprise Signals work relies on aggregated, de-identified usage data and automated systems to classify message content; no employee reviews customer messages. That is useful privacy and methods context. The same page mentions a global classification sample of more than 10 million messages, but the new EU report does not say that this is its denominator or identify which figures reuse that sample. It would be misleading to attach the 10-million count to every EU chart.
Selection is the larger issue. ChatGPT Enterprise and Business customers chose OpenAI and are likely more engaged with its products than a typical EU firm. Companies that use Microsoft, Google, Anthropic, local models or no generative AI are absent from the central usage measures. The sample can reveal behaviour among OpenAI customers; it cannot estimate the share of all European businesses using agents.[1][2][3][4]
Why the small-business claim needs care
OpenAI's small-business definition—200 or fewer associated users—is operationally convenient but not the EU's standard statistical definition. Eurostat's enterprise surveys classify businesses using employment and sector rules, while EU policy definitions can also consider turnover and balance-sheet totals. An OpenAI account can represent only part of an organisation, and associated product users are not necessarily the same as employees.
The 8.1-fold growth also spans a product change. ChatGPT Work did not exist at the start of the May baseline and launched in July, so the combined user series partly captures a new route into agentic activity. That makes the acceleration real as product use, but it does not isolate organic growth in a fixed service, new customer acquisition, migration from ordinary ChatGPT or repeated use by the same firms.
Eurostat found that 20% of EU enterprises with at least ten people employed used at least one AI technology in 2025. That official survey covers a defined population and a broad set of technologies, not only generative agents. It provides useful economy-wide context, but its 20% cannot be compared directly with OpenAI's 55% token share or 8.1-fold user index. They use different populations, units and periods.[2][3][5]
What the frontier-firm comparison does and does not mean
OpenAI labels the top 10% of sampled firms by output tokens per active user as frontier firms. In September, EU frontier firms generated 6.1 times the May token output of a typical firm, while frontier firms outside the EU reached 12.9 times that baseline. EU organisations accounted for roughly 10% of the global frontier group, compared with 65% from North America and 13% from Asia-Pacific.
These figures show where heavy use sits within OpenAI's customer base, not where the world's most productive or technologically advanced companies are located. A firm can generate many tokens because employees are experimenting, an agent is verbose or a small active-user denominator magnifies intensity. Another firm can create high value with short outputs or use a competing system invisible to the data.
The report also defines a typical firm differently from the related Enterprise Signals page: the new EU chart describes typical firms as a comparison series without restating the full percentile band, while the methodology page places them between the 45th and 55th percentiles. Publishing the firm counts, distribution, industry mix and sensitivity to alternative intensity measures would make the comparison more interpretable.[2][3][4]
What this changes for workers and managers
For workers, the important change is from asking a model for text to delegating a sequence of actions. That can move effort away from first drafts and repetitive computer operations towards specifying goals, supplying context, checking sources and handling exceptions. It can also create hidden labour when an employee must reconstruct what an agent changed, correct errors or obtain approval for actions that cross financial, legal or safety boundaries.
Managers should not use token volume as a productivity target. Rewarding more tokens could encourage longer, costlier runs without better outcomes. A safer scorecard starts with the task: completion against a verified standard, error severity, rework, time saved after review, worker experience, security incidents and effects on customers. High-consequence actions need permissions, logs, rollback and a named person accountable for release.
The report says countries differ in the kinds of work people bring to ChatGPT, including legal research, equipment diagnosis, teaching, financial explanation and healthcare research. Those labels are relative frequencies from classified messages, not counts of jobs transformed. They cannot establish that a national workforce has adopted a tool successfully or that one occupation benefits more than another.[1][2][3]
Commercial interests and what remains unproven
OpenAI produced the report about its own products and has a direct commercial interest in presenting agentic adoption as a European economic opportunity. The data access is valuable because an outside researcher could not reproduce the internal product logs independently. That same asymmetry makes disclosure of denominators, exclusions, classification accuracy and uncertainty especially important.
No comparison group shows what happened in similar firms that did not use agents. The report measures no revenue, profit, service quality, worker wellbeing, wage, employment or productivity outcome. It also does not report task failure, hallucination, security, privacy, energy use or human-review burden. The conclusion that deeper use will produce stronger businesses and meaningful benefits for workers is an aspiration, not a tested result of these charts.
Country differences may reflect language, sector mix, customer acquisition, pricing, product availability, coding intensity or a few large accounts as well as national adoption culture. The report does not decompose those explanations. Readers should treat the charts as descriptive signals from one provider, not a league table of European readiness.[1][2][3][4]
What evidence would change the assessment
OpenAI could strengthen the evidence by publishing anonymised firm and user counts, country suppression rules, distributions around averages, product-migration effects and a reproducible methods appendix. A stable cohort analysis would separate increased use by existing customers from the arrival of new accounts and from the July launch of ChatGPT Work.
Independent studies should compare matched teams with and without agentic tools on prespecified tasks. They should measure verified completion, time after review, serious errors, costs, security events, worker control and customer outcomes over long enough periods to include novelty and learning effects. Results should be reported by occupation, organisation size, language and consequence of error rather than compressed into token totals.
The assessment would become more positive if deeper agent use produced sustained net gains without increasing serious mistakes, surveillance or uncompensated review work. It would weaken if token growth mainly reflected longer runs, product bundling or experimentation that failed to improve outcomes. For now, the report establishes rapid growth inside OpenAI's EU customer base—and leaves the economic and human result open.[1][2][3][4][5]
What this means for people
- Workers may spend less time on first drafts and repetitive computer steps, but more time specifying goals, reviewing actions and correcting failures.
- Managers need task-level quality, rework, security and worker-experience measures; token volume is not a productivity or value metric.
- Public officials should not treat a selected vendor's customer logs as an EU-wide adoption rate or evidence that benefits are shared across countries and firm sizes.
Global context
The analysis concerns OpenAI product use in EU member states and compares a selected customer group with firms outside the EU. It does not cover businesses using other providers, local models or no generative AI, so its country and regional rankings cannot be generalised to the global economy.
What the evidence does not yet show
- The report was produced by OpenAI about its own commercial products and was not independently audited or peer reviewed.
- It does not disclose the number of firms, users, messages or tokens behind the main EU and country estimates.
- Output tokens and weekly active product users do not measure task success, productivity, revenue, job quality or worker benefit.
- ChatGPT Work launched after the May baseline, complicating interpretation of the reported 8.1-fold small-business user growth.
- The enterprise, small-business and individual-message figures come from different populations and cannot be combined into one adoption rate.
What to watch next
- Publication of customer and user denominators, country suppression thresholds and uncertainty ranges
- Stable-cohort analysis separating new customers, product migration and increased use by existing firms
- Independent controlled evaluations of completed work, errors, rework, security and worker outcomes
- Comparable evidence across providers, local models, industries, languages and smaller organisations
Living evidence record
Impact record IAI-1RCJVQ2
Evidence stage
Observed
Confidence
Supported
Reporting basis
Source analysis
Independent or research support
Not yet
Record status
Monitoring
Last checked
14 October 2026
Source trail
5 direct sources across 1 source type.
People impact
Documented in this record.
Uncertainty
Limits and next checks are explicit.
Stages describe the evidence available—not whether a technology is good or bad. See the public method.
Related-source reporting disclosure
This record analyses 5 linked source records around the same underlying development. The extra records add method, date or context, but they do not by themselves constitute independent replication of every performance claim or predicted outcome.
Evidence trail
Sources used for this report
Links checked 14 October 2026
Research question, topics and regions
What OpenAI's EU usage data show about agentic products, and why tokens, growth multiples and customer rankings do not establish economy-wide adoption or productivity
Topics: AI agents · Work · Small businesses · Enterprise adoption · Productivity · Skills · Measurement
Regions: European Union · Europe · Global
This report is labelled source analysis. We summarise and analyse source material in our own words; company statements remain attributed claims until independently supported. Translated summaries preserve the meaning of the original source and link back to it. Read our editorial standards.
The Impact Brief · Free
Keep following the stories that matter to you.
Get our weekday briefing on what changed and why it matters. Choose your topics below.
Continue the story
There’s more to this story
Work & Skills
Who is gaining from AI at work?
A Gallup survey of 15,482 US employees finds that regular AI users report more speed, creativity and work quality, but frequent use is concentrated among graduates, managers and workers who already have better jobs. The results are self-reported associations, not a productivity trial.
8 min · 2 sources
Work & Skills
AI on the factory floor: who gets the gains, and who carries the risk?
A robot may take over a task while a worker keeps responsibility when something goes wrong. IBM’s survey of 800 leaders points towards job redesign, but its promises of higher productivity and fewer staff are expectations, not measured results.
9 min · 2 sources
Work & Skills
IDB projects an AI growth dividend for Latin America, with wages dependent on mobility
Preliminary scenarios put regional output between 0.3% and 5.1% higher after a decade. Workers' ability to move into expanding roles determines whether wages rise or fall in the model.
3 min · 2 sources
Reader commentary
What does this mean where you are?
Tell us what you have seen at work, in your community or in everyday life. A question is welcome too. No account is needed; we review notes before publishing them.
Explore commentary across the portal →Published reader notes
0No published reader notes yet. You can start the evidence-led discussion above.
Prefer a private correction or response? Contact the newsroom.