Two numbers, one dataset, opposite moods. Agent use inside European small and medium-sized companies is climbing fast. The share of European firms that use AI heavily stays small and is not moving at the same speed.
The figures come from OpenAI's October 2026 economic research update. They cover Codex and ChatGPT Work, the two products OpenAI counts as agentic, where the model executes a task instead of answering a question.
The numbers OpenAI published
Three figures carry the story. Weekly active small-business users of the agent tools rose more than eightfold across every EU member state. Agents now generate more than half of all output tokens on OpenAI's European enterprise and business platforms, up from a minority position earlier in the year. And the EU as a whole accounts for roughly 10% of the world's high-intensity AI business users, while North America holds about 65%.
OpenAI classifies business users by how intensively they use the tools rather than by headcount. The company has not published the threshold behind the "high-intensity" label, so the 10% and 65% figures describe a distribution it defines itself. Treat them as a direction, not a census.
Eight times more weekly agent users
What changed on the ground is less about model quality than about job shape. A chatbot answers. An agent completes a sequence: reading a ticket queue, pulling from an internal API, editing a file, writing the result back. That pattern only became practical for small teams once per-call costs fell and context windows grew long enough to hold a whole task.
Small firms were the fastest movers because they have the least to lose. There is no change-management programme to run when twelve people already share one tool. Across the EU, the growth in weekly agent users was spread over all member states rather than concentrated in the usual tech hubs.
The gap runs inside Europe
Country-level data makes the average less useful. The share of enterprise output tokens produced by agents ranges from 22% in Slovenia to 85% in Lithuania. That 63-point spread is wider than the gap between Europe and North America, and it sits between two EU members.
The same pattern shows up by firm type. European "frontier" technology leaders produce more than four times the output tokens per active user than the average European firm, according to the same research. Token consumption measures activity, not value. A company can burn a lot of tokens on a badly designed workflow and get nothing back. Still, a fourfold difference in volume suggests most European firms are using agents for side tasks while a small group has rebuilt core processes around them.
What one agent task costs
Token share is also a spending question. Take a mid-range frontier API at roughly $2–3 per million input tokens and about $10 per million output tokens. A single agent task that reads 50,000 tokens of context and writes 10,000 tokens of result runs about $0.20 at those rates — roughly €0.17 at an exchange rate of $1.16 per euro on 15 October 2026. Ten thousand such tasks a month lands near $2,000, or about €1,700, before caching or batch discounts.
That is a scenario, not a benchmark, and the token counts are an assumption. The shape of the arithmetic matters more than the exact figure: agents multiply the number of model calls per task, so per-token pricing carries far more weight than it did in the chat era. Cheaper tiers exist. Cached input is normally billed at a fraction of fresh input, some providers run time-limited launch discounts, and others halve their rates outside peak hours. Open-weights embedding models under permissive licences cover the retrieval step that usually sits in front of the agent, and carry no licence fee.
For short, repetitive steps with little context, local inference on a 16 GB card is still the cheapest way to run a loop, which is what our AI Arena rig exists to measure. Long-context frontier work stays with the APIs.
Deadlines that moved
European buyers now operate under rules, not voluntary pledges. General-purpose AI model obligations under Chapter V have applied since 2 August 2025. The Article 50 transparency duties — Chapter IV — apply from 2 August 2026. The European AI Office has held oversight powers since 2 August 2026, and the Council extended the compliance deadline for standalone high-risk Annex III systems from 2 August 2026 to 2 December 2027 under the 2026 Digital Omnibus simplification package.
For a company deploying agents, the practical effect is uneven. Documentation duties sit mainly with model providers, which means a European firm buying API access inherits fewer paperwork obligations than one fine-tuning a model and putting its own name on it. Article 50 transparency applies where AI interacts with people or where generated content is published, so an internal-only agent workflow faces a different set of questions than a customer-facing assistant. The Commission maintains a consolidated overview of the framework.
Practical starting points
Three things are worth measuring before buying anything else. Agentic token share, to see whether the tools are doing real work or sitting in a sidebar. Cost per completed task, which is a better number than cost per seat once output tokens dominate the bill. And the split between workflow steps that can run on a small local model and those that need a frontier API.
Data residency is the other decision. Mistral offers an EU-based API for teams that need it, and open weights under permissive licences let a company run retrieval and classification in its own rack. That is a procurement question, not a technical one, and it is usually cheaper to answer before an agent touches production data.
Does eightfold growth mean European firms have caught up?
No. The growth started from a small base inside small businesses, and the EU's share of the most intensive business users is still around 10% against 65% in North America. Growth rate and installed base are different measurements.
Do AI Act transparency rules apply to an agent used only inside a company?
Article 50 is aimed mainly at cases where AI interacts directly with people or where generated content is published. A purely internal workflow sits in different territory than a customer-facing assistant, but the company still carries general data protection duties for whatever the agent processes.
Can local hardware replace API calls for agent work?
For short steps with limited context, yes, and the cost per task drops sharply. Tasks that need long context and reliable multi-step tool use generally still go to hosted frontier models, so most teams end up running both.
The widest distance in OpenAI's own country-level data is not between continents. It runs from 22% agentic token share in Slovenia to 85% in Lithuania, two EU member states measuring the same thing in the same month.
