In short
- Among OpenAI's paying business customers in the EU, agents now produce more than half of all AI output. That says little about the typical European or Dutch SME.
- Official statistics show most smaller firms are not behind on agents. They are behind on the step before: AI as a fixed part of one working process. In the Netherlands, 27 per cent of firms with 10 to 50 staff used any AI in 2025.
- Use the self-check below to see where you stand. The next step is almost always one process with an owner and a review point. Often that is a simple workflow, not an agent.
Are European SMEs falling behind on AI agents? For most of them, the question comes too early. The gap rarely sits with agents. It sits one step earlier: AI that carries out a defined task inside a working process, rather than a one-off conversation in a chat window.
OpenAI's new report and the official statistics seem to contradict each other, but they measure different things. Here is what each one shows, and how to work out where your business stands.
What does OpenAI's report say about agents in Europe?
On 14 October 2026 OpenAI published The EU's next chapter in AI is agentic, with a longer report as a PDF. "Agentic" here means AI that takes a task from plan to finished result, rather than answering a question. OpenAI measures it through two products: Codex, for software work, and ChatGPT Work, the agent that has carried out tasks across apps and files since 9 July 2026.
The headline figures:
- For EU firms in the sample, Codex and ChatGPT Work accounted for more than half of output tokens by the end of September. Output tokens measure how much text a model produces.
- By country, that share ranges from 22 per cent in Slovenia to 85 per cent in Lithuania. The PDF puts the Netherlands at 68 per cent.
- Among EU small and medium businesses, defined here as firms with 200 or fewer associated users, weekly users of these tools grew more than eightfold since early May. In the Netherlands the figure was 8.5 times.
- "Frontier firms", the top 10 per cent by output tokens per active user, produced more than four times as many tokens per user as typical EU firms in September. Frontier firms outside the EU were more than twice as high again. Only about 10 per cent of frontier firms worldwide are based in the EU, against roughly 65 per cent in North America.
OpenAI's conclusion: Europe should turn broad adoption into deeper use, by moving beyond experiments and building skills and confidence.
Why these figures do not describe your business
These are a vendor's figures about its own customers: ChatGPT Enterprise and Business accounts. A firm without a business subscription, or one that works with Copilot or Gemini, is not counted. Three further caveats:
- Growth from a low base. The comparison starts in early May, but ChatGPT Work only launched in July. Part of the eightfold growth is simply a new product rolling out. That is our reading; the report gives no absolute numbers.
- Tokens measure volume, not value. An agent that searches and writes for an hour produces a lot of tokens. Whether the result was any use is not measured.
- The yardstick is the vendor's. OpenAI charges for usage. Defining leaders as the firms that use the most tokens per person makes sense for OpenAI. It is not automatically your goal.
The useful signal is different: in firms already using these tools, AI is shifting quickly from answering questions to doing work.
Where do SMEs actually stand with AI?
Official statistics cover all firms, whatever tool they use. According to Eurostat, 20 per cent of EU enterprises with ten or more employees used at least one AI technology in 2025: 17 per cent of small firms and 30 per cent of medium-sized ones. Among firms that considered AI but did not adopt it, a lack of relevant expertise was the most common reason.
The Dutch statistics office CBS shows the same pattern. In 2025, 27 per cent of Dutch firms with 10 to 50 people used AI, against 45 per cent for 50 to 250 and 66 per cent for 250 or more (CBS, provisional, in Dutch). Three quarters of Dutch firms use no AI and have not considered it. Of those that considered it but did not start, 73 per cent cite lack of experience as the main reason.
Two notes: these figures cover 2025, largely before ChatGPT Work existed, and they ask about AI technology in general, not agents.
So are you behind? Compared with OpenAI's frontier customers, yes, but they are not necessarily your competitors. Compared with firms of your size, probably not: most have not gone beyond loose experiments. The real gap is opening now, and it is about processes rather than subscriptions.
From chat to AI that finishes the job
In a chat, someone pastes something in, gets an answer and copies it somewhere else. The work stays with that one person and nobody sees what happened. AI that finishes the job takes a task from start to end within set limits: it fetches data, drafts a result, queues it for approval and writes it back into your system.
That needs four things:
- A defined task with a clear start and end, for example from quote request to draft quote in your CRM.
- Access to the data the task needs, and no more.
- An owner who decides what good enough looks like.
- A review point for anything that leaves the building.
Little wonder so many firms cite lack of experience. You need process knowledge, AI skills and technical delivery, and they rarely sit in one person. How to organise that is covered in AI implementation: the bottleneck is people.
Self-check: where are you with AI?
Find the row that fits your business best. If you are torn between two, pick the lower one.
| Stage | You recognise it by | Next step |
|---|---|---|
| 1. Occasional | A few people try things, no agreements | Agree which tools and data are allowed |
| 2. Loose chats | Many people use AI, each their own way | Pick one process with lots of manual work |
| 3. Fixed workflow | AI does a set step in one process, a person approves | Measure the result, connect your systems |
| 4. Agent with limits | AI chooses its own steps, with its own access and a log | Expand only once review and logging work |
The statistics suggest most SMEs are at stage 1 or 2. That is a starting point, not a failing. The biggest gain is in moving from 2 to 3, not in jumping to 4.
When an agent is the wrong step
More tokens is not a goal in itself. What counts is whether a task gets done faster, better or with fewer errors. An agent that runs for a long time uses more and is harder to check.
If you can write the steps down in advance, such as processing a sign-up or drafting product copy, a fixed workflow with AI at a few points is usually more predictable and cheaper. An agent only fits when the route differs case by case and the software has to act in your systems itself. The limits to set are covered in Using AI agents safely: five questions before you start.
How should an SME start with AI?
Choose one process with a lot of manual work, waiting or errors. Name an owner, agree which data the AI may use and where a person approves, and decide up front what you will measure. If it works, move to the next process. If you want to make that choice with your leadership team, the session Introducing AI in your organisation ends with a prioritised list of opportunities and a 90-day plan.
Want to know which process should go first in your business? Book a call. If it fits, we build a first working prototype within five working days, free of charge. It tests the idea; it is not yet a production version.
Sources
- OpenAI, The EU's next chapter in AI is agentic, 14 October 2026
- OpenAI Economic Research, Europe's agentic opportunity (PDF), October 2026
- OpenAI, ChatGPT for your most ambitious work, 9 July 2026
- Eurostat, Use of artificial intelligence in enterprises, data extracted December 2025
- CBS, Bedrijven gebruiken AI vaakst voor marketing of verkoop (in Dutch), 12 December 2025
