AI agents in European SMEs: the eightfold jump — and what Iberia still lacks
AI agents already generate over half the output of OpenAI’s EU business accounts. What those numbers say — and do not say — for Iberian small businesses.
What OpenAI’s report actually says
On 14 October 2026, OpenAI published the report “Europe’s Agentic Opportunity”, from its economic research team. The central thesis is simple: in Europe, AI has stopped being mainly a tool for asking questions and has become a tool for delivering work. An AI agent, in this context, is not a chatbot that answers: it is a system that plans steps, uses tools, works with files and carries a task through to completion, with human oversight where needed. The numbers behind the thesis all come from the same source: OpenAI’s ChatGPT Enterprise and Business accounts in the European Union. By the end of September 2026, the company’s agentic tools — Codex, its coding agent, and ChatGPT Work, launched in July — already accounted for 55% of the output tokens generated by those business accounts — 34% from Codex and 21% from ChatGPT Work. In other words, inside OpenAI’s paying customer base, work delivered by agents now outweighs conversational answers. The second number is the one behind this article’s headline: between the week of 1–7 May and the week of 29 September–5 October 2026, weekly active users of the agentic tools at small and medium businesses in the sample grew 8.1 times. OpenAI defines “SMB” for this report as organisations with 200 or fewer associated users — an accounts criterion, not the European criterion of headcount or turnover. Growth happened in every member state, with Belgium (12.1 times) and Italy (10.8 times) leading, followed by Luxembourg. There is a third number that matters: the distance between frontier firms and the rest. Frontier firms — the top 10% by output tokens per active user — generated in September more than four times the tokens per user of typical European firms. But frontier firms outside the EU generated more than twice as many as European frontier firms. EU firms make up about 10% of frontier firms globally; North America, about 65%.
The honest reading: what these numbers do not prove
Before drawing conclusions, the obvious needs saying: these are a vendor’s data about its own tools and its own paying customers. They measure activity inside OpenAI, not the European economy. A company that does not use ChatGPT Enterprise or Business does not appear here; a company that uses another vendor does not either. The report is a good map of how agents spread inside one customer base. It is a poor map of Europe as a whole. Then, tokens measure output volume, not business value. A long agentic task — reviewing a code repository, preparing a report from dozens of files — consumes far more tokens than a short answer. When the agentic share rises, that can mean deeper work, but it can also mean merely longer work. Growth in use is not demonstrated productivity, and the report itself does not measure profits, deadlines met or hours saved. There is also a calendar effect that inflates the 8.1×. Measurement starts in May, before ChatGPT Work existed, and ends in October, after its July launch. Part of the growth is therefore adoption of a new product inside the same customer base — not necessarily eight times more European companies using agents than in May. OpenAI itself says a major part of the growth followed the introduction of ChatGPT Work. Finally, the report’s SMB definition deserves a note. “200 or fewer associated users” is not the same as 200 employees: a 30-person company may hold 30 accounts, and a 500-person company may hold 150 product users. Comparing this 8.1× with official SME statistics compares different samples. None of this invalidates the report — the movement is real inside the sample — but it rules out the triumphalist reading: it is not “Europe multiplied its productivity by eight”, it is “OpenAI’s business customers use agents eight times more than in May”.
Where Iberia sits in these numbers
The agentic share varies widely from country to country inside OpenAI’s sample: from 85% in Lithuania to 22% in Slovenia, in September 2026. Portugal and Spain sit in the lower half of the table. In Spain, 30% of the output tokens of the included business accounts were agentic in September; in Portugal, 29%. These are numbers from OpenAI’s customer base, not from every company in each country — but they say something about intensity of use among those already paying for these tools: at Iberian firms in the sample, the agent is still a minority of the work. Growth tells the other half of the story. Between May and early October, weekly active agent users at SMBs in the sample grew 8.8 times in Spain — above the European median of 8.1 times — and 6.5 times in Portugal, below it. In other words: Spain starts from a low share but accelerates above average; Portugal starts from a similar share and accelerates below average. Lithuania, the share leader, grew only 6.8 times; Slovenia, last on share, grew 9.2 times. Share and growth barely relate: starting late does not prevent accelerating fast. Outside OpenAI’s sample, the Iberian picture is even more modest — and this is where official numbers come in. In 2024, according to Eurostat, only 8.6% of Portuguese firms with 10 or more employees used AI, against a 13.5% European average. OpenAI’s own report records the European rise: in 2025, 20% of EU firms with at least 10 employees used AI, up from 13.5% in 2024. Portugal therefore starts from behind — and only about 54% of Portuguese SMEs reach the basic level of digitalisation. In Spain, the IndesIA Barometer 2026, built from 131,554 companies, finds that only 4.1% of Spanish industrial SMEs use AI — it was 2.1% in 2024 and 2.9% in 2025. Adoption grows every year, but not even one in twenty industrial firms has got there. By size, medium firms (50 to 250 employees) are at 6.5%, small firms (10 to 50) at 5.1% and micro firms (fewer than 10) at 3.6%. By region, Madrid leads at 6.3%, followed by Catalonia at 4.9% and the Basque Country at 4.2%. The table puts Iberia side by side:
| Indicator | Portugal | Spain |
|---|---|---|
| Agentic share, OpenAI accounts, September 2026 | 29% | 30% |
| SMB user growth, May → October 2026 | 6,5× | 8,8× |
| Firms with 10+ employees using AI (Eurostat 2024 / IndesIA) | 8,6% | 4,1% (PME industriais) |
Portugal answers with training: PME Digital+
If AI literacy has been a legal duty under Article 4 of the EU AI Act since February 2025 — the subject of last week’s article — the next question is practical: where does a small business learn? In Portugal, part of the answer arrived on 20 October 2026, with the launch in Leiria of the pilot programme for AI training for micro and small businesses, inside the IAPMEI’s PME Digital+ programme. The pilot is deliberately small: about 50 workers in the first phase, two editions and 100 participants in total. It is promoted by IAPMEI, with support from NERLEI, the Leiria region business association, and the training — the “AI Dive for Business” cycle — is delivered by Shaken Academy. The focus is practical: using generative AI tools on each company’s concrete challenges, experimentation in real business contexts, and producing openly accessible knowledge, methodologies and use cases. The pilot is the first step of a larger project: training around 100,000 micro and SME workers across the country in AI over three years, according to the Ministry of Economy and Territorial Cohesion’s statement. PME Digital+ targets SMEs in the Norte, Centro and Alentejo regions, and enrolment and the calendar open gradually over 24 months — which means the sensible step today is to watch the IAPMEI portal rather than count on a fixed date. No company should plan its training around a calendar that is still opening. It is worth connecting the two articles: the Article 4 duty requires each company to be able to show its people have sufficient literacy for the systems they use. A public training programme does not discharge the duty by itself — the company still has to record who took part, when and on what topics — but it gives a place to start without hiring a consultant on day one. For a Portuguese SME that does not use AI yet, the Leiria pilot is less a news item than a signal: training is reaching the business fabric, not just large companies.
Spain answers with grants: Murcia’s Cheque TIC
In Spain, this week’s most concrete public response has a name, a deadline and an amount. The Cheque TIC, from Murcia’s Instituto de Fomento (INFO), opened applications on 22 October 2026 and closes them on 30 December 2026. The total budget is €900,000 and grants run from €2,000 to €20,000 per beneficiary, depending on the service and, in some cases, the company’s average headcount. SMEs, self-employed workers and joint-property entities in any sector may apply, except primary agricultural, fishing or aquaculture production, provided the subsidised activity takes place at a workplace open in the Region of Murcia. What the grant pays for is exactly the kind of first step the IndesIA Barometer suggests is missing: external specialist services in AI automation and management-process optimisation, business intelligence tools, production management and control systems (MES), digital twins, cybersecurity and data processing. It is not a cheque for buying software: it is a cheque for buying specialist knowledge the company does not yet have in-house. The contrast with Portugal is instructive. Portugal is betting, at this stage, on large-scale public training — skilling people. Murcia is betting on co-financing specialist services — buying skills. Neither response solves the problem alone: training without a process to apply it to evaporates within weeks; a contracted project without anyone inside the company able to follow it dies when the consultant leaves. The company that makes best use of these windows is the one that combines both: someone inside who understands enough to choose, and a specialist outside to deliver the first project. One calendar warning: 30 December is a short deadline and coincides with year-end. Anyone in the Region of Murcia who wants to apply should file in November, not in the final week — the paperwork for an AI project, however small, is not improvised between Christmas and New Year.
What a small business can do on Monday morning
None of this requires a revolution. The most common SME mistake with agents is trying to automate everything at once — and the second most common is starting nowhere. There is a middle path, and it starts with a question: which company process is clear, repetitive and measurable? First, pick a process, not a tool. Draft proposals from meeting notes, classifying invoices and emails by topic, summarising calls, answering customer FAQs: processes where the input is known, the output is checkable and an error costs little. An agent that drafts something a human reviews before sending is a good first agent. An agent that decides alone on prices, credit or dismissals is a bad first agent. Second, measure before automating. Count how many drafts, classifications or replies the team produces per week and how long each takes. Without that baseline, any “saving” a vendor promises is advertising. With it, in two months you will know whether the agent helped or merely moved the work elsewhere. Third, keep sensitive decisions under human control. Customer data, health data, credentials and decisions about people never go into a prompt without clear permissions — and when they do, it is with tools the company controls, not personal accounts. This is also where transparency comes in: since August 2026, Article 50 of the EU AI Act requires people to know when they are interacting with AI. A notice at the start of the chatbot or WhatsApp assistant conversation meets the requirement; pretending the bot is human does not. Fourth, record the training. The Article 4 duty is not met with good intentions: it is met with records — who took part, when, on what topics. A practical one-hour session, with examples from your own business, and a sign-in sheet are worth more than a generic course with no trace. A concrete example of an agent with this profile is WhatsApp customer service: answering FAQs, quoting prices and hours, booking appointments and saving interested contacts, presenting itself as a virtual assistant from the first message. That is what our ZapDesk demo does, which you can try live — and it is a good test for your company: if an agent can absorb that repetitive work without pretending to be human, your first process is chosen.
ZapDesk
See how a customer-service agent absorbs repetitive work: live demo, no signup.
What to watch in the coming months
Three windows will tell whether this autumn was a turning point or just an adoption spike for a new product. The first is the PME Digital+ calendar: as enrolment and editions open over the next 24 months, it will show whether training reaches SMEs in Norte, Centro and Alentejo or stays concentrated in the usual hubs. The second is the Cheque TIC closing on 30 December: a small programme, but a good thermometer of real SME appetite for contracted AI projects. The third is the next Eurostat release. If the share of European firms with 10 or more employees using AI keeps rising from 20% in 2025, and if Portugal closes the gap with the European average, the movement OpenAI sees inside its customer base will have confirmation outside it. If not, the story is different: agents grow fast where paying customers already exist, and slowly where even basic adoption has not arrived yet. The practical conclusion does not change with any of these scenarios. The European opportunity, in the report’s own framing, is to move from breadth to depth: not more companies trying AI, but companies using it intensely enough to change a process. For an Iberian SME, depth does not mean buying the most expensive tool: it means picking a process, measuring it, having a human review what the agent produces, and recording who learned what. The rest — market share, country tables, multipliers — is context. The process is the company.
Frequently Asked Questions
What sets an AI agent apart from a normal chatbot?
A chatbot answers questions; an agent carries out tasks. The agent plans steps, consults files and tools, and delivers a result — a draft, a classification, a booking — that a human can then review. In OpenAI’s report, “agentic” means specifically Codex and ChatGPT Work, not any chatbot.
Does the 8.1× growth mean European SMEs are 8 times more productive?
No. It means weekly active users of OpenAI’s agentic tools at SMBs in its sample grew 8.1 times between May and October 2026. The report measures use inside OpenAI’s customer base, not productivity, profits or hours saved across the whole economy. Part of the growth follows the July launch of ChatGPT Work.
Does my company count as an SMB in OpenAI’s numbers?
Only if it is a ChatGPT Enterprise or Business customer with 200 or fewer associated users. It is an accounts criterion, not a headcount one: a 30-person company with 30 accounts qualifies; a 500-employee company with 150 product users does too. Do not confuse these numbers with Eurostat’s official SME statistics.
How can I join PME Digital+ or the Cheque TIC?
PME Digital+ targets SMEs in the Norte, Centro and Alentejo regions and enrolment opens over 24 months: watch the IAPMEI portal, as there is no single date. The Cheque TIC is only for the Region of Murcia, with applications from 22 October to 30 December 2026 through INFO’s online office. If your company is in neither geography, use the programmes as a reference for what to ask your region for — not as an application you can file.
Do agents replace the Article 4 AI literacy duty?
No — quite the opposite: the more agents a company uses, the more literacy it needs. Article 4 requires staff who operate and use AI systems to have a sufficient level of skill to do so knowing the risks — and requires the company to be able to show it with records. An agent with nobody able to review what it produces is a risk, not a saving.