Romania's AI Adoption Crisis: 5.2% vs 20% EU Average – What It Means for Your Business

In December 2025 Eurostat published its data on artificial intelligence (AI) in enterprises, and Romania is last in the EU again: 5.2%, against an average of 20%. The gap does not come from a lack of technology but from how companies keep their data; it is the problem we solve with the platform before any agent. This analysis compares 2024 with 2025, shows what companies using AI actually do and where you start in the first two months.

Romania is last in the EU for AI adoption, for the second year running

Eurostat, December 2025: 5.2% of Romanian enterprises with at least ten employees use AI technologies, against the 20% EU average; last in 2024 as well. The ranking has the same ends as in January 2025, the north on top and the east at the bottom. Denmark is at 42.0%, Finland at 37.8%, Sweden at 35.0%; at the other end sit Bulgaria with 8.5%, Poland with 8.4% and Romania with 5.2%.

Eurostat counts enterprises with at least ten employees, so the figure says nothing about micro-firms. It does say something about companies that already have an office, an accountant and a flow of invoices: one in twenty uses an AI technology.

It is not a statistic about robots. Eurostat asks about text analysis, language generation, speech recognition or machine learning, in other words about things done on the documents the company already has.

The gap is widening: the EU rose 6.5 points in a year, Romania 2.1

The EU average rose from 13.5% to 20% in one year, Romania from 3.1% to 5.2%; the distance grew by more than four points. The two Eurostat releases, from January and December 2025, tell the story in two figures. The EU added 6.5 percentage points; Romania added 2.1, our own calculation from the two series.

Denmark rose by 14.5 points, Finland by 13.5, Lithuania by 12.5 (Eurostat, December 2025). The countries already at the top accelerated; those at the bottom took small steps. A year on pause does not leave you where you are, it leaves you further from the market you compete in.

Pace matters more than level. At the 2025 pace, Romania would reach today's EU average only in 2032, by our calculation; the EU average does not stand still, though, so the target moves.

European companies use AI on text, documents and language, not on robots

The most used AI technologies in the EU in 2025 are text analysis and language generation; they work on invoices, reports and emails, not machinery. Eurostat, December 2025: 11.8% of EU enterprises use analysis of written language, 9.5% generation of images, video or audio, 8.8% generation of written or spoken language and 7.2% speech recognition. Autonomous physical movement, meaning robots and vehicles, is at 1.4% (Eurostat, Statistics Explained, December 2025).

In a company, that means: an AI technology reads supplier invoices into the system, writes the monthly report's commentary from existing figures, turns notes dictated on site into text. None of this needs a data department; it needs data in order.

Size matters, but less than it seems. Eurostat again, on the same page: 17.0% of small EU enterprises use AI, 30.4% of medium ones and 55.0% of large ones. Small European firms are at more than three times the Romanian average.

Romania stays behind for four reasons, none of them technical

Romania has the same vendors, the same tools and engineers as good as Denmark's; the difference lies in four management decisions that keep being postponed. None of them shows up in any report, because all four are ways of not deciding.

The first is “it works well enough”. The manual process appears nowhere as a cost line: the salaries of the people consolidating data are “administration”, and the Friday report is “how it's done”. The cost becomes visible only when someone counts the hours.

The second is overestimating complexity. “AI is for Google, not for us” confuses research with use. The technologies Eurostat measures are features of tools the company already pays for or can rent by the month.

The third is underestimating the consequences. “Our competitors work in Excel too” was true a few years ago. When the client gets a quote the same day from a competitor and in three days from you, they do not care why.

The fourth is booking IT as a cost without its counterpart, the value of the process. The question “how much does the platform cost?” is asked on its own, without “how much do the months of manual reporting cost?”. Without the second question, the first always gets the answer “too much”.

You start in three phases, over twelve months

The first two months bring visible wins on one report and documents; months three to six automate core processes; only then come forecasts and agents. You do not start with everything. You pick the most painful and most visible process, usually a recurring report or invoice entry, and fix it with a figure measured before and after.

Phase two connects the sources: data from sales, stock and accounting lands in one place, and the operations dashboards update themselves. It is the phase in which Excel is closed and the team goes through the Valley of Despair; it is normal and it is temporary.

Phase three is the only one where the word “intelligence” is earned: forecasts from your own history, risk signals on projects, an AI Agent that answers from the company's files. It works only on the data put in order in the first two phases; that is why the order is not skipped.

Five mistakes that turn the advantage into a cost

The mistakes that cost are not technical: custom software, too much ambition early, dirty data, goals without a figure and the project left to IT. We see them in the same order at almost every company that has tried once before and given up.

Custom software is the first. Existing tools, for reporting, document extraction or approvals, cover the current needs of a company with a few dozen employees. Building from scratch costs more and arrives later.

Ambition is the second: automating the whole company at once stalls in month two, when nobody knows what is finished. Dirty data is the third: an Excel full of errors, automated, produces errors faster.

The goal without a figure is the fourth. “We want to be more automated” cannot be measured; “the monthly report comes out in hours, not days” can be measured and seen. The fifth is treating it as an IT project: whoever runs the process defines the need, IT implements it.

In practice: a distributor with 70 employees, the first two months

A fictitious distributor with 70 employees, three branches and 1,400 supplier invoices a month starts with documents and with the sales report. In month 0, Ana M. and Radu T. type the invoices in by hand, and the sales report by branch comes out on the 12th, after three days of consolidation. Delivery confirmations sit in email, and approvals for large orders are given by phone.

Nothing is in crisis; everything is slow. Phase 1 picks two things: invoice extraction from PDF and the sales report generated from the system, with the hours measured before starting.

In month 2, Ana M. checks only the exceptions, about 60 a month, and the report is ready on the 2nd, with an automatically written commentary. Radu T. has moved to delivery disputes, where the company was losing money without knowing how much. The figures are fictitious, but the proportions are the ones we see at clients: the entry work disappears, the checking work stays, and the decision arrives ten days earlier.

Where it does not apply

The Eurostat ranking describes companies with at least ten employees and documents to process; in three situations the gap is not your problem now.

• When the company's processes are not written down anywhere yet. If every order goes differently, AI has nothing to work on; first digitisation, then automation.
• When the volume is small. A company with a few dozen invoices a month and a one-page report does not recover the cost of an implementation. A good invoicing tool is enough.
• When the gap is in sales, not in operations. If the problem is that orders are not coming in, a faster report does not bring them; the technologies in the Eurostat statistic work on what already exists.

What next

Pick a single reporting process, the one you wait for every month, and measure its hours over one cycle: who, how long, from which sources. With the figure in hand you have three options: do nothing and stay in the 94.8%, wait to see what others do, or start next month with that process. For the concrete steps, the business process automation guide shows how the first process is chosen and measured.

Frequently asked questions

Sources

1. 20% of EU enterprises use AI technologies · 2025
2. Usage of AI technologies increasing in EU enterprises · 2025
3. Use of artificial intelligence in enterprises (Statistics Explained) · 2025