Artificial Intelligence in Romanian Companies

The purpose of this report is to give a more detailed picture of AI use in the Romanian private sector, starting from a survey of 1,003 companies with at least 10 employees, carried out by MIA between 30 June and 28 July 2026 on behalf of Edge Institute and the Concordia Employers' Confederation1. Eurostat data place Romanian enterprises last in the EU for the use of AI technologies (5.2%).

What is at stake in integrating AI is productivity, closely tied to the question of competitiveness: in European companies, AI adoption raises labour productivity by around 4%, without reducing headcount. For the Romanian companies that do not yet use AI systematically, that would mean an indicative opportunity cost of around 32 billion lei a year, as we set out in section 4.7.

The frequency and intensity of AI use in Romanian companies

Companies in Romania use AI more than official statistics show, but use remains fairly shallow. 61.7% of companies with at least 10 employees use AI in one form or another, but only 13.5% use it regularly, and 3.3% have integrated it into the company's processes. In most cases, use means one person writing texts with a chatbot, with no rules or processes behind it. In 56% of the companies that use AI, the technology reaches fewer than a quarter of employees.

The data show that institutional preparation is weaker still. Only 20% of companies have any employee training measure, 26% have any responsible-use measure, and 15% consider AI an important component of their strategy. Among the enterprises that use AI, the companies with the fewest internal rules are those that use it occasionally.

The factor that determines AI adoption

AI adoption depends more on the company's level of digitalisation than on its size2. AI use rises from 33% among companies with no digital tool at all to 91% among those with three or more, a difference of 58 percentage points. Between small and large companies, the difference is only 25 points. At the same level of digitalisation, size no longer has a significant effect on AI adoption.

One of the problems that came to the surface is precisely the lack of digital infrastructure: 46% of companies use none of the eight digital tools measured, and among companies with 10-25 employees the share rises to 55%, against 3% at large companies. Investment intentions, in turn, deepen the gap: 30% of companies want to invest in AI, from 6% of those with no digital tools to 75% of those with five or more. External funding appears as an identified need for 78% of the companies that want to invest in AI.

Five groups of companies

By digital foundation and intensity of AI use, companies fall into five groups:

Public policy recommendations

The current instruments either expire soon or do not reach most companies. The only grant scheme for companies that want to equip themselves digitally, the PNRR's “Digitalisation of SMEs”, has a payment deadline of 30 October 20263, and continuing the programme in its current form or in a recalibrated one requires a new source of funding. The other programmes mainly fund research and innovation, for example 58% of the POCIDIF calls for 20264, and the only funding line with “adoption of advanced technologies” in its title is reserved for ICT SMEs, with a minimum grant of 200,000 euro5.

The government sworn in on 20 October 2026 announced, in the 2026-2028 government programme, vouchers for the adoption of AI, cloud and cybersecurity in SMEs6. These would continue the support offered through the PNRR, but the budget and timetable of this new programme have not yet been specified.

Future funding should follow three principles.

1. 1

   Public money, to the companies that use AI

   First, public money for AI should go mainly to the ordinary companies that use it, not only to those that develop it, especially given that the companies that innovate are few: only 14.6% of companies with 10+ employees had innovation activities in 2016-20187. Romania looks poorly placed to win the race to develop AI, but it can win the race for targeted use aimed at raising productivity by equipping local companies.

2. 2

   Basic digitalisation, before AI

   Second, basic digitalisation has to come before AI. For almost half of companies, the start means a first digital tool for basic functions, such as electronic document management. Direct funding of AI at non-digitalised companies would have the effect of enlarging the group of unequipped but active companies, which already numbers 8,094 companies, a direction that carries a high risk of adopting AI without necessarily raising these enterprises' productivity.

3. 3

   Support also for the companies that are not looking for it

   Third, support has to reach the companies that are not looking for it as well. 39% of companies say they have not looked into AI and do not consider it relevant to them, either because it does not suit them or because they are far from any technology: three quarters have no digital tool at all, and only 2% want to invest in AI. Basic digitalisation and a free digital diagnosis, offered at scale, as well as applied examples of AI use in companies from the same sector of activity, are good directions for action.

At present, there are 27 financial instruments managed by the Ministry of Investments and European Projects and an EBRD grant covering 70% of the cost of digitalisation consultancy. These reach mainly the companies that can take out a loan, but only 10% of Romanian companies have access to bank financing, against a European average of 20-30%8.

Support has to be matched to enterprises' digital needs:

• Subsidised diagnosis and consultancy
  for companies that have not started adopting AI
• Financial instruments
  for companies planning digitalisation and AI implementation projects
• Grants
  where no other funder reaches

Artificial intelligence (AI), especially in the form of “large language models”, is already changing the global economy, and Romania will not be spared this wave, described as the next industrial revolution.

Following the 2024 Draghi report, which underlined the technology and productivity gap the EU has to close, the Competitiveness Compass, the “AI Continent” Action Plan and the “Apply AI” Strategy were drawn up. Among the Digital Decade targets for 2030 is the digital transformation of enterprises: at least 75% of EU companies should use at least one of three technologies, cloud, data analytics or artificial intelligence9. In 2025, 63% of European companies used at least one of the three technologies, but only around 20% used artificial intelligence10, even though AI is considered the technology that can close the productivity gap between the EU and the other major global powers.

Although research into this technology's impact on company productivity is still at an early stage, there are already analyses that quantify, in various forms, the benefits of AI adoption. At the level of work tasks, a generative AI assistant raised by 14% the number of requests resolved per hour by the 5,179 customer support agents of a large American software company, and among novice agents the gain was 34%11. At the level of organisational performance, the European Investment Bank shows that AI adoption raises labour productivity by around 4%, without reducing headcount. Moreover, the same study shows that companies that adopt AI are also more innovative, and their employees earn higher wages12. In the United States, companies that invested more in AI grew their sales and their headcount by around 20% more than other companies between 2010 and 2018, mainly through new products13.

At the level of the whole economy, estimates vary considerably. The OECD estimates that, over the next ten years, AI could add between 0.4 and 1.3 percentage points a year to labour productivity growth in the most exposed G7 economies, at a time when productivity in OECD countries has grown by only 1-1.5% a year in recent decades14. Other estimates are far more cautious: Daron Acemoglu calculates a rise in total factor productivity of only around 0.7% over ten years15. Among the large companies surveyed by Deloitte, two thirds already see efficiency and productivity gains, but only one in five has grown its revenue thanks to AI16. In section 1.4 we put forward a possible explanation: the gains depend on the investment the company makes in digitalising the organisation and developing its employees' technical skills, and using AI does not automatically guarantee a rise in productivity.

The definition we use is deliberately broad and covers both large-language-model tools, such as ChatGPT, Claude or Copilot, and data analysis systems, automation or robots. More importantly, the survey's central question explicitly asks that individual or informal use be included as well, that is, the situation in which an employee uses a tool on their own, without the company having set out a formal framework for its use.

This methodological choice explains why we find more frequent AI use than official statistics do. For example, Eurostat reports AI use of 5.2% for Romanian companies with at least 10 employees17, while the European Investment Bank's EIBIS survey reports an AI use rate of 22% for Romania18. By comparison, the data from our survey suggest AI use in 61.7% of Romanian companies.

The three figures do not contradict each other; they describe the same economy at different thresholds. The Eurostat result starts from a narrow definition of AI use, asking the company whether it uses a fixed list of AI technologies, which assumes that the person answering knows exactly which technologies are in use. The European Investment Bank asks to what extent data analytics and AI are used in the company, with three options: not at all, in parts of the business, or the whole business organised around them. AI users are those in the last two categories, hence the 22% figure. Our questionnaire works with the broadest definition of AI use, measuring cases of individual and informal use as well.

The data presented in the report capture further details of AI use. The existing measurements give an aggregate estimate, how many companies use AI, but do not tell us what kinds of companies do or do not use AI and in what way, and they do not link, at the level of each company, AI use to digital equipment, to declared skills, to the intention to invest or to the existence of a formal framework for its use. Moreover, they draw no clear distinction between companies that integrate AI into their organisational processes and those in which employees work with a chatbot on their own.

Romania performs poorly on AI adoption first and foremost because it performs poorly on the digitalisation of its economy. Eurostat again, through its digital intensity index, shows that in 2025 54.5% of Romanian companies had very low digital intensity, against a European average of 27.9%. At the other end, only 2.7% of Romanian companies reach very high digital intensity, against a European average of 10.1%19. In other words, Romania has twice as many poorly digitalised companies and four times fewer highly digitalised companies than the Union average.

The same figures also become targets. Through the Digital Decade, Romania has committed to objectives for 2030, measured at company level, including the 75% threshold for cloud, AI and data analytics. On cloud, Romania stands at 15.5%, against a European average of 39%20. The state has therefore taken on a measurable target, but the instruments through which an ordinary company could contribute to it are few and are coming to an end (section 1.5 and chapter 8). The report's central argument, that digitalisation comes before AI, holds in other countries too. In Romania it matters more, because we start from lower down.

The decision to invest in technology is taken in a difficult economic context. On top of the digital gap comes financial pressure: the net profit margin of non-financial companies fell to 4.6% in June 2025, and SMEs collect their receivables in 98 days, against 75 for corporations21. With such a thin margin, any discretionary investment in technology competes directly with working capital, and the pressure is greatest at small companies, precisely where the digital gap is widest.

A paradox thus emerges: the productivity gain from digitalisation and AI adoption is all the more valuable the thinner the margins, but it is precisely these thin margins that make the initial investment harder to cover. The companies that would have the most to gain are the ones that can least afford the first step. Section 7.2 gives the full figures and shows how this dynamic shapes companies' intentions to invest in AI and the sources of funding they have in mind.

Recent studies reach a common conclusion about the companies that adopt AI. The productivity gain described in section 1.1, however, depends on the company's level of digitalisation: in the European Investment Bank's study, the gains are significantly larger at companies that have made complementary investments in software, in data and in employee training, so they depend on the company's capacity to integrate the technology22.

The same paper shows how unequal the starting point is. AI adoption stands at 45% among companies with more than 250 employees and at 24% among those with 10-49, and the financially less developed European economies, Romania among them, sit well below the average of the developed ones. The OECD, BCG and INSEAD study of enterprises in the G7 reaches the same conclusion: adoption is concentrated among large companies and those that already had digital infrastructure, and the lack of data maturity is described as a fundamental barrier. The obstacle companies cite most often is the impossibility of estimating the return on an AI investment before committing the money23.

As for small companies, the OECD survey of SMEs is the closest to this report. Among SMEs that use generative AI, 91.6% use it to generate text, only 28.7% in the company's core activities, and more for simple, occasional tasks than for complex, repeated ones24. Marketing and sales lead; legal and finance come last. The most frequent reason for not using it is the belief that AI does not fit what the company does, cited by 57%25. In our data, the share is even higher: 77% of Romanian companies that do not use AI say they have not looked into it and do not consider it relevant to them (section 3.4). The questions differ between the two surveys, so the comparison is indicative. The size gap is documented in the same terms by the OECD paper on AI adoption in SMEs26.

Our data from Romania show the same thing. In addition, we show directly that the effect of company size as such on AI use disappears completely when we compare equally digitalised companies. We also measure informal use, by employees, with no decision and no budget, which official statistics leave out. The report returns to these sources in the chapters where our results confirm or qualify them, without going over them again at length. The sources can be consulted in the bibliography, in annex E.

The European regulation on artificial intelligence, the AI Act, applies in stages. The penalty regime has applied since August 2025, and from August 2026 the transparency obligations apply as well. The obligations for high-risk systems, among them the use of AI in recruitment and in employee evaluation, have been postponed to December 202727.

For a company with 10-49 employees, that is, 84% of the companies in the survey, most obligations do not apply directly. As a rule, the company is a user, not a provider of AI, but the transparency obligations still concern it, that is, informing people when they interact with an AI system or when content is artificially generated. Our data show that companies are not ready for this level of transparency: 74% have no responsible-use measure for AI at all.

At home, the National Artificial Intelligence Strategy 2024-2027 was adopted in July 2024. Its diagnosis matches ours and it proposes targets too: the share of SMEs using at least one AI technology, from 1.4% in 2021 towards 8%. At 5.2% in 2025, Romania is between the starting point and the objective28.

The content of the strategy is good, but the funding needed to meet the objectives it sets is missing. As the Edge Institute call for the digital transformation of Romania also shows, the strategy rests on other programmes, and implementation has remained an abstract effort29.

The instrument that best fits the companies in our survey comes from the PNRR: the “Digitalisation of SMEs” scheme, with grants of up to 100,000 euro for enterprises that want to equip themselves digitally. The payment deadline has been extended to 30 October 2026, so the scheme is in its final stage30.

The other grant programmes do not address this need. They mainly fund research and innovation: 58% of the POCIDIF calls for 2026 go to research and innovation31. Annex D describes each instrument, and recommendation 2, in section 8.2.2, suggests what should follow the PNRR scheme.

Beyond grants there are also financial instruments. The Ministry of Investments and European Projects manages 27 such instruments, five from the PNRR and 22 from the cohesion funds, in the form of guarantees, equity investments and blended finance, with a grant component of up to 40%. They reach companies through the Investment and Development Bank, FNGCIMM, the EBRD, the EIB and the EIF, and seven of the eight regions already have instruments of their own32. Public statements do not say how many of them can fund the adoption of digital technology. Moreover, a financial instrument assumes that the company can take out and repay a loan, and according to the EBRD only 10% of Romanian companies have access to bank financing, against 20-30% at European level33.

The government sworn in on 20 October 2026 announced, in the 2026-2028 government programme, vouchers for the adoption of AI, cloud and cybersecurity in SMEs, to replace the PNRR scheme, and a shift in support for companies from grants towards guarantees, loans and blended finance34. The budget and timetable of the vouchers have not yet been specified. Our data point to two conditions for these measures to reach the right companies: the vouchers should also cover basic digitalisation, not only AI, and financial instruments cannot replace grants for small companies, given that only 10% of companies have access to bank financing (section 8.2.2).

Chapter 2 describes the survey and the companies: how the sample was built, who answered, what variables we measure and how our figures compare with Eurostat and with the European Investment Bank's survey. Chapter 3 measures AI use, how widespread it is, how deep, what it is used for and how many employees it reaches. Chapter 4 sums up the argument in one figure and points to the following chapters for details. Whoever reads the first four chapters has covered the substance of the report, in brief.

Chapters 5-7 give more detail. Chapter 5 deals with the digital foundation and where it comes from, chapter 6 with human capital, that is, the understanding of AI, skills and governance in relation to employees, and chapter 7 with investment, funding and perceived barriers. Chapter 8 sets out the recommendations, for the five groups of companies defined in section 4.6. The annexes contain the methodology, the analytical tables, the questionnaire in short form, the public policy instruments, the bibliography and the research directions left open.

This report is based on a new nationally representative survey in Romania, carried out among companies with at least 10 employees. The survey consisted of 1,003 computer-assisted telephone interviews (CATI) and ran between 30 June and 28 July 2026. The interviews, of around 15 minutes, were conducted only with decision-makers in companies (owners, founders, directors, heads of IT/digitalisation or other top managers). The survey measured (1) the company's profile, (2) the digital tools companies use, (3) knowledge and adoption of artificial intelligence, (4) skills, consultation and practices towards employees, (5) investment and funding, and (6) barriers, from costs and skills to legal risks and the lack of support from the state.

As in any survey, the answers are self-reported: they show use, practices and intentions as the decision-makers report them, not a technical inventory of the systems present or used in companies. The full methodology, together with the weighting scheme, the effective sample and the significance tests, is in annex A, and the questionnaire, in short form, in annex C.

An important methodological choice is the broad definition of artificial intelligence: tools such as ChatGPT, Claude, Microsoft Copilot or Gemini, predictive analytics systems, intelligent process automation, image recognition or other solutions that can generate, analyse, interpret or automate content, data or processes. The definition also covers the individual, informal use of generative tools. Official statistics often leave it out, and the survey sets out to measure it and to separate it from the organised adoption of AI.

Two limitations apply to every figure in the report. The survey measures perceptions, which do not necessarily match reality, and it compares companies at a single point in time (June-July 2026), so it cannot follow the same company over time. Both matter most in the chapter on barriers, where they are discussed at length in the box in section 7.3.

Before analysing digitalisation and the adoption of artificial intelligence, we describe the companies in the survey.

After weighting, the universe of Romanian companies with at least 10 employees is overwhelmingly made up of small companies (figure 2.1). A company has on average around 43 employees, but the median company has only 18. The picture changes, however, if we look at the distribution of total employment instead of the distribution of companies: companies with 10-49 employees (83.6% of all companies) account for only 36.5% of all the employees in the research universe, medium-sized ones gather 28% of all employees, and those with 250 or more employees, although only 2.9% of companies, employ 35.4% of the workforce.

The discussion therefore depends on two dimensions: in terms of the number of companies, we are interested in the capacity of small companies to invest, but in terms of the number of people affected, large companies weigh almost as much as all the small ones put together. Public policies aimed only at large companies would miss 97% of companies (even if they would still reach a third of employees). Conversely, public policies aimed only at small companies would miss almost two thirds of employees, those in medium-sized and large companies.

By field of activity, the economy is concentrated in traditional sectors: trade, retail and distribution account for 21.9% of companies, manufacturing 19.7%, construction and real estate 13.4%, professional, scientific and technical services 10.5%, transport and logistics 8.3%, HoReCa and tourism 8.1%, education, health and social services 5.8%, and agriculture 5.2% (figure 2.2). At the opposite end, the sectors we usually associate with technology are marginal in number of companies: IT, software and telecommunications account for only 1.2% of the universe, financial services 0.9%, and media and the creative industries also 0.9%. The results on artificial intelligence therefore describe Romanian trade, production and construction, that is, the economy as a whole. The technology sector is too small a share to influence the national averages.

As for the financial size of companies, 37.7% of companies had a turnover below 2 million lei in 2025, 35.4% were between 2 and 9 million, 20.9% between 10 and 49 million, and only 6% exceeded the 50 million lei threshold (figure 2.3). Almost three quarters of companies with at least 10 employees therefore have a turnover below 10 million lei, that is, around 2 million euro. At this level, an investment in technology often competes with the money needed for day-to-day activity.

As for business ownership, 91% of companies have fully Romanian capital and a further 6.8% majority Romanian capital, so only around 2% of the universe is under foreign control, majority or full, that is, 32 companies in the sample, too few for separate analyses. If, however, we group all companies with foreign participation, full or partial, we obtain 112 companies, enough for a significant result, which we use in section 4.2. Few companies export: 73.3% do not export at all, 17.4% export only occasionally and only 9.3% export regularly (figure 2.4). Exporting matters for the rest of the report (chapters 4 and 5), because exporters have richer digital infrastructure than companies oriented exclusively towards the domestic market.

Regionally, companies are distributed fairly evenly across the five macro-regions used in the survey: Bucharest and Ilfov account for 24.1% of companies, Transylvania 23.7%, Muntenia 21.3%, Moldova 15.5% and Banat-Oltenia 15.3% (figure 2.5). Although the capital remains the densest market, three quarters of Romanian companies with at least 10 employees operate outside Bucharest and Ilfov, so a digitalisation policy built around Bucharest alone would bypass most of the economy.

The typical company in the survey is therefore a company with fewer than 20 employees, with Romanian capital, with a turnover below 10 million lei, which sells on the domestic market, operates in trade, production or construction and is run directly by its owner. These are the conditions under which AI adoption is decided in Romania, and they largely explain why, in the following chapters, the individual and sporadic use of AI tools comes well ahead of their integration into the organisation's processes.

Finally, a few words about the people who answered: 72.6% of the interviews were conducted with owners, founders or general managers, 12.3% with heads of IT, digitalisation or data, 10.7% with other top managers and 4.4% with HR, legal or compliance managers (figure 2.6). Moreover, 67.5% of respondents directly decide on investment in digitalisation or artificial intelligence and a further 25.0% take part in these decisions, so over 92% of the answers come from people directly involved in the company's technology choices. So when these respondents say they do not use, do not test or do not understand AI, they are describing their own decisions, which affect the whole company. That is why their answers matter for public policy. The role of the person answering nevertheless influences some figures: companies in which a head of IT answered have on average 3.5 digital tools, against 1.2 where the owner answered. The difference is mostly about the companies, not the respondents, because digitalised companies have someone in charge of this (annex A.6).

The table below shows the report's main variables. The full version, with all variables grouped by theme, is in table B.1 in annex B.

Almost all the variables record low values. The typical company in this survey has 10-49 employees, 1.57 digital tools out of a possible 8 and no measure to develop digital or AI skills. Almost half of companies lack the digital base on which AI could be built.

Two rows of the table sum up the report. 62% of companies use AI in one form or another, but only 17% use it regularly or integrated into their processes. The difference between these two figures is the subject of chapter 3.

Section 1.2 pointed out that our survey finds AI use in the private sector of 61.7% where Eurostat reports 5.2%. The difference is more than tenfold, but both figures are correct, because they measure different things.

Eurostat asks the company whether it uses AI technologies from a fixed list, from machine learning and natural language processing to image recognition and autonomous robots. The question requires the person answering to know exactly which of them are used in the company, and for Romania the result is 5.2%35. The European Investment Bank, through the EIBIS survey36, asks to what extent data analytics and AI are used in the company, with three options: not at all, in parts of the business, or the whole business organised around them. AI users are those in the last two categories, and for Romania the result is 22%37.

We ask what the level of use of AI tools is, “even at an individual or informal level”, and the interviewer stresses this wording. The result is 28.6% structured use (11.8% testing/piloting, 13.5% regular use in some activities, 3.3% use integrated into processes) and 61.7% any form of use.

Set side by side, the three sources give close figures where the definitions come close. The EIBIS figure for Romania, 22%, lies between regular or integrated use (16.8%) and use that also includes piloting (28.6%). That is to be expected, because “used in parts of the business” is a more demanding threshold than piloting, but a more permissive one than full integration. By size, the figures are closer still: the European Investment Bank finds 24% adoption among companies with 10-49 employees, and we find 26% structured use in the same category38.

None of the external sources covers the bottom level: the 33 percentage points of individual and sporadic use. This report measures it and shows that it represents an important part of the economy. An official statistic that counts only formal adoption misses the companies in which AI has come in through employees, with no decision, no budget and no rules, that is, precisely the group in which rules of use are most often missing (section 8.2.4).

When it comes to the use of artificial intelligence in the private sector, the survey results show that six in ten companies already use it. The figure is representative, but it hides large differences between companies. A use rate of 62% also includes the third of companies for which “using AI” most often means one or two employees using a chatbot now and then, with no organisational decisions or processes behind it. This chapter breaks that figure down into several types of use and shows us where the Romanian private sector actually stands: how often, how deeply and in what way AI is used in companies.

The survey asks companies not only whether they use AI, but how. The answers fall on five steps, from “we do not use it at all” to use “integrated into the company's processes”. The distribution is as follows:

Adding up the last four rows, we get the headline figure: 62% of companies “use AI”. But 33 of those 62 percentage points are in fact ad hoc uses. Only 29% of companies have moved past the threshold of individual use: they are testing concrete applications, use AI regularly or have integrated it into the organisation's processes. On the basis of these results, the rest of the chapter divides Romanian companies into three types of AI use: no use, informal use and structured use. Informal use describes companies that use AI exclusively at an individual and sporadic level, while structured use includes both the companies piloting AI use (12%) and those that have come to use AI regularly and systematically (17%).

We can check in the data whether the “informal use” label is apt: if a company uses AI exclusively ad hoc, it should have no institutional processes behind it (strategy, training, rules of use). The data confirm this.

Only 6% of companies with informal use have included AI as a priority in the company's strategy, against 44% of companies with structured use, a sevenfold difference. Employee training measures follow the same pattern: 13%, against 53%.

The leap in formalising AI use happens right at the threshold between individual use and piloting. Companies testing AI have it as a priority in their strategy five times as often as those with individual and sporadic use (31%, against 6%) and have a skills training measure four times as often (50%, against 13%), almost as often as companies with systematic use (56%). What companies testing AI lack, on the other hand, is breadth of use: only 4% of them say that over 50% of employees use AI, against 36% of companies with systematic use. So AI testing in Romanian companies tends to involve a small number of employees.

Even so, a few good practices exist even among companies with informal use. In this category, 3 in 10 companies say they inform their employees when AI is introduced and used in the organisation, or keep human control over AI-assisted decisions. Companies that inform their employees also have a strategy or a training programme 5-6 times as often. For almost a third of companies using AI informally, management is nevertheless involved by setting a minimum framework, even if actual use remains modest. For the remaining 69%, however, the company's management remains very little involved in developing a framework for the use of the emerging technology.

Intuition might tell us that small companies remain stuck at informal use because of limited resources, while larger ones move on to structured use. This reasoning is partly wrong.

The rate of informal use is practically identical regardless of size, around a third of companies, whether they have 10 or 500 employees. Size does not influence who starts using AI ad hoc, since the early exploration of the technology has low upfront costs. The difference appears at structured use: only 26% of small companies reach this level, against 38% of medium-sized ones and 58% of large ones. But it comes from digitalisation, not from size: at the same level of digitalisation, large companies do not reach structured use more often than small ones (chapter 4). So a policy for small companies does not need to persuade them to try AI, because they already do, but it should instead support them in moving to the next step.

Compared with company size, the structure of capital has a stronger effect on AI use, but the reason is not very clear. Companies with foreign capital, even partial, do not have more digital infrastructure than those with exclusively Romanian capital. However, starting from the same technological base, companies with foreign capital reach structured use far more often (38%, against 28%), and overall they use AI significantly more often (odds ratio 2.89 unadjusted, 2.43 after taking into account company size, digital score, sector and region; p=0.0009).

By turnover, the picture is more complicated: informal use is more frequent at medium-sized companies (10-49 million lei, 39%) than at very small or very large ones. We therefore cannot speak of a linear progression, from companies with low turnover to those with high turnover, when it comes to AI use. The peak of informal use recorded in the middle may suggest that these companies are exposed enough to experiment, but not yet organised enough to formalise.

Asked which option best describes the stage of AI adoption in their organisation, 39% of all companies in Romania choose the answer “we have not looked into it and do not consider it relevant to us”. Most of these say they do not use AI at all, but the same answer is also given by 15% of the companies that do use the technology, albeit occasionally.

This perception is not randomly distributed. It is several times more frequent at small companies than at large ones, at companies with low turnover than at those with high turnover, and at companies with exclusively Romanian capital than at those with foreign capital.

The group overlaps strongly with companies with no digital base at all: 75% of the companies that say AI is not relevant have a very low level of digitalisation (they use none of the digital tools measured in the survey), against 46% in the general population. A necessary caveat: for small companies in trade, construction or transport, the reasoning may be perfectly legitimate for their business model. The fact that a company's representatives do not see AI as relevant to the business does not automatically mean they are wrong. The data do not let us separate the companies that are mistaken from those that assess correctly, but they show us where the Romanian private sector stands: at a considerable distance from the systematic use of AI, but also from any digital foundation. An information campaign should target the most sceptical profile, the small company with low turnover and exclusively Romanian capital, and focus predominantly on basic digitalisation before AI adoption (recommendation 1).

Even among the companies that use AI, the depth of use remains modest. 56% of them have under a quarter of employees involved and only 4% reach over three quarters of staff. The great majority of Romania's “AI use”, even where it exists, is concentrated in the hands of a small number of people in the company.

Large companies reach over half of their staff more than twice as often as small ones (24%, against 9%). As in section 3.3, the difference comes from digitalisation: at the same level of digitalisation, size no longer matters. How far a company goes with AI depends on the digital foundation it has, not on its size.

Asked what exactly they use AI for, the companies that answered yes most often name generic tasks: searching for and summarising information, drafting texts and documents. At the bottom of the list are the uses that involve real technical integration: programming, cybersecurity and human resources.

The result confirms the rest of the chapter. Romanian AI use is still, for the most part, individual and general-purpose. AI remains a tool for writing and searching, not one integrated into the company's technical processes.

A related question shows which part of the company AI reaches. Marketing, sales and customer relations lead by a wide margin, at 63%, followed by the administrative area, at 52%. The profile changes very little from one sector to another: AI reaches roughly the same places, whatever the company's field of activity. The only real differences are production and operations, more frequent in industry, and research and development.

AI enters first where work means text and the relationship with the customer, and only then the technical processes. The same pattern appears in the OECD survey of SMEs39: among SMEs that use generative AI, 91.6% use it to generate text, and marketing and sales is the area where it is used most, with legal and finance in the last places. The same report observes that only 28.7% use AI in the organisation's core activities.

The three types of use have different needs.

Companies with no use at all (38% of the total, concentrated among small companies with low turnover and exclusively Romanian capital) first need basic digitalisation and targeted information. Many of them say AI is not relevant to them, but three quarters of these do not have even one digital tool (section 3.4).

Companies with informal use (33%, evenly spread regardless of size) do not need to be persuaded to try AI; they already do. Instead, they need support to move from a few employees using AI ad hoc to a practice the company recognises, formalises and supports.

Companies with structured use (29%, more frequent among large companies and those with foreign capital) more often have a strategy and the necessary resources. For them, the next step is governance, discussed in chapter 6.

The recommendations in chapter 8 develop these three tracks.

In this section we briefly illustrate the mechanisms of AI adoption in Romanian companies. Figure 4.1 shows the main factors. The most important variable is the digital foundation40: among companies with no digital tool at all, 33% use AI, against 91% of those with three or more tools, a distance of 58 points. The likely explanation is that AI needs digital infrastructure and data to be useful.

At first sight, company size is relevant too: 60% of companies with 10-49 employees use AI, against 85% of those with over 250 employees, a distance of 25 points. The difference by digitalisation, however, is more than twice as large, and the analysis shows that size matters through digitalisation. Comparing companies with the same level of equipment, the difference almost disappears: among companies with three or more digital tools, 91% of small ones use AI, against 95% of large ones (figure 4.2). A statistical model that takes digitalisation into account confirms that the effect of size becomes practically nil (annex A.8).

A company that grows may come to use AI more often, but not always, because there are also large non-digitalised companies. Still, large companies digitalise more easily or reach this need sooner: 3% of companies with over 250 employees have no digital tool at all, against 51% of those with 10-49 employees. Company growth remains a variable to watch, but the level of digitalisation is the most direct road to AI adoption.

The same pattern appears in other countries. The European Investment Bank finds that the productivity surplus from AI adoption, around 4% on average, is significantly larger at companies that have made complementary investments in software, data and training41. The OECD finds that AI adoption is concentrated among large companies and those already digitalised, and that the lack of data maturity is a fundamental barrier42. Our data show, in addition, that the effect of the organisation's size as such disappears completely when we compare companies with the same level of digitalisation.

The other factors are linked to AI use in different ways. Some factors, such as the declared level of knowledge, seem to go both ways: a company learns about AI by using it, but it is also more inclined to use the technology if a minimum level of knowledge and digital skills exists. Other factors, such as employee training or the intention to invest, appear mainly after the company has already adopted AI, but can in turn support it. Figure 4.1 marks with double arrows the links whose direction we cannot establish and with single arrows those that appear after adoption. Future studies can look deeper into the relationship between these factors, in particular through longitudinal data collection.

Foreign capital is the only exception to the mechanism described above. Companies with foreign participation are no more digitalised than Romanian ones: in both groups, 26% have three or more tools. And yet they use AI far more, 81% against 60%, and the difference holds when we take into account size, technology, sector and region. So the effect of foreign capital on AI adoption does not run through digitalisation. Section 3.3 sets out three possible explanations, which the survey cannot test.

Exporting is linked to the structure of capital: companies with foreign capital export far more often (62%, against 23%). On AI use, however, exporting behaves like size: the apparent difference of 14 points disappears completely at the same level of digitalisation, so exporting acts entirely through digitalisation.

If large companies digitalise more easily, the perceived barriers may show why. Barriers fall steadily with company size: 61% of small companies report three or more major barriers, against 51% of medium-sized ones and 37% of large ones. The fall is not explained by digital equipment. The effect of size on barriers remains even after we take the level of digitalisation into account, which is not the case for AI use. It is the only place in the analysis where size keeps an effect of its own.

Barriers are therefore linked to the company's level of digitalisation, but they do not separately explain non-use of AI: the link between barriers and AI almost disappears if we take into account how digitalised the company is. Among companies that do not use AI, perceived barriers fall steadily as the company digitalises, from an average of 3.76 to 2.70, on a scale from 1 to 5. We cannot say, however, what lies behind this dynamic: whether digitalisation reduces the real obstacles or whether digitalised companies simply perceive them as smaller. Companies do not distinguish between the seven barriers in the survey, so we do not rank them (section 7.3).

The level of understanding of AI declared in the company is closely linked to use of the technology. The link, however, runs in both directions: the level of understanding rises both with the company's digital equipment and with AI use. On a scale from 1 to 5, the average is 1.77 at companies that do not use AI at all, 2.96 at those that use it and 3.42 at those that use it regularly or integrated. Most likely, both mechanisms are real: a company learns about AI by using it, but can start earlier if it already has the necessary technical skills. In the absence of longitudinal data, this report cannot separate the two mechanisms.

The intention to invest in AI and the training of technical skills are strongly associated with AI use. Only 3% of companies that do not use AI have any skills development measure, against 32% of companies that use it and 56% of those that use it regularly or integrated. The effect of the organisation's size on the existence of skills training measures disappears when we compare equally digitalised companies. The factor most closely linked to investment in professional training is use itself.

The same dynamic applies to technology investment in AI: 6% of companies that do not use AI intend to invest, against 71% of those that use it regularly or integrated. Responsible-use policies follow the same pattern, although the gap between types of AI use is not as wide.

So these three factors do not in themselves explain AI adoption in companies; rather, they describe what changes in the organisation once the technology is in use, and can in turn support it. Consequently, the recommendations on investment in AI, professional training and responsible use are aimed rather at the companies that have already taken the first step towards integrating AI.

The factors above divide the private sector into five groups. The groups are defined by us, by two simple criteria: whether the company has at least one digital tool and how much it uses AI. The division is useful because each group has different needs.

Disconnected companies (29.5%) have neither a technological foundation nor AI. Unequipped but active companies (14.8%) use AI without a digital foundation and resort to shallow uses, such as generating texts with chatbots. Equipped but inactive companies (7.8%) are digitalised, but do not yet use AI. Experimenting companies (30.6%) use AI occasionally or in pilot projects. Integrated companies (14.6%) use AI regularly or have integrated it into the company's processes and have a digital foundation.

Weighting the groups by employment and turnover, we obtain different results. Disconnected companies account for almost a third of companies, but only 14% of employment and 11% of turnover. Conversely, equipped but inactive companies are the smallest group in number of companies, but hold 12.7% of turnover, more than disconnected companies. And experimenting companies, together with integrated ones, account for two thirds of the turnover of the economy covered by the survey.

The total turnover of the 54,674 companies in the survey universe is around 1,060 billion lei, that is, 19.4 million lei per company on average. The companies that do not use AI systematically account for 75.9% of this sum, around 805 billion lei. If we were to apply, indicatively, the 4% gain in labour productivity calculated by the European Investment Bank43 to these companies' turnover, the surplus would be around 32 billion lei a year. For equipped but inactive companies, which already have the digital foundation and account for 135 billion lei, the surplus would be around 5.4 billion lei44.

Chapter 8 sets out recommendations for each of the five groups.

Turnover does not influence AI use separately from digitalisation. Neither the level of digitalisation nor AI use varies significantly by sector of activity or by macro-region.

The digital infrastructure a company already has can largely determine its capacity to integrate artificial intelligence: for instance, records kept in digital systems, customer management through a CRM or the use of cloud can make the adoption of AI tools easier. A company still operating on paper or in scattered files usually has to digitalise first. The survey recorded the use of 8 types of digital tools in line with the Eurostat categories, namely (1) ERP, (2) CRM, (3) business intelligence, (4) an e-commerce platform, (5) paid cloud services, (6) industrial robots or autonomous vehicles, (7) digital document management and (8) software automation for internal processes.

The digitalisation of the economy is only just beginning: the most widespread tool is the digital document management system (35.9% of companies), followed by software automation for internal processes (20.8%), CRM (20.5%), paid cloud services (20.4%), e-commerce platforms (19.1%) and ERP (18.4%), while business intelligence software is present in only 12.9% of companies, and industrial robots or autonomous vehicles in 9.3% of companies (figure 5.1). The order says something about Romanian digitalisation: the most used tools are those that keep records and manage documents, that is, they digitalise internal bureaucracy, while the tools that analyse data for decision-making or automate production sit in the lower half of the usage ranking.

More important is the number of companies that use no tool at all: 46.4% of Romanian companies with at least 10 employees use none of the 8 tools measured. Almost half of companies therefore have no digital infrastructure on which to build uses of AI. The average is 1.6 tools per company (out of a possible 8), the median company uses just one tool, and only 26% of companies use 3 or more such tools. The full distribution (figure 5.2) is polarised: 14.6% of companies use a single tool, 12.9% two, 8.9% three and 17.2% four or more, so that the share of companies with no tool and that of companies with 4 or more tools together add up to almost two thirds of the universe, with the intermediate steps more thinly populated.

The most relevant differences relate to company size and are statistically significant for all 8 tools. Digital document management is used by 32% of companies with 10-49 employees, by 52.4% of those with 50-249 and by 71.8% of those with over 250 employees. Likewise, paid cloud services rise, by size, from 16.5% to 55.3%, and software automation from 17.7% to 50.6% (figure 5.3). On average, a large company uses each tool 2 to almost 5 times as often as a small company, depending on the tool.

Distinct categories of companies emerge: 51% of small companies use no tool at all, against 27.1% of medium-sized companies and only 3.4% of large ones, while the average number of tools used rises from 1.3 at small companies to 2.6 at medium-sized ones and 3.9 at large ones. The OECD documents such gaps in most developed economies45. In Romania, however, small companies have bigger shortfalls and account for 83.6% of all companies with 10+ employees, so their gap sets the level of the whole economy.

The differences are even larger by turnover (figure 5.4): 63.7% of companies with business under 2 million lei have no digital tool at all, against 45.4% of companies with 2-9 million, 28% of those with 10-49 million and only 8.3% of companies over 50 million lei. At the top of the distribution, almost two thirds of companies with a turnover over 50 million lei use digital document management (64.8%), and almost half use software automation (48.7%) and cloud (47.9%). Resources and digitalisation go together, probably in both directions: companies with resources can afford the technology, and the technology, in turn, supports growth and the acquisition of new resources.

A third dividing line, which holds at any size, is exporting. Companies that export (regularly or occasionally), that is, around a quarter of the universe, use at least one digital tool in 72.3% of cases, against 46.7% of companies oriented exclusively towards the domestic market. Exporters record an average of 2.3 tools against 1.3 for companies that do not export. The difference is notable for 6 of the 8 tools measured (figure 5.5) and is concentrated in the tools that connect the company with partners and customers, namely cloud (36.6% against 15%), CRM (28.5% against 17.6%) and ERP (25.9% against 15.2%), while digital document management, a tool for internal use, is at the same level in both groups (37.5% against 34.7%).

The gap between exporters and the rest holds in every size category (figure 4.4): 65.9% of small exporters have at least one tool, against 43.4% of small non-exporting companies, and at medium-sized companies the ratio is 89.1% against 64.6%. Exporting therefore seems to bring technology. Exporters also use AI more often, by 14 percentage points, but the difference disappears completely when we compare them with equally digitalised companies. In other words, exporting matters for AI through the technology it brings (section 4.2).

As for sector of activity or macro-region, the differences disappear almost entirely: of all the sectoral comparisons only three pass the threshold of statistical significance at a 95% confidence level, and of the regional ones only two. So the digital gap between Romanian companies is almost entirely a matter of the company's size, resources and external openness, and not of geography or activity profile. The breakdowns are in figures B.1 and B.2.

The comparison with European data confirms Romania's weak position. The European Commission's State of the Digital Decade report shows that 15.5% of Romanian companies use cloud services, against a European average of 39%46. We find 20.4%, but the two figures cannot be compared directly. We asked a single question, with examples such as Azure, Google Cloud or AWS, whereas the European statistic asks separately about each type of service. However we compare them, Romania remains at no more than half the European average. Likewise, an earlier Edge Institute report showed, on the digital intensity index, that only 23% of Romanian companies reach a high level, against 34% across the Union47. Whatever the source, the conclusion is that Romanian companies enter the competition for artificial intelligence with poorer digital infrastructure than their European competitors.

The lack of this foundation does not completely prevent AI use, but it limits it. Among companies with no digital tool at all, 33.4% already report some form of AI use, but almost entirely individual and sporadic (25.6%). Regular or integrated use is only 4.0% in this group, against 38.5% at companies with 3 or more tools (figure 5.6).

Industrial robots and autonomous vehicles are one of the eight digital tools measured above. We treat them separately, because the companies that have them look different from the rest of the economy. They are 9.3% of companies, 14.2% in industry and 34.3% of those with over 250 employees. Of these, 90% use AI, against 59% of other companies, and 45% use it regularly or integrated, against 14%.

These companies are, in fact, the most digitalised in the economy: besides robots, they use on average 4.2 other digital tools, against 1.2 at the rest of companies. The difference also shows in how they use AI. Chatbot-type tools are equally widespread everywhere: text generation appears at 59% of the companies with robots that use AI and at 60% of the others. The difference appears in process automation, 54% against 24%, and in production and operations, 47% against 33%. These companies also invest more in training: 30% have a formal training programme for digitalisation or AI, against 16% of the other companies that use AI. The differences, however, are not about the robots themselves, but about the overall level of digitalisation and size. If we compare them with companies that are just as large and have just as many digital tools, the differences disappear.

For the recommendations, this means that training needs differ by how digitalised the company is. In highly digitalised companies, AI enters processes and production, and training more often takes the form of a formal programme. In the rest of the economy, AI means mainly writing and summarising. The survey does not measure what training contains, but the types of use suggest different needs, which a single digital skills programme can hardly cover both of.

Chapter 3 shows that, in many companies, AI use has run ahead of the preparation to use the technology effectively and responsibly. This chapter measures that preparation on five planes: how much respondents say they know about AI, how present AI is in the company's strategy, whether the company invests in training its employees, whether it consults them and whether it has minimum responsible-use measures.

Few companies have taken these steps. Half of respondents say they know nothing about AI or do not know how it would apply in their company. 61% of companies have not included AI in their strategy even as a mention. 80% have no employee training measure, 76% have not consulted their employees, and 74% have no responsible-use measure, such as human control over decisions or written policies. Even the companies that already use AI are not much better off: only one in five declares advanced knowledge and only one in four has AI as a priority or an important component of its strategy. At the same time, the survey results show that these steps go together: a company that takes at least one of them tends to take the others too.

The figures in this section come from a single person in each company (the owner, a member of management or a top manager who answered the survey), not from an official assessment of the company. The level of knowledge is therefore that of the person who answered and may vary depending on who completed the questionnaire (annex A.6). Even so, the pattern is strong enough to matter. Asked directly how much they know about AI, almost half of respondents place themselves on the first two steps: they have no knowledge, or they have heard of AI without knowing how it would apply to their company. Only one in seven declares themselves advanced or expert.

Knowledge varies, like most indicators in the report, with the company's size and turnover. Large companies and those with high turnover declare themselves advanced three to four times as often as small ones. Borrowing the classification from chapter 3, only 3% of companies with no AI use declare advanced knowledge, against 8% at informal use and 39% at structured use. The structure of capital matters here only partly: companies with foreign capital more often declare basic knowledge, but not advanced knowledge (section 6.5).

If the declared level of knowledge about AI is low, the number of companies that include AI in the company's formal strategy is lower still. Only 15% of companies say AI is an important component or a clear priority of their strategy; 61% either have no formal strategy or have not included AI in it.

Here the structure of capital starts to matter: companies with foreign capital, even partial, have AI in their strategy more than twice as often as those with exclusively Romanian capital (29%, against 14%). The pattern repeats for employee consultation and responsible use, but not for training (section 6.5).

The rest of the chapter follows three concrete measures: whether the company invests in training employees' skills for digitalisation or for AI use (e.g. online courses, a training programme, hiring specialised staff), whether it consults them before or during the introduction of such tools (e.g. briefings, consultations with employee representatives) and whether minimum control measures exist (human oversight, the right to contest, written policies).

In short: only 20% of companies have at least one training measure, 24% have consulted their employees in some way, and 26% have at least one responsible-use measure. All three rise with size and with turnover, and consultation and responsible use also with the presence of foreign capital. Sector and region do not matter.

Even at large companies, responsible use reaches just over half. Improving AI governance is therefore not a problem that solves itself with company size. Among the companies that actually use AI, the figures rise (38% consultation, 43% responsible use), but remain far from a universal standard (figure 6.1).

The five steps are not independent, but closely linked. Among companies with advanced knowledge, 59% also have AI as a priority or an important component of their strategy. Among those with a strategy, 67% also have training measures, and among those with training, 70% consult their employees and 63% have responsible-use measures. At each step, the chance of also having the next one is two to four times higher than the average of all companies.

The link is strong, but the data show only that these five things go together, not that one produces the others. It is just as plausible that better-organised companies accumulate all five at the same time. The link between the steps does not change the fact, however, that almost four in ten companies that use AI have not yet reached any of them, and most have reached at most one: use has run ahead of preparation.

One result does not follow the pattern in which the five steps go together, and it matters for the recommendations.

Companies with foreign capital, even partial, far more often have AI in their strategy (29%, against 14%) and more often consult their employees (44%, against 22%). They also more often have responsible-use measures (39%, against 24%). But they do not invest more in training their employees: the difference (28%, against 19%) is not statistically significant. On knowledge, the result is split. Respondents from companies with foreign capital far more often declare at least basic knowledge of AI (79%, against 48%), but not more often advanced knowledge (17%, against 14%).

In short, foreign capital brings rules and processes, but not more investment in people. Training employees and building governance structures are two different levers, which do not automatically move together. A recommendation that deals with only one of them risks solving only half the problem.

The data in this chapter show a ladder that few companies have climbed: six in ten have taken none of the five steps, and eight in ten have taken at most two. Knowledge, strategy, training, consultation and responsible use grow together, but all start from a low level. Foreign capital is a partial exception: it more often brings strategy, consultation and rules of use, but not more employee training.

These figures also matter for the obligations described in section 1.5. The transparency obligations of the AI Act apply from August 2026, and those for high-risk systems, among them the use of AI in recruitment and in employee evaluation, from December 2027. The results above show how little prepared companies are for them.

Section 4.4 shows that understanding of AI and its use influence each other: a company learns by using AI, and one that knows more starts using it earlier. The ladder in this chapter shows which steps companies have climbed, but not in what order.

Adopting AI costs money. Beyond the digital infrastructure, the company has to pay suppliers, hire people and prepare processes and a strategy. This chapter analyses the intention to invest in AI and the barriers that reduce it. Only 30% of companies intend to invest in AI, and 78% of these need at least one source of funding from outside their own budget.

Companies that already have technology see the barriers as smaller. The difference appears at the first digital tools: on a scale from 1 to 5, the barriers index falls from 3.61 at companies with no tool at all to 3.28 at those with one or two, after which it stays almost unchanged. Barriers also fall with company size: 61% of small companies report three or more major barriers, against 37% of large ones. Small companies intend to invest less often and can rely less often on their own money: only 51% of those that want to invest would use their own budget, against 69% of large companies.

Only 30% of companies intend to invest in AI49. The intention rises strongly with the number of digital tools: from 6% in companies with no tool at all, to 38% in those with one or two, 56% in those with three or four and 75% in those with five or more. The intention also rises with company size, but less so: 26% at small companies, 46% at medium-sized ones and 65% at large ones. Digitalisation matters more than size. Even among small companies, 60% of those with three or more digital tools intend to invest, against only 5% of those with no tool at all.

The companies that intend to invest are, in fact, a different kind of company from the rest. They have on average 3.3 digital tools, against 0.9. Nine in ten already use AI, against one in two. Over half have employee training measures, against only 6% of the others. The intention to invest therefore appears mainly where AI is already in use, as section 4.5 also shows.

Among the companies that intend to invest in AI, only 19% would use their own budget exclusively, and 78% need at least one outside source50. European funds and public grants are mentioned by 52%, almost as often as the company's own budget, at 53%. Small companies look to European funds to the same extent as large ones, 53% against 53%, and rely less on their own budget, 51% against 69%. Bank credit, leasing and other private financing appear at 19% of the companies that intend to invest. The small share is not surprising: these instruments usually finance physical assets or broader projects to digitalise the business, such as automating a production line or integrating digital flows into the company's processes.

The small share of credit is not only about the nature of investment in AI. The National Bank of Romania's Financial Stability Report51, of June 2026, shows that Romanian companies finance themselves externally mainly through trade payables, that is, by delaying payments to suppliers, not through bank credit. Moreover, according to the EBRD, only 10% of Romanian companies have access to bank financing, against 20-30% at European level52. Plans to invest in AI therefore follow a general pattern. There are nevertheless signs of appetite for investment: investment loans grew by 19.5% in a year, to March 2026, and reached 30.8% of the portfolio.

Nor is there more of the companies' own money than before. According to the same report, the net profit margin of non-financial companies fell to 4.6% in June 2025, and the debt ratio rose to 140.5%. Operating costs rose by 10.7%, and wage costs by 15.4%. The pressure also shows in the risk signals. The non-performing loan ratio of companies rose to 5.6% in March 2026, by 1.2 percentage points in a year. The number of companies in insolvency has been rising since the second half of 2025. Against this background, for many companies, the need for outside money is a constraint rather than a preference.

The pressure is greater at small companies. SMEs collect their money from customers after 98 days on average, but pay their suppliers after 80, so they end up financing almost three weeks of activity from their own resources. At corporations, the difference is only three days (75 and 72 days). That could explain why small companies rely less often on their own budget. With a net margin below 5% and money tied up with customers, an investment in technology competes directly with the money needed for day-to-day activity.

To note for the recommendations: the need for external funding is high at all companies, but its role differs by size. Large companies more often intend to invest and more often combine their own money with external sources. Small companies intend to less often, and when they do, almost half of them do not mention their own budget at all, against a third of large companies. For them, external support is more often the only possible source.

Companies that intend to invest report somewhat smaller barriers than the others: on a scale from 1 to 5, their index is 3.29, against 3.42, and 49% of them have three or more major barriers, against 60%. Barriers nevertheless remain high for all companies.

The difference is mainly about who these companies are. The ones that see smaller barriers are the same ones that already have digital tools, trained people and experience with AI. Probably the obstacles are not smaller for them; rather, the companies are better prepared to overcome them. This shows in how the index falls. Taken together, companies see far smaller barriers once they have their first digital tools: the index falls from 3.61 at those with no tool at all to 3.28 at those with one or two, after which it stays practically unchanged. Likewise, the share of companies with three or more major barriers falls from 68% to around 50% and then no longer changes.

The explanation is that companies that already use AI perceive barriers almost the same, however many digital tools they have. Among companies that do not use AI, on the other hand, barriers fall gradually as the company digitalises, from 3.76 to 2.70. Barriers also fall with company size, from 3.47 at small companies to 2.86 at large ones. Here size matters in its own right too: large companies see smaller barriers even when they have the same level of digitalisation as small ones. For AI use, on the other hand, size matters only through the digitalisation it brings.

Beyond the limitations in the box below, these results have to be read with one caution. The survey asked companies about seven different barriers, from costs and the lack of specialists to legal risks and resistance to change. Companies, however, did not distinguish between them: whoever rated one as high rated them all, as a rule, the same, and one in six companies gave exactly the same answer to all seven. The answers therefore show how hard they find it, in general, to adopt AI, not which particular obstacle stops them. That is why we do not rank the barriers against each other.

How many companies could invest in AI if they had access to funding? Of the 54,674 companies with at least 10 employees, around 16,400 intend to invest in AI, and 78% of these need at least one source of funding from outside their own budget. There are, then, approximately 12,800 companies that have already decided to invest but depend, at least in part, on external funding53. We cannot say how many of them would actually invest if they had access to money, because the survey measures the intention, not the investment.

Most of the companies that want to invest already use AI: almost nine in ten are in the experimenting and integrated groups (section 4.6). At the other end, for the 16,129 disconnected companies, with no digital foundation and no AI, a funding line for AI would suffer from a lack of demand: only 4% of them intend to invest in AI. Equipped but inactive companies, that is, the 4,257 enterprises that already have digital technology and do not use AI, are a group apart. These companies have something to build on and account for 12.7% of the turnover of the companies in the survey, but only 14% of them currently intend to invest. For them, money alone is probably not enough; information about what AI could bring them is needed too.

AI use in Romanian companies is widespread but shallow: 62% of companies use it in one form or another, but only 17% use it regularly or have integrated it into the company's processes. AI use depends first and foremost on the digital foundation, not on size. Larger companies use AI more often, but mainly because they are more digitalised. A medium-sized company with no digital tool at all uses AI far less often (29%) than a small company with three or more tools (91%). Knowledge, training and the intention to invest appear mainly at companies that already use AI, which suggests that they come after adoption rather than before it. Companies that do not use AI more often consider it irrelevant to them and perceive higher barriers.

Digitalisation is therefore the condition for real AI use. Companies without digital tools can use AI too, but rather at an individual and occasional level: only 4% of them use it regularly, against 38.5% of companies with three or more tools. That is why public policy should first support basic digitalisation and only then AI adoption.

The report divides companies into five groups, by two criteria: whether they have digital tools and how much they use AI. Each group has different needs. Disconnected companies need basic digitalisation before anything else. For unequipped but active companies, what matters are the systems and rules around a use that has already begun. Equipped but inactive companies already have the technology, but do not use AI. They are only 7.8% of companies, but account for 12.7% of turnover, and their digital foundation makes them the easiest to bring towards AI, even if few intend to invest for the moment. Experimenting companies need to move from individual use or pilot projects to integrating the technology into the company's processes. Integrated companies need deeper use, skills and governance rules. The recommendations below address these groups in turn and say, for each, how many companies it targets.

The result: AI use rises from 33% at companies with no digital tool at all to 91% at those with three or more. Size no longer matters when we compare equally digitalised companies: at the same level of equipment, a large company does not use AI more often than a small one (odds ratio 1.54 without controlling for digitalisation, 1.00 with). Company growth remains useful, because large companies digitalise more easily, but support can go directly to digitalisation, at companies of any size.

Target group: disconnected companies (16,129 companies, 29.5% of companies, 11.4% of turnover)

A large part of this group will not seek support on its own. 39% of all companies say they have not looked into AI and do not consider it relevant to them, and among disconnected companies the share reaches 82%. The answer may reflect a correct assessment for their business model, but also a long distance from technology adoption: three quarters of the companies that give this answer have no digital tool at all, only one in five declares at least basic knowledge of AI and only 2% intend to invest in it. Almost nine in ten are companies with 10-49 employees, from every sector. For them, a funding line for AI would have no effect, because almost nobody would apply for it.

Singapore shows how such a diagnosis can work at scale. The SMEs Go Digital programme, launched in 2017, had reached by 2024 around 88,000 SMEs, almost a third of Singapore's SMEs. Since 2022, the CTO-as-a-Service platform starts with an online self-assessment of digital maturity, after which the company receives free consultancy and can choose from over 450 pre-approved digital solutions, with grants covering 50-80% of costs. An impact evaluation found that grants for basic solutions raised value added per employee by 3%, and at micro-enterprises by up to 6.4%. Solutions designed for a particular sector were almost twice as effective as generic ones57. The lesson for Romania is the order: it is important for the diagnosis to come first, and for funding to follow the objectives set, or even be conditional on achieving them.

The result: 30% of companies intend to invest in AI, and 78% of these need at least one source of money from outside their own budget. Small companies rely far less often on their own budget than large ones (51%, against 69%), although they look to European funds just as often (53% in both cases). The economic context could explain the difference: companies' net margin is 4.6%, and SMEs collect their money from customers after 98 days, against 75 at corporations.

Target group: equipped but inactive companies (4,257 companies, 7.8% of companies, but 12.7% of turnover). They already have the digital foundation, so an investment in AI would have something to rest on. Only 14% of them currently intend to invest, so funding should come with information about what AI could bring them.

The right instrument exists, but it is in its final stage. The PNRR's “Digitalisation of SMEs” scheme, with grants of up to 100,000 euro for the adoption of digital technologies, has a payment deadline of 30 October 202658. We recommend that the scheme that follows it be launched as soon as possible and that the programme's continuity be ensured, which matters more than novelty: a scheme companies already know can produce effects faster than a new project.

In the absence of this scheme, a gap in public funding remains. SMEs are eligible almost everywhere in POCIDIF, but 58% of the calls for 2026 fund research and innovation, 24% defence capabilities and 10% the development of new products. For the company that just wants to buy and use existing technology, almost no money is left59. The condition to innovate excludes most of the economy: according to INS, only 14.6% of companies with at least 10 employees carry out innovation activities (2016-2018)60. That is why the scheme that follows should fund the adoption of existing technology, without requiring the company to innovate. It is the most direct way to reach equipped but inactive companies and, through the basic digitalisation of recommendation 1, disconnected ones: together, 37% of companies and 24% of turnover.

The state, however, is not the only funder, and each type of instrument suits a different group. Most of the companies that want to invest already use AI: almost nine in ten are in the experimenting and integrated groups. For their broader digitalisation projects, in which AI is integrated into workflows or into production, the existing financial instruments fit, that is, guarantees, loans and blended finance through the Investment and Development Bank, FNGCIMM, the EBRD, the EIB and the EIF, as well as through commercial banks (section 1.5). For equipped but inactive companies, which for the most part do not intend to invest, subsidised consultancy and diagnosis fit better, such as the EBRD grant that covers, through partner banks, 70% of the cost of digitalisation consultancy61. Public grants remain necessary where no other funder reaches, that is, at small companies without access to bank credit.

Spain offers a model for the next funding scheme. The Kit Digital programme, funded from the European recovery funds with 3.067 billion euro, had granted by the end of 2024 around 530,000 aids to SMEs and self-employed persons63. A company with 10-49 employees receives a voucher of up to 12,000 euro, which it uses for solutions from a catalogue: a website, e-commerce, customer management, electronic invoicing or cybersecurity. The solutions are implemented by accredited providers, around 11,000, mostly small IT companies, which have to provide the service for 12 months64. The model answers two of the details above: support also covers use, not only purchase, and the amount is calibrated to the company's size.

The 2026-2028 government programme announces a scheme built on the same model: vouchers for the adoption of AI, cloud and cybersecurity in SMEs, replacing the PNRR scheme, with a simplified procedure and providers accredited through framework agreements (section 1.5). The proposed scheme directly targets this recommendation, because it funds the adoption of existing technology, without requiring the company to innovate in order to access funding. The programme does not, however, specify four aspects that will shape its success: a budget and timetable that ensure continuity after the PNRR scheme; coverage of the costs of use and maintenance, not only of the purchase; amounts calibrated for companies with 10-49 employees; and the inclusion of basic digitalisation, with an assessment of digital maturity before support for AI, so that AI comes after the digital foundation, not in its place (recommendation 1).

The result: understanding of AI is closely linked to use. On a scale from 1 to 5, the declared level of knowledge rises from 1.77 at companies that do not use AI to 3.42 at those that use it regularly or integrated. Training measures, on the other hand, come after adoption: only 3% of companies that do not use AI have any, against 56% of those that use it regularly. Training does not start adoption, but knowledge can.

Target group: all companies, with priority for the 16,729 experimenting ones, where management is now debating whether to go further with AI adoption. Closest to the next step are the around 5,400 companies in this group that have reached the stage of piloting the technology. They have on average 3.4 digital tools, more than twice the economy's average, over half already have an employee training measure, and 69% intend to invest. What these companies lack, on the other hand, is the move from pilot to process: only 4% of these companies use AI with more than half of their employees.

Training should address owners and managers, not only technical staff. In 73% of companies the owner or general manager answered, and their level of knowledge about AI is the lowest of all roles. The decision is therefore taken where knowledge is scarcest. Support works best when it results in a written strategy. The first mention of AI in the strategy separates the companies that use AI from those that do not (92%, against 41%), and turning it into a priority separates regular use from occasional use (57%, against 19%). For pilot companies, the second step is the one that matters: 77% already have AI mentioned in their strategy, but only a third have it as a priority.

The existing programmes show that demand exists, the limiting factor being the size of public support. The PNRR scheme “Advanced technology skills for SMEs” trained 2,453 employees from 984 SMEs, almost two and a half times the initial target of 1,000 employees65. Even so, our survey finds that 80% of the around 54,700 companies with at least 10 employees have no training measure for digitalisation or AI. A future programme should be designed on a different order of magnitude and aimed at companies' management as well.

An example of a programme aimed at management comes from the United Kingdom. Help to Grow: Management, launched in 2021, is a 12-week course for SME leaders, with 12 modules, individual mentoring and peer learning, taught by 66 business schools. The participant pays 750 pounds, with the rest of the cost covered by the state. Between 2021 and 2025, 9,847 SME leaders took part. The official evaluation shows that participating companies survived more often, but does not yet find effects on growth or productivity, which usually appear after at least five years66.

To these is added the link with the training of IT specialists, where Romania sits below the European target. The survey does not measure the effects of AI on jobs, and the European literature finds no reductions in the short term. The European Investment Bank recommends that retraining programmes emphasise the skills that make people and AI complementary, such as working with prompts, managing data and keeping the final decision with a human67.

The result: 74% of companies have no responsible-use measure for AI at all, such as human control over decisions, the right to contest or written policies. Among the companies that use AI, the weakest situation is at those that use it individually and occasionally: only 32% have any measure, against over half of those that use it regularly. Rules therefore appear mainly where use has spread, and the period between the first uses and the first rules is the one with the highest risk.

Target group: unequipped but active companies (8,094 companies, 14.8% of companies). They use AI with no digital systems behind it and, almost all, with no rules: only 12% have any responsible-use measure.

For a small company, a responsible-use policy written from scratch is a cost it rarely takes on. The solution is for a minimum set of guidelines, ready to apply, to come with the tool, either from the provider or through a template made available free of charge by state institutions. Data preparation and cybersecurity should be treated as conditions of AI adoption, not as steps after the technology is implemented. The starting point is informing employees and human control over decisions taken with the help of AI, because they are the simplest to put into practice. The same principle applies to the obligations of the AI Act: for small companies, they become achievable and easier to navigate when they come together with the tool68.

The four recommendations have to be applied in order. Applied in reverse, they do not work: funding without a digital foundation risks enlarging the group of unequipped but active companies, and guidelines written before the first use stay on paper. A disconnected company needs the first two, an experimenting one the last two.

To the logical order is added one of timing, since the PNRR scheme that funds the digitalisation of SMEs has a payment deadline of 30 October 2026, and the current national strategy covers the period up to 2027. The window in which recommendations 1 and 2 can be implemented with the existing instruments is therefore short. The next Multiannual Financial Framework, 2028-2034, is the moment when they can be rebuilt, and the discussion has already begun: the Ministry of Investments and European Projects has prepared with the World Bank two documents for the National and Regional Partnership Plan, the framework for European funding in the coming period69. The recommendations in this report should enter that discussion now.

The timetable of this discussion is already set: according to the 2026-2028 government programme, the National and Regional Partnership Plan 2028-2034 is to be sent informally to the European Commission by 31 December 2027, after public consultation. The 2027 consultation is therefore the moment when recommendations 1 and 2 can enter the programming of the next seven years.

The four recommendations describe instruments, but governance and accountability remain questions that cannot be sidestepped. The National AI Strategy 2024-2027 shows why this matters. The diagnosis the document starts from matches ours, its objectives are reasonable and the targets it proposes are clear. What it lacks, however, is a budget of its own and a structure accountable for putting it into practice, and implementation has remained, so far, an abstract effort (section 1.5). For the recommendations in this report not to end up in the same situation, each needs a responsible institution, a budget and a timetable.

The Edge Institute call for the digital transformation of Romania argues that the main blockage is institutional accountability and proposes a set of concrete steps: an explicit mandate at deputy prime minister level, a Government Chief Information Officer (CIO), a digital council with binding decisions, backed by government capacity on AI matters, with guidelines, a procurement framework and a regulatory sandbox. The estimate in the same document is that coherent, coordinated and predictable digitalisation can bring annual growth of 1-1.5% of GDP70.

The problem was topical in September 2026 as well, including as regards budget allocations: the minister then negotiating the future National and Regional Partnership Plan said he had a mandate to negotiate, but not the institutional mandate to decide the allocations, because every ministry or party in the coalition would want control over a part71.

The 2026-2028 government programme moves the coordination of digital transformation and AI to the Prime Minister, with a Government CIO, as the Edge Institute call also proposed72, but only for the administration's projects. For AI adoption in companies, however, responsibility remains divided between several ministries. We recommend that it fall to a single institution, accountable from one end to the other. The role suits the minister delegate for digitalisation and artificial intelligence in the Niculescu Government73: the coordination of the vouchers, of training and of the financial instruments for AI adoption in companies would thus come together in one place, with a budget, a timetable and public indicators.

The four recommendations presuppose that someone is accountable for their implementation, that there is an identifiable funding line, administrative capacity and that companies actually come to benefit from the support offered.

The report has the limits of a single survey wave, so we cannot say in which direction the links we see run, that is, whether companies that understand AI adopt it earlier or whether using the technology leads to the development of the necessary technical knowledge. Moreover, we measure the intention to invest in AI, but not how much companies actually spend. At the same time, we are not in a position to say whether AI use brings any measurable benefit, because the survey contains no financial data. Annex F describes what would resolve each of these limits: a second wave applied to the same companies, linking to balance sheets, investment audits on a sub-sample, sectoral samples and qualitative research.

What separates the companies that use AI from those that do not is, first and foremost, the digital foundation, and company size matters mainly because it brings that foundation with it. Almost half of Romanian companies use none of the eight digital tools measured. For them, the first step is not an artificial intelligence project but a first digital tool, for example a document management system, the most widespread of them.

The difference between digitalised companies and the rest shows how much room for growth there is. Among companies with no digital tool at all, 33% use AI, almost all occasionally. Among those with three or more tools, 91% use it. The five groups in section 8.1 show where each company stands on this road and what it needs for the next step.

Part of the gain can come from small changes. Companies that mention AI in their strategy, however vaguely, use it far more often than those that do not mention it at all (92%, against 41%). Those that have made it a priority use it regularly far more often (57%, against 19%). And the most frequent answer of companies that do not use AI is not a concrete obstacle, but the fact that they have not looked into it and do not consider it relevant (39.1% of all companies). For some of them, the conclusion may be correct. For others, the blockage is one of perception and can be reduced through information, examples and good practices, at low cost to the state.

The existing policies can cover a large part of the road, if they are continued and improved. The PNRR scheme for the digitalisation of SMEs answers exactly the need described in this report, the adoption of existing technology, and deserves a successor that keeps this emphasis, unlike other public funding in the field of digitalisation, oriented mainly towards innovation. Just as important are simpler access to the existing programmes and a lighter administrative burden, which frees up time and money for small companies, where resources are most limited.

Human capital remains essential. Employee training appears mainly after companies start using AI, and knowledge about AI is lowest precisely among owners, who take the decisions. Public support for training should therefore start with companies' management and remain closely tied to the needs and demands companies express through consultations.

Not everything rests exclusively with the state. Many of the steps are within reach of companies and of the organisations that represent them: the first digital tool, a mention of AI in the company's plan, a minimum set of rules of use, training the owner or the management, the exchange of examples between companies in the same sector. The companies that have already taken these steps show what the rest of the economy can look like.

We thank our partners at the Concordia Employers' Confederation for their collaboration and contribution to this study.

We thank the MIA team for their support during the data collection stage.

Last but not least, we thank the respondents who took the time to answer the survey.