The adoption numbers

Percentages circulate on this subject that look as though they contradict one another. One Italian worker in two says they use AI at work; 16.4% of Italian firms use at least one AI technology; 8% of Italians aged 16 to 74 say they have used it for work. These are three true measures, from the same year, with three different bases. Here are the numbers with their scope alongside them, the year of the survey and who collected them.

How many firms use it

Eurostat and Istat ask firms the same question across the Union: do you use at least one of eight AI technologies, from written language analysis to speech recognition, from text or image generation to workflow automation? The sample is firms with at least ten employees, excluding the financial sector. Last year surveyed: 2025.

Eurostat, survey on the use of information technologies in enterprises, 2025 data, recalculated line by line on the database on 3 October 2026. The Italian figure is the one Istat sends to Eurostat, and indeed the two sources agree. The series is standing still: the data were last updated on 15 June 2026 and 2026 figures do not yet exist.

Italy is eighteenth among the twenty-seven member states, three and a half points below the European average. At the top is Denmark with 42.03%, at the bottom Romania with 5.21%. Istat shows the other face of the same figure: 83.6% of Italian firms with at least ten employees use no AI technology at all.

The line that explains the others is the one on size classes. Above 250 employees more than one firm in two uses it, below 50 one in seven. In Italy the second category is almost every firm there is, and that is where the distance sits between the story told about AI and the country's productive fabric. Both values doubled in a year, and the gap between them stayed where it was.

Below ten employees Eurostat does not look, and that floor now has a number. On 8 September 2026 the Bank of Italy published its survey on the financial and digital literacy of micro-enterprises: 5,000 owners of firms with up to nine employees, a sample stated to be representative, surveyed in 2025. Artificial intelligence is adopted, extensively or in a limited way, by 5% of these firms; another 19% are trialling it or expect to adopt it within a year; 73% say they are not familiar with the technology or have planned no use for it. It is the floor missing from the table above, which starts at ten employees and leaves out most Italian firms.

A figure published on 15 September 2026 says one in five, and it does not contradict the 16.4%. Unioncamere and Dintec have published an analysis of more than 20,000 assessments collected over two years with SELFI4.0, the self-assessment tool of the Punti Impresa Digitale: the share of firms saying they use AI solutions goes from 13.1% to 19.5% between 2024 and 2025. These are firms that choose to fill in a digital assessment, not a drawn sample, and the question is not Eurostat's. The 19.5% and the 16.4% do not compare and do not replace one another: they are two different measures of the same country.

Large Italian firms: where the chain thins out

The Osservatorio Artificial Intelligence of the Politecnico di Milano interviews large firms, SMEs and workers every year, and presents the results in February. These come from the 2025 survey, presented on 5 February 2026. Put in descending order they describe a chain that loses pieces at every step.

Between the first line and the last there are seventy-five points. Buying the licences is something almost the whole set has done; measuring whether they are worth anything is attempted by little more than half; doing it regularly is managed by one in nine; having someone answerable for how AI is used in the company remains rare.

SMEs are an order of magnitude behind: 8% have AI projects, 15% among medium-sized firms and 7% among small ones. The Italian AI market is worth €1.8 billion in 2025, 50% more than in 2024. 19% of that spending comes from the public administration, which is a share of the market and not a share of administrations using AI.

People

Eurostat asks people aged 16 to 74 whether they have used generative AI tools in the three months before the interview. The 2025 European survey rests on about 330,000 individuals.

The percentages are calculated on the whole population aged 16 to 74, in work and not. That is why the 8.00% for work sits without contradiction alongside the 47% of workers the Osservatorio finds inside firms: the two bases are different.

Italy is second to last in the Union, ahead of Romania alone. Among young Europeans aged 16-24 the share rises to 64%.

On the supplier side the figures exist, but they are three different yardsticks. On 11 September 2026 OpenAI wrote, in an engineering piece on storage, that it serves "over 1 billion people each week"; on 31 March 2026 it stated more than 900 million, again weekly, with more than 50 million subscribers, and that subscriber number has not been updated since. On 11 August 2026 Google stated that the Gemini app had passed one billion monthly users. On 30 July 2026 Microsoft stated more than thirty million paid seats of Microsoft 365 Copilot. Anthropic and Meta do not publish dated, comparable user figures.

Weekly, monthly and seats sold are three different things, and putting them in a league table says nothing. A published definition does exist, and it is OpenAI's: in the study on ChatGPT adoption of 30 June 2026 a user is active "if they have sent a message in the seven days before the start of each month". It says when a user is active, not whether a user is an account or a person, it holds for that study and it is not a methodological note attached to the announcements. No third party verifies any of these figures.

Around the world: how many say so, how many see it in the accounts

McKinsey's annual survey on the state of AI is the source almost every global adoption percentage in circulation comes from, including Stanford's AI Index, which republishes them. Latest edition: "On the road to ROI", 25 August 2026, 1,719 respondents in 97 countries, questionnaire in the field from 4 May to 8 June 2026. The previous edition, of 5 November 2025, counted 1,993 in 105 countries and closed at the end of July 2025: these are two distinct samples, not a series.

The first line and the fifth come from the same questionnaire and often end up in the same headline. The first says that somebody in the company has opened an AI tool. The fifth says whether this shows up in the accounts, and among those who answer yes only a minority reach 5% of EBIT: they are the "high performers" of the last line, about six in a hundred, a share McKinsey reports as unchanged on the year before.

Nothing is subtracted between the 2025 edition and this one. Two lines from 2025 are not in this one: the share of organisations trialling at least one agent, 62%, and the share reporting at least one negative consequence, 51%. The 2026 edition measures something else: who has scaled agents. And the question about negative effects moves from organisations to people, where 47% of middle managers and individual contributors report them against 31% of executives. On enterprise scaling McKinsey now writes "44%, up from 38% a year ago", while the 2025 edition wrote "about a third": the difference between the two readings is a restatement of the estimate, not measured growth.

There is an official statistic too, and it updates every fortnight. The US Census Bureau asks a drawn sample of firms whether they have used AI in any business function in the last two weeks. In the 24 August-6 September 2026 round, published on 24 September, 23.8% answered yes, against 23.2% in the previous round and 22.4% in the one before that; 9.7% answer that they do not know. Asked whether they will be using it in the next six months, 27.6% say yes. It is a different base and a different question from the McKinsey questionnaire, and that is why the 23.8% and the "nearly nine in ten" sit together without contradicting each other.

Whoever fills in a survey about AI is usually whoever deals with it: the sample is a panel of executives who agree to respond, not a random draw of firms. Eurostat, Istat and the US Census do not have this problem, McKinsey and the Osservatorio do. Use the first to know how many firms, the second to know what those who do it are doing.

Work

The book explains why the high percentages measure automatable hours of work and not jobs that disappear. Here are the figures, each with the definition that belongs to it.

Hours, not jobs. The McKinsey Global Institute, in its report "Agents, robots, and us" of 25 November 2025, splits the hours worked in the United States in 2024 as follows.

The 57% that circulates is the sum of the last two lines. The report writes that the estimate "reflects the technical potential for change in what people do, not a forecast of job losses". The scope is the United States alone and paid hours alone, which are about 45% of waking hours. And it is the cautious scenario among those considered: in the fast one, agents and robots reach 60-70% of hours worked worldwide.

The same institute has redone the count for Europe, and it comes to 58%. The European version of the report, of 12 May 2026, estimates that 58% of hours worked in ten European countries could theoretically be automated with existing technology: 44% by agents and 14% by robots. The total resembles the American one, and the composition changes with the mix of industries.

The forecast balance stays positive, and the source has not moved. The World Economic Forum's Future of Jobs Report estimates 170 million jobs created and 92 million eliminated by 2030, with a net balance of 78 million. As at 3 October 2026 the latest edition is still the one of 7 January 2025, and the Forum's publications index, read again today, stops at the Travel & Tourism Development Index of 25 September 2026: the series runs to five, from 2016, 2018, 2020, 2023 and 2025, and a 2026 report does not exist. If you find a "Future of Jobs Report 2026" cited, it is other work published by the WEF in January 2026, with a different scope and method.

The Italian estimate, and the order in which to read it. In October 2025 the Osservatorio of the Politecnico di Milano presented to the Chamber of Deputies two figures that need to be kept the right way round: about 50% of full-time-equivalent jobs in Italy is theoretically automatable already today, and 18% could be so in practice by 2033, that is the work of about 3.8 million people. The two figures often circulate the wrong way round, with the 18% attributed to today and the 50% to 2033: the numbers are right, the year and the adverb are not. "Full-time-equivalent jobs" are units of work, not heads. And the same research sets beside that 18% a number that almost always disappears from the citations: by 2033 Italy would be short of 5.6 million equivalent jobs for demographic reasons.

What has already been seen to happen. Erik Brynjolfsson's group at the Stanford Digital Economy Lab measures employment on ADP payroll records, millions of American workers, rather than on forecasts. The version revised on 12 August 2026, with data up to June 2026, says six things.

1. There is no evidence of a job displacement spread across the whole economy.
2. Employment among 22-25 year olds in the occupations most exposed to AI is 19% below the level it would have reached had it followed the path of their peers in less exposed occupations. For experienced workers there is no comparable gap.
3. The gap has been widening since it was first measured, in August 2025.
4. It comes through hiring that does not happen, more than through people leaving.
5. It concentrates where AI substitutes for human work; where it works alongside it, employment is stable or grows, especially for those with experience.
6. The adjustment happens on employment and not on base pay.

The authors state that they regard these as "early descriptive indicators - canaries in the coal mine - and not causal estimates". The data are American, and the effect softens once level of education is taken into account.

Two different measures, and they do not subtract. Earlier versions of the study headlined a regression estimate adjusted for firm-level shocks: 13% on July 2025 data, 16% on September 2025 data. This version headlines the descriptive divergence, which requires no modelling choices, and on that measure the shortfall was 15% on July 2025 data and is 19% on June 2026 data. The 16% and the 19% belong to two different measures and should not be read as a three-point worsening. In levels, between November 2022 and June 2026 employment of 22-25 year olds fell by about 11% in the two most exposed quintiles of occupations and grew by about 10% in the three least exposed.

An independent check, and it points the same way. On 1 September 2026 the Federal Reserve Bank of New York published the results of its August surveys of firms in the New York and northern New Jersey region: 61% of service firms say they used AI this year, against 40% last year, and among manufacturers 51% against 26%. Only 4% of service firms laid anybody off in response to AI over the past six months; 15% say they hired fewer people than they otherwise would have. It is a regional survey, not a national one, so it does not compare with the Census's 23.8%, but the mechanism it describes is the canaries' own: it works through hiring, not through lay-offs.

Job postings. In Italy, in 2025, about 44,000 positions out of 3.2 million postings published asked for artificial intelligence skills, 93% more than the year before. That is 1.4% of postings. The Italian figures come from the Osservatorio of the Politecnico di Milano, which analysed more than 17 million postings published in Italy between 2019 and 2025 with Lightcast. The same measure read by Stanford's AI Index from Lightcast's international archive gives 1.33%, and the two figures agree because they go back to the same source, not because they confirm one another. For comparison, again in 2025: Singapore 4.69%, United States 2.56%, United Kingdom 1.93%, Germany 1.13%, France 0.99%.

Stanford's updated indicators →

How to read an adoption percentage

When a new number arrives, four questions are enough to work out whether it says what it seems to say. They are the same ones I used to build this page.

What is the base?	Firms, workers, population, job postings. "One worker in two" and "16.4% of firms" look as though they contradict each other and they do not: they count different things.
What does the question ask?	"Has bought licences" (84%), "has a project" (71%), "uses it across several functions" (one in five) and "sees an effect on EBIT" (37%) are four different questions on the same subject.
Which year is the survey from?	Not the year of publication. The Osservatorio data presented in February 2026 were collected in 2025; the McKinsey questionnaire published in August 2026 closed on 8 June.
Who answered, and why?	A statistical sample drawn at random (Eurostat, Istat), a panel of executives who agree to fill in a questionnaire about AI (McKinsey, Osservatorio), or an archive of job postings and payroll records that asks nobody anything (Lightcast, ADP). The three do not mix in the same sentence.

What has changed

3 October 2026 · revision

The US Census has published a new round, and the share has risen for the third time running. Between 24 August and 6 September 2026, 23.8% of American firms had used AI in the previous two weeks, up from 23.2% in the round before. The last four rounds give 21.8%, 22.4%, 23.2% and 23.8%: a few tenths at a time, always in the same direction. McKinsey's "nearly nine in ten" stays a long way off, for the reasons given earlier on this page.

The other sources I reopened have not moved. The Eurostat data on firms and on individuals, recalculated on the database on 3 October 2026, match to the second decimal place; the latest edition of the World Economic Forum's Future of Jobs Report is still the one of January 2025; the results of AgID's new survey of the public administration have still not appeared.

19 September 2026 · revision

McKinsey has published the 2026 edition, and the picture has moved from use to scale. Nearly nine organisations in ten report regular use of AI, as they did the year before, but the share saying they have scaled it across the enterprise rises to 44% and the share attributing any EBIT effect to it stays at 37%. Those getting at least 5% of EBIT from it have been at about six in a hundred for a year. Between buying and earning, the whole distance this page describes is still there.

The number that was missing for the smallest Italian firms is now there, and it is 5%. European business statistics start at ten employees, and in Italy that means leaving out almost everybody. On 8 September 2026 the Bank of Italy measured that floor on 5,000 micro-enterprises: five in a hundred use AI, nineteen are trialling it or expect to adopt it within a year, and seventy-three are not familiar with the technology or have planned nothing.

Two corrections to this page, and they come from rereading the sources, not from new facts. The Stanford study does not carry a gap "from 15-16% to 19%": the 16% was a regression estimate the authors no longer headline, and on the measure they use today the shortfall runs from 15% to 19%. And OpenAI has published a definition of an active user, on 30 June 2026, inside a study on ChatGPT adoption.

14 August 2026 · first publication

Italian adoption has doubled in a year, the size gap has not. Firms using AI go from 8.2% to 16.4%, but the distance between those with more than 250 employees and those with fewer than 50 stays at almost forty points. If you work in a large firm, what you see around you is not the country.

The Stanford payroll study was revised on 12 August 2026. The gap for young people in exposed occupations rises from 15% to 19%, and the first of the six findings remains that a widespread displacement of jobs is not visible. Anyone citing that study to announce the end of entry-level work is citing half the result.

On employment forecasts there has been nothing new for nineteen months. The figures 170 / 92 / +78 million you find everywhere come from the January 2025 report and have not been revised. When a number stays still this long while the subject moves, it counts more as an order of magnitude than as a forecast.

Sources

• Eurostat, use of artificial intelligence in enterprises, 2025 data, database queried on 14 August 2026 and recalculated on 3 October 2026; data last updated on 15 June 2026; Eurostat news release of 11 December 2025.
• Eurostat, use of generative artificial intelligence by individuals, 2025 data, database queried on 14 August 2026 and recalculated on 3 October 2026; data last updated on 5 June 2026; news releases of 16 December 2025 and 10 February 2026.
• Istat, "Imprese e ICT - Anno 2025", statistical report of 15 December 2025.
• Osservatorio Artificial Intelligence of the Politecnico di Milano, release on the results of the 2025 research, presented on 5 February 2026; release of October 2025 on the research presented to the Chamber of Deputies for the estimates to 2033, and for the analysis with Lightcast of more than 17 million Italian job postings 2019-2025. Consulted on 14 August 2026 and read again on 19 September 2026.
• Banca d'Italia, "Indagini sull'alfabetizzazione finanziaria e le competenze di finanza digitale in Italia: piccole imprese", 8 September 2026, on 5,000 micro-enterprises with up to nine employees surveyed in 2025.
• Unioncamere and Dintec, release of 15 September 2026 on the analysis of more than 20,000 SELFI4.0 assessments from the Punti Impresa Digitale.
• McKinsey, "The state of AI in 2026: On the road to ROI", 25 August 2026, and the previous edition "The state of AI in 2025: Agents, innovation, and transformation", 5 November 2025.
• McKinsey Global Institute, "Agents, robots, and us: Skill partnerships in the age of AI", 25 November 2025, and the European version "How AI reshapes work and skills in Europe", 12 May 2026.
• World Economic Forum, "Future of Jobs Report 2025", 7 January 2025; series page consulted on 14 August 2026 and the Forum's publications index read again on 3 October 2026 to check that no later editions exist.
• Brynjolfsson, Chandar, Chen, "Canaries in the Coal Mine? Six Facts about the Recent Employment Effects of Artificial Intelligence", Stanford Digital Economy Lab, version revised on 12 August 2026, consulted on 14 August 2026 and read again in the PDF on 3 October 2026.
• US Census Bureau, Business Trends and Outlook Survey, cycle 202619, reference period 24 August-6 September 2026, published on 24 September 2026, read in the national data file on 3 October 2026.
• Federal Reserve Bank of New York, Liberty Street Economics, "Businesses Are Using AI to Transform Work, Not Cut Jobs", 1 September 2026, on the August surveys in the New York and northern New Jersey region.
• Stanford HAI, AI Index Report 2026, chapter 4, for the Lightcast data on job postings.
• OpenAI, release of 31 March 2026 and engineering piece of 11 September 2026, for the figures stated on users, and study on ChatGPT adoption of 30 June 2026 for the definition of an active user.
• Google, release of 11 August 2026 on monthly users of the Gemini app; Microsoft, release of 30 July 2026 on paid seats of Microsoft 365 Copilot.
• AgID, "L'Intelligenza Artificiale nella Pubblica Amministrazione", report of June 2025 on a survey carried out between 25 September and 10 October 2024; AgID news and publications archive read again on 3 October 2026.