The Machine and the Hour

Essay · economics · October 2026

Keynes expected technology to bring a fifteen-hour week by 2030. Artificial intelligence has revived the hope, and the fear. The first careful studies find that AI makes some tasks much faster, but that workers put the time saved back into other work: earnings and hours have not moved. In Portugal, people use the new tools more than the European average, but firms adopt them far less. Whether AI shortens the working week will depend less on the technology than on who decides what to do with the time it saves.

Reading time ≈ 10 min The record · exposure · adoption · productivity · the hours · the choices

The question

In 1930 Keynes predicted that, within a century, technology would make people so productive that a working week of fifteen hours would suffice. The Fifteen-Hour Week showed that productivity grew as he expected but hours did not fall as far, and The Fifth Day looked at the four-day week. Artificial intelligence is now presented both as the technology that will finally free our time and as one that will take our jobs. The Accelerant looked at what AI does to society; this essay looks at what it does to paid work: which jobs it touches, how fast it is spreading in Portugal, what it has done so far to productivity, pay and hours, and what could make it shorten the week.

9–30%of Portuguese jobs exposed to automation or AI, depending on the definition

20% vs 12%Portuguese adults using generative AI for work, against firms using AI, 2025

2.8%of work hours saved by Danish chatbot users; under 10% of them took the time off

0.07–1.5percentage points a year: the range of forecasts of AI’s effect on productivity

The record

Machines took tasks, not work

Fears of technological unemployment are old, and so far they have been wrong about work as a whole. In 1900, 41 per cent of Americans worked in agriculture; in 2000, 2 per cent did, and the others found other work.[1] About 60 per cent of American jobs in 2018 had titles that did not exist in 1940.[2] In textiles, steel and cars, automation first raised employment, because cheaper products sold far more, and cut it only once demand stopped growing.[3]

Particular workers did lose. In the United States, each additional robot per thousand workers lowered the share of people in work by 0.2 points and wages by 0.4 per cent between 1990 and 2007, concentrated in the places where the robots went.[4] In Germany, the jobs lost to robots in manufacturing were fully replaced by new jobs in services, and across fourteen European countries, Portugal included, more robots went with more employment, not less.[5] Since 1980, though, the jobs destroyed by automation have weighed more and the new work it creates less than in earlier decades.[2]

Hours did fall: annual hours per worker in France fell from about 3,200 in 1870 to about 1,400 in 2000, and similarly elsewhere. But no study attributes the fall to automation as such; it came from laws, unions and rising incomes. Most of the gain from productivity was taken as more consumption: an American in 2015 could earn the average income of 1915 by working about 17 weeks a year, and almost no one does.[1][6]

Exposure

How many jobs AI touches depends on who counts

Every estimate of how many jobs AI will affect measures something different (Figure 1). The OECD finds that 8.7 per cent of Portuguese jobs are at high risk of automation if the test is that a quarter of the skills used can be automated, and 30 per cent if the test is being in the quarter of occupations most at risk. A study for the Fundação Francisco Manuel dos Santos classifies 29 per cent of private-sector jobs as “collapsing professions”; the Fundação José Neves puts 18 per cent of employees in the jobs most exposed to AI. None of these is a forecast of jobs lost; “exposed” means that AI can do some of the tasks, which may make the worker more productive rather than redundant.[7][8][9]

Fig. 1 — Portugal: share of jobs exposed to automation or AI, by five definitions, per cent. The definitions, years and populations differ; the numbers are not comparable and cannot be added. Data: OECD, Employment Outlook 2023, Fig. 3.5, and Job Creation and Local Economic Development 2024, Portugal note; FFMS Policy Paper 3 (2025); Fundação José Neves, Estado da Nação 2024.

What is new is who is exposed. Earlier automation threatened routine work in factories and offices; generative AI reaches the work of graduates. In Portugal, the jobs at high risk from the older kind of automation are concentrated in the North and Centre, where manufacturing is; the jobs most exposed to generative AI are concentrated in Lisbon, with its offices and services (Figure 2).[8] Among Portuguese employees, 56 per cent of those with a master’s degree or doctorate are in highly exposed jobs, against under 4 per cent of those with basic schooling, and women (21 per cent) more than men (16 per cent).[9]

Fig. 2 — Portugal by region: jobs at high risk of automation (at least a quarter of skills automatable) and jobs exposed to generative AI (at least a fifth of tasks can be done twice as fast), 2022, per cent of employment. Data: OECD, Job Creation and Local Economic Development 2024, Portugal country note, Figs. 13–14.

Adoption

Portuguese people use it; Portuguese firms do not

In 2025, 39 per cent of Portuguese aged 16 to 74 had used a generative AI tool in the previous three months, and 20 per cent had used one for work, both above the EU averages of 33 and 15 per cent. Among graduates aged 25 to 64, 47 per cent used it for work, against 34 per cent in the EU.[10] Firms are another matter. Only 11.5 per cent of Portuguese firms with ten or more employees used any AI technology in 2025, against 20 per cent in the EU, 21st of the 27 countries (Figure 3). Portugal kept pace until 2023 and fell behind as other countries accelerated. The gap is in small firms: 9 per cent of those with 10 to 49 employees use AI, against 17 per cent in the EU; among firms with 250 or more, the difference is small.[11]

Fig. 3 — Share of firms with ten or more employees using at least one AI technology, and share of adults aged 16–74 using generative AI for work, EU countries and Norway, 2025, per cent. A European survey of 2024 found far lower personal use in Portugal (4 per cent of workers); it asked a different question a year earlier. Data: Eurostat (isoc_eb_ai, isoc_ai_iaiu).

This matters because the gains from a new technology come mainly when firms reorganise work around it, not when individuals use it on their own. A country of small firms with few managers trained to reorganise them, as What Schooling Bought described, risks having workers who use AI to do the same job a little faster, and firms that do not change how they work.

Productivity

Large gains on tasks, small ones in the economy

The controlled studies show large gains on particular tasks. Customer-support agents with an AI assistant resolved about 15 per cent more problems per hour, and the least experienced about 30 per cent more. Professionals writing short documents took 40 per cent less time and produced better work. Consultants using AI on tasks within its capabilities worked a quarter faster and better, but on a task beyond them they were 19 points less likely to get the answer right than those without it. In all three, the least skilled gained most.[12]

Forecasts for the economy as a whole range from almost nothing to a lot (Figure 4). Daron Acemoglu estimates that AI will raise total factor productivity by at most 0.66 per cent over ten years, less than 0.1 points a year; the OECD, 0.25 to 0.6 points a year in the United States, about half that in France and Italy; Goldman Sachs, about 1.5 points a year for labour productivity.[13] The low estimates assume AI will do a modest share of tasks only somewhat cheaper; the high ones assume it will spread fast and reorganise whole jobs. There is no estimate for Portugal. For comparison, Portuguese output per hour grew by 0.17 per cent a year from 2013 to 2025.

Fig. 4 — Forecasts of the extra annual productivity growth from AI, percentage points; bars are ranges, dots central estimates. Acemoglu’s estimate is 0.66 per cent over ten years. Total factor and labour productivity are different measures; the forecasts cover different countries and periods. The dotted line is Portugal’s measured growth of output per hour, 2013–2025. Data: Acemoglu (2024); Filippucci, Gal and Schief, OECD (2024); McKinsey Global Institute (2023); Goldman Sachs (2023); Eurostat.

The hours

The time saved goes back into work

The most careful evidence so far on pay and hours comes from Denmark, which links surveys of chatbot use to official records of earnings and hours. In eleven occupations exposed to AI, workers who used chatbots in 2023 and 2024 saved on average 2.8 per cent of their working time. That time did not become free time: 80 per cent said they used it for other work tasks, a quarter spent it doing the same tasks more carefully, and fewer than one in ten took breaks or leisure (Figure 5). The effect on recorded earnings and hours was zero, precisely enough to rule out changes larger than 1 to 3 per cent, and only 3 to 7 per cent of the productivity gains reached pay.[14] In the United States, the time saved amounts to about 1.4 per cent of all work hours.[15]

Fig. 5 — Denmark: what workers in eleven AI-exposed occupations who saved time with chatbots did with it, 2023–2024, per cent of those who saved time; answers can overlap. Data: A. Humlum and E. Vestergaard, “Large Language Models, Small Labor Market Effects”, BFI Working Paper 2025-56, and NBER Working Paper 33777 (revised March 2026).

This is what earlier technologies did too. A worker cannot, on their own, turn a faster task into a shorter day: the hours are in the contract, and the employer decides what fills them. The evidence is early, and the effects are small because adoption is still recent. But nothing in it suggests that AI will shorten working hours by itself.

Who gains

Narrower gaps at work, wider ones in income

Two effects pull in opposite directions. Within jobs, AI narrows differences: the least experienced gain most, as the studies above show. Across the economy, the experience of earlier automation points the other way. Replacing workers in particular tasks accounts for between half and seven-tenths of the change in the American wage structure since 1980, and added little to productivity.[16] The IMF finds that in its scenarios inequality of capital income and wealth “always increase” with AI, because owners of capital capture part of the gain; whether workers’ incomes also rise depends on whether AI complements them or replaces them.[17] Who Gets the Profits showed how concentrated the ownership of Portuguese capital already is.

The choices

Taxing machines, shortening weeks, training people

Three kinds of policy are proposed. The first is to tax automation. Models that work out the best robot tax find it small and temporary: about 5 per cent falling to under 1 per cent within two decades in one, 1 to 4 per cent in another.[18] A related argument is that tax systems already favour machines: in the United States, labour was taxed at an effective 25 to 33 per cent in the 2010s, equipment and software at about 10 per cent and less after 2017; equalising the two would raise employment.[19] Critics reply that a robot is hard to define and that taxing productivity slows growth; the European Parliament discussed a robot tax in 2017 and left it out.[20]

The second is to turn productivity gains into shorter hours by law or agreement. An American bill of 2024 would have cut the standard week to 32 hours, presented as a way of sharing the gains of AI and automation; it was not passed.[21] Portugal regulates how algorithms are used to manage workers, and in January 2026 adopted a national AI agenda with 32 initiatives, mainly on training and public services; neither includes any measure on working time.[22]

The third is training. Since February 2025, the European AI Act has required firms that use AI to ensure their staff know enough about it, though the Commission has since proposed simplifying its rules.[23] Training programmes help, but slowly: across more than 200 studies, their effects on employment are close to zero in the first year and larger two or three years later.[24] In Portugal, where 42 per cent of adults read at the lowest level, the skill of checking what an AI produces is not evenly spread.

The balance

A technology that will not decide the hours by itself

The evidence so far supports neither the fear nor the hope. Past automation did not end work, and the early evidence on AI shows no loss of jobs or pay; but it also shows the time saved being absorbed by more work, as it was before. Portugal’s particular risk is not that AI will spread too fast but too slowly through its firms, so that it adds little to the low productivity growth that holds Portuguese wages down.

Keynes was right that technology would make shorter hours possible; it has, several times over. Whether AI brings the fifteen-hour week closer depends on the same things that decided it in the past: whether the gains are large, who captures them, and whether workers, through law or bargaining, choose to take part of them as time. On that, AI changes nothing by itself.

On method and tools

This piece was written collaboratively with Claude Opus 5.5 (Anthropic): human specification, editorial direction and critical review; machine data research, analysis and drafting. The figures are computed by scripts/automation.py. Exposure measures are reported with their own definitions and are not combined. Adoption data are Eurostat’s 2025 surveys of firms with ten or more employees and of individuals aged 16 to 74. Several of the studies cited are working papers; where the published version was not read, the working paper is given. The productivity forecasts are reported as their authors state them; Acemoglu’s ten-year figure is divided by ten for the chart. Results are in docs/automation-results.json; the downloaded sources, with page or table for each number, are kept with the script’s data.

The cover photograph is BMW Leipzig, Karosseriebau by BMW Werk Leipzig; CC BY-SA 2.0 DE, via Wikimedia Commons, cropped.

Authored by: Luis Matos Ferreira — Physicist, Developer, Writer

Related essays on this blog

1. The Fifteen-Hour Week — Keynes’s prediction and what happened.
2. The Fifth Day — the four-day week after the trials.
3. The Accelerant — social media and AI as accelerants of social change.
4. The Stalled Hour — why Portuguese output per hour stopped growing.
5. Work, Time and Money — the reading guide to the whole series.

Sources

1. D. Autor, “Why Are There Still So Many Jobs? The History and Future of Workplace Automation”, Journal of Economic Perspectives 29(3), 2015, pp. 5–8.
2. D. Autor, C. Chin, A. Salomons and B. Seegmiller, “New Frontiers: The Origins and Content of New Work, 1940–2018”, NBER Working Paper 30389, 2022, pp. 12, 43 and 46.
3. J. Bessen, “AI and Jobs: The Role of Demand”, NBER Working Paper 24235, 2018, pp. 2 and 13.
4. D. Acemoglu and P. Restrepo, “Robots and Jobs: Evidence from US Labor Markets”, Journal of Political Economy 128(6), 2020, p. 2188.
5. W. Dauth, S. Findeisen, J. Suedekum and N. Woessner, “The Adjustment of Labor Markets to Robots”, Journal of the European Economic Association 19(6), 2021, p. 3104; D. Klenert, E. Fernández-Macías and J.-I. Antón, “Do robots really destroy jobs?”, Economic and Industrial Democracy 44(1), 2023.
6. M. Huberman and C. Minns, “The times they are not changin’”, Explorations in Economic History 44, 2007, Table 3.
7. OECD, Employment Outlook 2023, chapter 3, Fig. 3.5 (StatLink xkr98z).
8. OECD, Job Creation and Local Economic Development 2024, Portugal country note, pp. 13–16.
9. FFMS, Automação e Inteligência Artificial no Mercado de Trabalho Português, Policy Paper 3, 2025, pp. 16–17; Fundação José Neves, Estado da Nação 2024, pp. 55–60.
10. Eurostat, use of generative AI by individuals (isoc_ai_iaiu), 2025.
11. Eurostat, artificial intelligence by size class of enterprise (isoc_eb_ai), 2021–2025; Eurofound, European Working Conditions Survey 2024: first findings, 2025, Fig. 26.
12. E. Brynjolfsson, D. Li and L. Raymond, “Generative AI at Work”, arXiv 2304.11771 v2, 2024, pp. 2 and 16; S. Noy and W. Zhang, Science 381, 2023, pp. 187–192; F. Dell’Acqua et al., “Navigating the Jagged Technological Frontier”, HBS Working Paper 24-013, 2023, pp. 4 and 18.
13. D. Acemoglu, “The Simple Macroeconomics of AI”, NBER Working Paper 32487, 2024, p. 2; F. Filippucci, P. Gal and M. Schief, “Miracle or Myth?”, OECD Artificial Intelligence Paper 29, 2024, pp. 3 and 39; McKinsey Global Institute, The economic potential of generative AI, 2023; J. Hatzius et al., Goldman Sachs, 26 March 2023, p. 1.
14. A. Humlum and E. Vestergaard, “Large Language Models, Small Labor Market Effects”, BFI Working Paper 2025-56, pp. 2–18; NBER Working Paper 33777, revised March 2026, pp. 2 and 16.
15. A. Bick, A. Blandin and D. Deming, “The Rapid Adoption of Generative AI”, NBER Working Paper 32966, revised February 2025, pp. 2–6.
16. D. Acemoglu and P. Restrepo, “Tasks, Automation, and the Rise in US Wage Inequality”, NBER Working Paper 28920, 2021, pp. 1–6.
17. M. Cazzaniga et al., “Gen-AI: Artificial Intelligence and the Future of Work”, IMF Staff Discussion Note 2024/001, pp. 17–19.
18. J. Guerreiro, S. Rebelo and P. Teles, “Should Robots Be Taxed?”, Review of Economic Studies 89(1), 2022 (NBER WP 23806, p. 4); A. Costinot and I. Werning, “Robots, Trade, and Luddism”, Review of Economic Studies 90(5), 2023 (NBER WP 25103, pp. 4 and 20).
19. D. Acemoglu, A. Manera and P. Restrepo, “Does the US Tax Code Favor Automation?”, Brookings Papers on Economic Activity, Spring 2020, p. 234.
20. European Parliament, resolution of 16 February 2017 on civil law rules on robotics, P8_TA(2017)0051, para. 44.
21. US Senate, Thirty-Two Hour Workweek Act, S.3947, 118th Congress, 14 March 2024.
22. Lei 13/2023 (Agenda do Trabalho Digno), amending the Labour Code, arts. 3, 24 and 106; Resolução do Conselho de Ministros 2/2026, Diário da República 1st series no. 5, 8 January 2026.
23. Regulation (EU) 2024/1689 (AI Act), art. 4 and art. 113(a).
24. D. Card, J. Kluve and A. Weber, “What Works? A Meta Analysis of Recent Active Labor Market Program Evaluations”, Journal of the European Economic Association 16(3), 2018 (NBER WP 21431, pp. 2–4).