AI promises a productivity windfall – who will reap the benefits?

When Amazon Web Services’ (AWS) acting vice-president for the worldwide public sector, David Appel, opened the firm’s UK Public Sector Symposium earlier this month, he drew a historical parallel to artificial intelligence (AI). Three miles west of the City of London venue we were at is the Crystal Palace, a million square feet of glass and iron that housed the Great Exhibition of 1851. What filled it, he said, was proof of what happens when ordinary people – the engineers whose products were showcased – are given access to the means to create and invent. The Crystal Palace, in Appel’s telling, represented “the infrastructure they couldn’t build alone” of the first industrial revolution. And that, he implied, is what AI (via AWS’s cloud infrastructure, of course) offers us today. It is a seductive parallel, but only half the story. The Great Exhibition was the victory lap of a revolution that had, for the previous four decades, paid capital before it paid labour. Between 1800 and 1840, British output per worker rose strongly while real wages flatlined, and the profit share of national income roughly doubled – a period the economic historian Robert Allen calls “Engels’ pause” . TL;DR – AI and the lessons of ‘Engels’ pause’ and ‘the general intellect’ Supplier AI utopias ignore that early industrial gains flowed entirely to capital while wages flatlined – “Engels’ pause”. Like the electric dynamo in industrial revolution 2.0, AI requires costly organisational restructuring before productivity gains materialise. By absorbing collective human knowledge – “the general intellect” – into capital-owned machines, AI shifts wealth upstream. Augment versus automate – the choice between sharpening human expertise or deskilling workers. The Turing Trap – a rush to replace humans rather than augment their strengths risks high costs and low productivity. Deliberate choices – shaping the future requires active decisions on wealth distribution, public services and augmenting human endeavour. The same machinery the Crystal Palace celebrated had enriched its owners first; the workers whose labour powered it were made to wait. In the decade before the exhibition opened, Friedrich Engels had been in Manchester documenting the consequences . The Crystal Palace and the Manchester slums were two views of the same moment. The question is which of these two histories the AI revolution will repeat. History shows that a general-purpose technology’s gains arrive late and flow to capital first, and a 200-year-old idea – “the general intellect” – explains why. Today, AI presents a fork in the road – augment or automate, genuine productivity or so-so automation, shared or concentrated wealth – and which path we take is a choice, not a destiny. Supplier claims and workplace impact There is evidence that AI is making its impact on work. Klarna has replaced around 700 full-time customer service agents with an AI assistant. BT, meanwhile, cut around 10,000 roles and attributed it to AI and digitisation. At the same time, the hyperscalers – Microsoft, Amazon, Google, Meta – funnel record sums (not far off $1tn this year) into the datacentres and chips that make all of it possible. The promise, oft-repeated by IT suppliers, is that AI will absorb drudgery, lift productivity, and free human beings for better work. The windfall, we are often told, will be shared for the good of us all. That claim appears to be doing some heavy lifting, because the historical record suggests productivity gains and their shared benefits are not automatic. There’s a productivity paradox at the heart of industrial revolutions, and that’s where the trouble begins. That was summed up, famously, in 1987, when Nobel economist Robert Solow quipped: “You can see the computer age everywhere except in the productivity statistics.” Nearly four decades later, with AI, the same observation keeps recurring – the technology is everywhere, but the measured numbers are hard to come by, or are underwhelming. Part of that is measurement – a great deal of digital value is free, and free goods weigh almost nothing in GDP. But most of it is something older and more interesting. That is, that a general-purpose technology revolution does not pay off on arrival. How long before the gains show? Economist Paul David made the point in a 1990 essay that the electric dynamo was commercially viable by the 1880s – Edison’s Pearl Street Station opened in 1882, and factory electric motors followed through the 1890s. Yet US manufacturing productivity did not surge until the 1920s. The obstacle was not the technology, it was the factory. Early adopters wired one big electric motor to existing steam-age shafts and belts. The productivity gain only materialised when factories were rebuilt around what electricity actually allowed – small motors on individual machines, single-storey factory layouts, moving assembly lines, and a different kind of worker and manager. “The gains,” as the lesson is usually summarised, “were in the reorganisation, not the motor.” Erik Brynjolfsson, director of Stanford University’s Digital Economy Lab, has given that insight a formal, modern shape . He argues that a general-purpose technology follows a “productivity J-curve” – first comes a dip, then a sharp rise. The dip is not failure. It is the cost of “intangible” investments – new workflows, retrained staff, re-engineered business processes – that have to accumulate before the payoff shows. “That investment in intangible assets is huge,” Brynjolfsson says. “It can be much bigger, up to 10 times bigger, than the direct investment in the technology, the computers and the software. But it mostly isn’t measured.” His estimation of where we are now is the bottom of the curve. “Last year, 2025, productivity was about 2.2%, which was significantly higher, almost double what it had been averaging the previous 10 years,” he noted. “I anticipate that will grow.” So the first question – how long before the gains show up? – has a reasonably settled answer. They show up, but only after a costly, unglamorous reorganisation that almost nobody is measuring. The second question is harder, and it is the one the suppliers don’t seem to examine so deeply. Who gets the gains when they do come? Here, the relevant precedent is not electricity but steam, and the relevant period is not the 1920s but the early 19th century. Robert Allen’s “Engels’ pause” describes the period between roughly 1800 and 1840 when British output per worker rose strongly, but real wages flatlined. Over the same decades, the profit share of national income roughly doubled. Capital was paid first. Labour waited. Allen’s finding matters because it dismantles the assumption underneath most AI optimism – that productivity growth and wage growth move together. For the first four decades of industrial revolution 1.0, they didn’t. Productivity arrived, but higher wages did not. The reason was described, with remarkable prescience, before the pause had fully played out. In 1824, the Irish socialist William Thompson wrote that knowledge, once alienated into machines, becomes “a power inimical to workers”. Later, Karl Marx took up Thompson’s idea of the “general intellect” – the accumulated social knowledge, science and technique that gets “congealed” into machinery. We can now read to include AI, or more precisely large language models (LLMs), that have literally absorbed vast amounts of human knowledge. Because as an increasing amount of knowledge – the general intellect – is built into machines, more and more of the economy’s productive power comes from what Marx called “dead labour”. That equates to past labour stored in the machines rather than the “living labour” of the worker. And because they allow greater productivity, capital substitutes machines for people. This means fewer workers are required to produce more, and more of the output flows to the owners of the machines. “Already in the early 19th century, Thompson provided one of the first theories and political programmes regarding mental labour,” says Matteo Pasquinelli, associate professor in philosophy of science at Ca’ Foscari University in Venice. “The general intellect he was referring to, however, was not science in the abstract, but the knowledge produced by workers themselves.” Why it matters now is that it shows how productivity gains do not automatically reflect in higher wages. The practical translation is that when you automate a task and the skill moves into a machine owned by the firm, the money it earns follows the owner of the machine, not the worker. The trapping of collective knowledge into machines means the dividends of automation flow upstream to the capital-owning few. The consequence is that any AI-driven productivity surge may replicate Engels’ pause rather than break it, and leave living labour to fight over the scraps of a stagnant wage share. The ‘general intellect’ and AI Origin: Coined by William Thompson, an Irish Ricardian and utopian socialist, in An inquiry into the principles of the distribution of wealth (1824). His argument was that knowledge, once alienated into machines, becomes “a power inimical to workers”. Later use: Adopted and transformed by Karl Marx in the Grundrisse (1858), in the passage now called the “ Fragment on Machines ”. For Marx, the general intellect is accumulated social knowledge “congealed” into machinery, which he calls “dead labour” – as against the “living labour” of the worker. Co-opting: The Italian autonomist tradition later read the idea optimistically, as shared wealth that could one day free labour from the wage relation. Why it matters now: It names the mechanism by which productivity detaches from wages. The lineage from Thompson to Marx is reconstructed in Matteo Pasquinelli’s The eye of the master (2023): “Marx added that collective knowledge was also embodied across the mechanisms of industrial machinery. Both forms of knowledge, whether collective and mechanised, are of course forms of collective wealth.” The modern echo Here is where the 19th century stops being history. In 2013, economists Loukas Karabarbounis and Brent Neiman documented what they called “the global decline of the labor share” . Using a dataset that covered 59 countries, they found that the share of corporate income paid to labour had fallen by roughly 5% since the early 1980s, and that the decline was not confined to a few countries or industries. Their explanation was strikingly close to that of Thompson and Marx. “The decrease in the relative price of investment goods, often attributed to advances in information technology and the computer age, induced firms to shift away from labor and toward capital,” they wrote. “The lower price of investment goods explains roughly half of the observed decline in the labor share.” In other words, cheaper computers and automation made it rational to substitute machines for people, and the income shifted accordingly – a greater share of production drifting towards “dead labour”, the “general intellect”. Brynjolfsson, whose optimism about productivity is genuine, has noticed the same thing. “Less of the GDP is going to wages and salaries, and more is going to capital,” he says. “For the past 200 years, that ratio has been almost constant. Recently, it looks like the labour ratio is beginning to fall.” His concern is not only income, but power. “I worry that the technology is disempowering a lot of workers, a lot of citizens, because most of us get most of our income from labor income, not from capital income.” The forks in the road If the rise in proportion of “dead labour” drives down the labour share of income, accepting that outcome is a policy choice, not a law of physics. This brings us to the fundamental forks in the road governing how AI is deployed. Automation … affects you not just because it affects the quantity of things you do, but also how expert they are. Eliminating work might make you more expert because it removes supporting tasks [or] less expert because it eliminates your specialised tasks David Autor, MIT The most important of these – as we’ve alluded to already – is who benefits? That resolves in large part to a question of to whom the benefits flow. But there is a related question about the likely effect of AI on work. There are forks in the road here, and a key one is whether AI will augment or automate and its effect on the workforce and society. For some, that’s also key to realising the productivity gains of AI. More expertise or less? MIT labour and employment professor David Autor frames the choice as being about expertise. Automation, narrowly defined, just means replacement. The real question, he argues, is whether new technologies such as AI “augment or eliminate” work and expertise. “Automation … affects you not just because it affects the quantity of things you do, but also how expert they are,” he says. “Eliminating work might make you more expert because it removes supporting tasks. It might make you less expert because it eliminates your specialised tasks.” He concludes that some jobs become more expert because inexpert tasks are removed or expert tasks are added, while some become less expert because expert tasks are removed or inexpert tasks are enabled. We can see this in software engineering, where senior roles can become more expert when coding agents and LLMs routinely handle boilerplate work like writing syntax, spinning up basic functions, and so on. For those roles, the job shifts upward into spec-driven development, system design, security auditing, agent orchestration, and so on. Meanwhile, coding has been “de-skilled” for the entry-level tier. A novice or even a non-programmer can use an LLM to spin up a functional web app or patch a basic bug in seconds. Automation destroys the scarcity of skills, says Autor – “taking something that was a valuable skill that people invested in to develop and all of a sudden makes it so easy and cheap that that skill is no longer valuable”. Autor’s conclusion is the fork in miniature: “We have a lot of choices about how we use these technologies. We can use them primarily to automate … or we can use them to enable new possibilities in science, in education, in health.” So-so automation The problem with so-so automation is you get the bad – displacement of workers – and you don’t get the good, which is productivity improvements or cost savings Daron Acemoglu, MIT Meanwhile, Daron Acemoglu, also of MIT , warns against what he calls “so-so automation … by which we mean automation that displaces workers from the tasks they used to perform but doesn’t actually reduce costs all that much”. His example is early automated customer service, which frustrated customers and failed to help firms. “The problem with so-so automation is you get the bad – displacement of workers – and you don’t get the good, which is productivity improvements or cost savings.” Much of today’s AI adoption, Acemoglu argues, is done in a rush – companies automate tasks before systems are genuinely ready, because the hype tells them they must. “If you rush and do with AI things that you could have done with workers before the AI systems are ready, you’re going to get a lot of so-so automation… That’s the recipe for getting a lot of the bad and none of the good.” Avoiding ‘the Turing Trap’ But to take the fork in the road signposted “augment” is real and achievable, says Stanford’s Brynjolfsson, who cites evidence from a study of call centres . His team found productivity gains of up to 30% within months when AI was used, and the beneficiaries were disproportionately the least-skilled workers, lifted towards the level of the best. That is augmentation working as intended. He contrasts that with what he calls “the Turing Trap” – the compulsion to design AI to imitate and replace humans rather than to complement them. “Way too many of the technologists … are very focused on using the technology to just replace humans,” he says. The remedy is to build machines that do what machines are good at, and leave humans to do what humans are good at: “We should work on having machines do things that are easy for machines and hard for humans… That would make them more of a complement.” The distribution fork: Who benefits? Even if individual firms successfully choose the augmentation path over the Turing Trap, it leaves a larger macroeconomic question unanswered: how are the aggregate gains distributed across society as a whole? Acemoglu and his long-time collaborator Simon Johnson, in Power and progress: Our thousand-year struggle over technology and prosperity , make the case that technology’s direction is not preordained. The directions AI will take – augmenting or replacing, shared or concentrated wealth – are a choice presented to firms, governments and workers. Pasquinelli reaches a similar place from the opposite direction. “By automating the collective sphere of knowledge, AI confirms that the collectivity is the source of value, and not just individual business,” he says. “This should compel everyone to consider the issue of its nationalisation or collectivisation. AI can only transform itself into a public utility.” The machinery that was built out of everyone’s knowledge cannot, in the end, belong only to a few, says Pasquinelli. In other words, if AI contains the myriad fruits of human knowledge, and meanwhile makes society vastly more productive and with a radically transformed relation of humanity to work, then all should benefit. That train of argument often manifests in a call for universal basic income (UBI), as argued for by “godfather of AI” Geoffrey Hinton of the University of Toronto and former OpenAI head of policy research Miles Brundage, among others. Aaron Benanav, labour historian and author of Automation and the future of work , argues that AI involves “real technological breakthroughs” that will probably have “substantial effects on productivity”. On jobs, he points out that “where these tools are being adopted, employment is growing rather than shrinking”. Automation removes some tasks, but creates bottlenecks that demand new workers elsewhere. On UBI, he is welcoming but wary. His book argues that a universal basic income, on its own, risks becoming a sticking plaster – a cash transfer that subsidises precarious work while leaving untouched the question of who owns the gains. In conversation, he makes the constructive version of the same point: “Universal basic income would certainly be something good, but so would a build-out of universal services” – higher taxes, a rebuilt NHS, the green transition. As for positive steps, one panellist at the AWS event – Laura Gilbert, formerly the UK government’s director of data science and now at the Tony Blair Institute – argued that AI’s biggest returns can be social and public, not commercial, and that improvements in the state and public sector can lift society as a whole. We have incredible agency, but we’re squandering that because we don’t understand well enough what our choices are Erik Brynjolfsson, Digital Economy Lab, Stanford University “The UK is the third most unequal country in the OECD developed nations,” she said. “And that costs us somewhere between £35bn and £70bn a year. Reducing inequality to the mid-level is predicted to reduce people being imprisoned by about 30%, mental health problems by about 5%, and murders by 33%. “A 10% efficiency gain in a benefits system, courts, or the NHS affects the whole population’s baseline quality of life,” she said. Conclusion For the C-suite reader, the practical lesson is that, historically, the benefits of productivity gains have tended to concentrate in few hands. If we want AI to bring social good, that has to be engineered deliberately – through procurement that favours augmentation over headcount reduction, through reskilling budgets that match the technology spend, through job design that protects what humans are good at, and through a willingness to argue for the institutions that can help the benefits flow to society more widely. Brynjolfsson puts it plainly: “We have incredible agency, but we’re squandering that because we don’t understand well enough what our choices are.” Read more about AI and society Interview with Paris Marx on why hyperscale datacentres pose greater threat than AI ‘science fiction’ : Ahead of the release of his book, Paris Marx tells Computer Weekly how tech companies’ insatiable datacentre demands are accelerating the climate crisis. AI hasn’t changed ‘core skills’ needed for work, say experts : AI is rapidly changing the world of work and how we navigate daily life, but the core skills schools need to instil in students haven’t changed.