This is a side essay from Lady Lovelace’s Objection*, a serialised history of computing I am writing here. New chapters and essays land on Tuesdays.*

In a café in Surry Hills, Sydney, on a Tuesday morning in April 2026, a barista in her late twenties is making the seventeenth flat white of the morning. She has been on shift since six. The customer is a regular. She started his order before he was through the door. The flat white costs five dollars and fifty cents. The bean is a single-origin Ethiopian roasted three days ago by a small operation in Marrickville. The cup is ceramic. She has written, in black marker on a small wooden token clipped to the saucer, the word cheers.

This transaction is not what an American economist would have predicted. By the standard story of automation, the espresso machine in front of the barista should have replaced her years ago. There are espresso robots that can produce a technically competent shot of coffee with no human intervention, end to end, for under a dollar in marginal cost. Australia has been told for two decades that they are coming. They have not arrived. They are not, in any visible sense, on the way.

Starbucks tried something close to the automated, scaled version of this transaction. The company opened eighty-seven stores in Australia between 2000 and 2008, betting that the local coffee market would behave the way the American one had: customers buying the same drink, in the same paper cup, from the same script, for a price slightly under the local independent. By 2008 sixty-one of those stores had closed. The losses, by Starbucks’s own reporting, ran to hundreds of millions of dollars. The chain reopened a smaller and quieter version of itself through the 2010s and into the 2020s, with the brand markedly less visible than it had been on day one. The flat white, which Starbucks added to its global menu in 2010, did not save them.

The thing that defeated Starbucks in Australia was not, in any easy sense, technology. The technology was Starbucks’s strength. It was the relationship. The barista in Surry Hills has fifteen regulars by ten in the morning. She knows their orders, their children’s names, their bad weeks. The five dollars and fifty cents is partly for the coffee. It is also, more substantially than most observers from outside the country quite register, for everything else.

This essay is about an economic argument advanced earlier this month by a behavioural economist at the University of Chicago Booth School of Business named Alex Imas. Imas published, on 14 April 2026, a Substack post called “What will be scarce?” The piece was covered briefly in Bloomberg and Fortune, and at length on a podcast called The AI Daily Brief, where its principal arguments went, by my listening, considerably over the head of a general audience. This is a translation. It is also, because the original is American and uses American examples, an exercise in seeing what the argument looks like through Australian eyes.

The historical analogy Imas uses to anchor his argument is the slow disappearance of agriculture from the American workforce. In 1900, roughly forty per cent of working Americans worked on farms. By 2026 the figure is under two per cent. The Australian numbers are not identical but tell the same shape of story: in 1901, the year of Federation, about a quarter of working Australians were employed in agriculture, forestry or fishing. By the early 2020s the figure was closer to two per cent.

In neither country did the disappearance of farming work cause the economy to collapse. Food got cheaper, and people did not run out of things to spend money on. They spent the money on different things. They spent it, first, on manufactured goods, which had become available at scales and prices their grandparents would have found impossible. Then, as manufacturing in turn became cheap and largely automated, they spent it on services. By 2026, the typical Australian household spends a small and shrinking share of its budget on food, a moderate share on manufactured goods, and the largest share on services of various kinds, from dining out to physiotherapy to the occasional architect.

This is the historical pattern Imas invokes for the AI transition. He does not predict that AI will leave the economy unchanged. He predicts that the economy will reallocate, the way it has reallocated twice before, toward sectors where the human element is the point. The first move was off the farm. The second was out of the factory. The third, he argues, will be toward what he calls the relational sector.

The technical name for this pattern is nonhomothetic demand, which is one of those phrases that an economist might use in a paper and an editor might cut from a magazine piece. The plain-English version is straightforward enough. As people get richer, they do not just buy more of the same things. They buy different things.

A household in inner-west Sydney earning a hundred thousand dollars a year spends most of its money on rent, utilities, transport and groceries. The grocery share alone might be twelve to fifteen per cent of the budget. The same household earning three hundred thousand dollars a year does not spend three or four times as much on groceries. It spends roughly the same on groceries in absolute terms and a much smaller share in relative terms, and the difference goes to a different category entirely. It goes to school fees, to a renovation, to a wedding, to a personal trainer, to a wine subscription, to a financial planner, to a dog walker, to a fortnightly cleaner, to dinners at restaurants whose chefs have names. The category that absorbs the increase, on the average, is services delivered by humans.

This pattern has been mapped, in the academic literature, in considerable detail. A paper by Diego Comin, Danial Lashkari and Marti Mestieri in 2021 estimated that more than three quarters of the long-run shift in employment between sectors is driven by income effects of this kind, and only about a quarter by changes in relative prices. As people get richer, in other words, they redirect their spending toward different categories at a much greater rate than the price of any individual category would predict.

The word Imas uses for the thing that the relational sector trades in is provenance. Provenance, in its older sense, is the verifiable history of who made an object and where it came from. Art dealers care about provenance. Wine collectors care about provenance. The owner of a vintage car cares about provenance. In each case the object’s value rests not only on its physical properties but on the chain of human hands that brought it to the present.

Imas’s argument is that provenance, in this older sense, is becoming an organising principle of much of the modern service economy. The barista in Surry Hills has provenance. So does the chef whose name is on the door at Sixpenny in Stanmore. So does the personal trainer who knows the history of your old shoulder injury. So does the architect who designed your renovation. So does the obstetrician who delivered your daughter. In each case, the customer is not simply paying for the service. The customer is paying for the service to have been performed by this particular person, with all the small adjustments to the customer’s particular case that this particular person could provide and a generic provider could not.

The same coffee, made by an espresso robot that arrives, identical, in cafés across the country, is not the same coffee. The same diagnosis, returned by an AI screening tool with a five-page output, is not the same diagnosis. The same wedding speech, drafted by a large language model from a list of details, is not the same wedding speech. Even when the technical output is identical, or arguably better, the value of the human-delivered version is, for a substantial population of customers, higher.

Imas pushes the argument one step further by drawing on the work of the late French theorist René Girard. Girard’s idea, advanced in a series of books from the 1960s onward, was that human desire is largely mimetic: we want things, in considerable part, because other people want them. The status of a thing depends on who else has it, and on who has been excluded from having it. A bag is a bag. A Hermès Birkin bag is a bag with a waiting list of years and a delivery decision that the company makes case by case. The waiting list and the exclusivity are not bugs in the Birkin. They are most of what the Birkin sells.

Imas, with his colleague Graelin Mandel, ran experiments on this. Subjects were asked to value pieces of artwork. The same piece, marked as exclusive (only ten existed, and the buyer would receive one), drew bids about forty-four per cent higher than the inclusive version (anyone who wanted one could have one). The mark-up for human-made art, however, was substantially larger than the mark-up for AI-generated art. AI art gained about twenty-one per cent in value when made exclusive; human art gained forty-four. The exclusivity premium, in the world of mass-produced AI output, was real but smaller than the premium that attached to human authorship. The implication is unsettling for a great many businesses currently building on top of generative AI. The work of a generic AI assistant is, by construction, available to everyone. By Girard’s logic, it is therefore worth markedly less than the work of the rare human who has been brought in specially.

Place the Imas argument in Australia, and several things look more obvious here than they do from Chicago.

The trades, for a start. A senior plumber in Sydney charges about two hundred dollars an hour. A senior electrician charges similar. The work is irreducible to AI. ChatGPT cannot fix a hot water system, locate a roof leak, or wire a new circuit. The shortage of skilled tradespeople has been a regular feature of the Australian press for at least a decade, and the wages reflect it. Imas’s argument predicts that this shortage gets worse, not better, as AI commoditises white-collar work. The status of the tradie, in Australian conversation, is already shifting in this direction. A degree from a sandstone university is no longer the obvious step up that it was for a generation that came of age in the 1990s.

The restaurant scene is another. Sydney and Melbourne have dining cultures that grew rapidly through the 2010s and survived the pandemic. The chef whose name is on the door, the wine list curated by a sommelier whose face you recognise, the maître d’ who remembers that you came in for an anniversary three years ago: these are the relational sector running at full power. None of it is the cheapest possible way to feed a person. All of it is what people who can afford to spend two hundred dollars on a Friday evening choose to do with that money.

The wine industry, particularly at the boutique end, sells provenance directly. A bottle of Margaret River cabernet is not chemically distinguishable, in a blind taste test, from a hundred other Australian cabernets at similar price points. What you are buying is the vineyard, the family, the soil, the year, and the label. Imas’s argument explains why a generic AI wine would not, even if the chemistry were identical, command similar prices.

There is, however, a part of the Australian economy that has already run a version of the experiment Imas is predicting at national scale, and the results are more complicated than his framework allows.

In 2008, Rio Tinto deployed the first commercially operational autonomous haulage trucks in the iron ore mines of the Pilbara, in the north-west of Western Australia. By the early 2020s, BHP and Fortescue had followed. The autonomous trucks, the autonomous drills, and the autonomous trains that ran the iron ore from the mines to the port at Cape Lambert, between them eliminated the great majority of the human jobs that had previously been required to mine and ship one of the world’s largest commodity exports. Iron ore production became, in the way Imas predicts production in general will become, very cheap.

What did not happen in the Pilbara was the reallocation Imas’s framework predicts. The relational sector did not flourish in Newman, Tom Price or Karratha. The towns that had supported the mining workforce in earlier decades did not, after automation, fill with personal trainers, bespoke restaurants, and wedding planners. The remaining mining workforce moved to fly-in fly-out rosters, twelve days on out of Perth and going home for their relational lives a thousand kilometres away. The towns themselves did not collapse, exactly, but they did not transform either. They supported a smaller skeleton population of permanent residents, mostly working in town services that had existed before automation, and a larger transient population that was not putting its discretionary spending into the local economy.

The lesson from the Pilbara is not that Imas is wrong. The reallocation he predicts did happen, but it happened in Perth and Sydney, not in the place where the productivity gain occurred. At national scale, Australia’s mining wealth has supported the same boom in services that Imas’s framework predicts. At regional scale, the picture is harder. The relational sector does not, on the available evidence, automatically follow the productivity gain to the place where the gain happened. It follows the spending to the place where people choose to live.

This is a non-trivial caveat. It suggests that the AI transition Imas is describing, even if it broadly works as he predicts at the level of national economies, may produce sharp regional dislocations in places that depend on the commoditised industries the transition is automating. Australia knows what this looks like, because it has lived through one round of it already. The next round is likely to be larger, faster, and to affect places that have not previously been on anyone’s list of vulnerable regional economies.

Imas’s framework, in its American version, leans heavily on healthcare and care work. In the United States, healthcare is a market sector that absorbs a large and growing share of household spending as people get richer; the relational sector argument applies very directly. In Australia, the rails are different. Medicare bulk-billing, the public hospital system, and the Pharmaceutical Benefits Scheme together provide most healthcare without direct out-of-pocket spending by the patient. Private healthcare exists and has grown, but it is not the safety net. The rich and the poor alike, in Australia, see the same GP for the same fee, paid by the government.

The implication for Imas’s argument is interesting and, as far as I can tell, undiscussed in his essay or its coverage. The relational sector in Australian healthcare is enormous. Healthcare and social assistance is the single largest employer in the country, with more than two million workers as of 2026. The work is, in Imas’s terms, deeply provenance-rich and substantially irreducible to AI. The patient who comes in to see a GP about a chest pain is not buying a generic medical opinion. They are buying the judgement of that particular GP, who knows their family history and noticed last time that something seemed off about their colour. Aged care workers, nurses, midwives, physiotherapists and a long list of allied health professionals do work that no current AI can do.

The catch, which Australian readers know well, is that the price signal Imas relies on does not flow through to these workers in the way his framework assumes. The Royal Commission into Aged Care Quality and Safety, which reported in 2021, found that aged care workers were paid in the order of twenty-five to thirty dollars an hour for work that was both essential and brutally difficult. The pay rates for nurses and personal care assistants have improved modestly since but are still, by international standards for skilled relational work, low. The reason is that the price signal is filtered through the federal budget rather than through individual customers. When the buyer is the government, the willingness to pay for the human element is mediated by political processes, not by mimetic desire.

I worked for twenty years inside this system. The lesson I would offer Imas, with respect, is that the relational sector is real, irreducible, and economically valuable, but in countries with publicly funded care it does not translate automatically into rising wages for the relational workers. The argument needs an extra step about how that value gets to the people doing the work. Australia has not, in two decades of policy debate, found the extra step.

Imas, to his credit, names one large caveat in his own essay. The framework he has built works for rich societies. It assumes a population whose income is rising fast enough to fund a transition into more expensive relational services. It does not work, or works much less well, for societies whose economies were built on producing commodities for rich countries. The Indonesian factory worker, the Vietnamese shipyard welder, and the Pacific Island freight operator are, in Imas’s framing, the people most exposed to the AI transition, because their economies depend on the production sectors AI is making cheap, and they do not have the income to redirect into the relational sector that follows.

Australia’s position, in this picture, is uncomfortable. Australia is a rich country whose wealth was built on exporting commodities, principally iron ore and coal, to the manufacturing economies of Asia. Those manufacturing economies are themselves now likely to be hollowed out by AI commoditisation, and the supply chain in which Australia is a producer of inputs runs through them. The transition Imas is describing is asymmetric. The countries best positioned to make it are the ones whose service sectors are already large and whose populations are already rich. The countries that will struggle are the ones who have spent the last fifty years tooling up to be commodity producers for the rich. Australia, at the level of its economy, sits awkwardly between the two.

This essay is a side piece from a book I am writing on the history of computing, Lady Lovelace’s Objection, which is being serialised on this Substack. The book takes its name from a sentence Ada Lovelace wrote in 1843, in notes attached to a French paper about a machine that did not yet exist. The Analytical Engine, she wrote, has no pretensions whatever to originate any thing. It can do whatever we know how to order it to perform.

The question Imas is answering, in 2026, is the next move in an argument almost two centuries old. AI does not think. AI produces. The economic question that follows, when the producing becomes very cheap, is what is left for humans to do that the machine cannot abstract. Imas’s answer, with the qualifications I have flagged, is that the relational sector, the parts of life where the human is the point, is what is left. The Sydney barista in the opening of this essay is, in this small framing, the version of Lovelace’s distinction that lives in everyday Australian life. She does what she can be ordered to perform, with extraordinary skill. She also does what nobody could order anyone to perform, because she does it as herself.