The missing market in the AI jobs debate
The AI and jobs argument now has a familiar beat. A new model does something impressive. That impressive thing is mapped onto an occupation. A chart appears. A number of jobs is put at risk. Then some labour market data arrives (graduate hiring down, entry-level postings thinner, unemployment up) and is bolted on as proof.
Cue the chorus: it’s the end of the world as we know it.
But two arguments are being fused together. One is about what the technology can do. The other is about what is happening to employment. I have written about both, but the thing I keep coming back to is almost absent from the debate: the market.
We are so busy predicting the end of our careers that we are paying too little attention to how the market for AI develops, and that market will largely decide what AI does to jobs.
You took me to the limit of my capabilities
Technologies do not spread just because they work. Steam took decades to move through the economy. Modern mass manufacturing and consumer culture had to wait for electricity. Yet no one doubts that both reshaped employment (and society).
A technology is adopted when someone can buy it, at a price that beats the alternative, through a channel that exists, with the boring but essential organisational apparatus around it: contracts, liability, assurance, skills, integration.
It must also make something possible: a new good or service, or an old one produced better, faster or cheaper.
AI capability tells us what is possible. It tells us much less about what is probable. That is a problem because most published AI job-exposure analysis measures this capability and presents it as a forecast.
The logic becomes: “because this model can do this task, it must replace this job — and everyone else’s.” Benedict Evans has argued that the exercise is not just hard, but close to impossible: back-test it and it fails. As he shows a century of automating accounting produced more accountants, not fewer. A 1995 model of internet exposure would almost certainly not have spotted taxi drivers being displaced by Uber, because the change was not simply about automating the known skills of a taxi driver.
The question is not simply whether AI can substitute for a person on a task. It is whether we understand the bundle of skills, knowledge and experience a person uses across many tasks. I am a pretty smart guy, but my writing is not always clear. AI has been superb for me because it makes my writing clearer. The ideas remain mine, rooted in my own reading and thinking. It has not replaced a writer I could not afford to hire. It has reduced the time and number of drafts I need, and replaced a task i am frankly awful at.
But the narrower point is this. Even if you could describe a job perfectly, and even if you knew exactly what a model could do, you would still not know the employment consequence. Because the consequence depends on price. And once you look at price, the story changes.
Rage Against the Machine
Replacing labour with capital is a relative-price decision. Which is cheaper: the person or the machine?
Steam was not adopted because people admired steam. It was adopted because it could do required work faster and cheaper. Whether it powered looms or steam hammers, the technology reduced costs or increased returns enough to justify the investment.
So the interesting question is not what AI can do, but what firms are willing to pay for it.
Currently, it is not much. According to The Economist, the median firm spent around $10.66 per worker per month on AI in June. Close to half of British businesses using AI don't pay for it at all, making do with free tiers or open-source models. Only about one in ten small and medium-sized businesses has paid for a dedicated AI tool. The average American executive uses AI for something like 1.7 hours a week.
We are being asked to believe that a technology capable of displacing professional salaries is being bought, by the median firm, for roughly the cost of two coffees per employee per month. Maybe you are cynical enough to believe that. I am more inclined to think the real subject of the AI and labour debate is not capability but the gap between what AI appears able to do and what firms are actually willing to pay for.
So what is going on? Strip AI out for a moment and a different picture appears: risk, taxes, the unwind of earlier hiring, and firms waiting to see which way the wind blows. AI may be part of the story. It may even be a useful excuse in some cases. But unless firms are willing to treat AI as a serious operating cost rather than a speculative experiment, it has not yet met the price test for replacing people at scale.
Nor is adoption obviously still climbing. Roughly a fifth of American firms report using AI in any business function; a third of British firms do, though the question is asked differently. The US Census Bureau noted this spring that usage had held broadly steady across many sectors over the previous six months, and one academic survey found workplace use falling from its 2025 peak. It also remains shallow: only about a tenth of euro-area firms using AI describe doing so intensively, and around half of German users apply it to 5% of working hours or less.
A big caveat: survey definitions differ; one apparent decline may be noise. But the overall shape is consistent. It is not the shape suggested by the displacement charts.
It's a Beautiful Thing
Steam and electricity had to produce something: either something new, or something old made better, quicker, or with more output per input. The technology itself is not the thing. It is the thing that makes other things possible.
What this means is simple: firms do not get value from a general-purpose technology by buying it. They get value by rebuilding around it. Processes, data systems, organisational charts, supplier relationships, skills, operating models: all of it must move to make the most of the technology.
Historically, for every dollar spent on computer hardware, firms have typically made five to ten dollars of complementary intangible investment. Applied to current AI capex, that implies complementary spending in the trillions each year. So far, that spending has not appeared. In fact, investment in US organisational capital has been falling as a share of GDP. The data-plumbing firms that should be obvious winners have revenues in the billions, not the hundreds of billions. And the share of American workers leaving their jobs sits near an all-time low — not what you would expect from a workforce being rebuilt around a new technology. And we have a similar story in the UK.
Reorganisation shows up in labour flows. These flows are unusually quiet.
Meanwhile, nine in ten executives say AI has had no effect on their firm’s productivity over the past three years, and few firms report saving money by replacing staff.
Technologies reach the labour market through firms reorganising themselves. That reorganisation is not happening at anything like the implied scale. You can believe it is coming — I do — while still recognising that a forecast which skips this step is not really a forecast.
AI can’t compete with History
Because weak labour market signals are so often read as AI effects, graduate hiring down, fewer entry-level roles, we are giving too little weight to at least four other explanations: the unwind of 2021 over-hiring, the end of near-zero interest rates, soft demand in slowing economies, and demographic ageing. These are rarely named, barely tested against, and almost never used to calibrate the claim that AI is replacing opportunity and jobs.
These other explanations do not point the same way, which makes the analysis harder. Unwinding over-hiring suggest jobs disappearing. Ageing points sharply in the opposite direction: the binding constraint of the next two decades may be too few workers, not too many.
The key point is that the same technology can produce opposite labour market outcomes depending on the economy it lands in. Right now the AI narrative is winning attribution by default because it is the most available story, not the best-identified cause. It’s a panacea for all because it fits easier than the alternative explanations for what you can do in response:
If entry-level hiring is weak because demand is weak, you reach for macro policy.
If it is AI, you reach for training and transition.
If it is ageing, you are managing a shortage, not a surplus. If capable AI arrives into a shrinking workforce, the story may be substitution for missing workers rather than mass displacement — something that might actually help. (Yet we are not looking closely enough at who the ageing workers are, what they do, or how to measure the expertise they will take with them.)
The wrong diagnosis is expensive, and right now too much of the diagnosis feels like vibes. I would argue that all explanations are valid, the extent of each will vary across nations, regions, sectors and parts of the economy. If I were to pick one - I would always argue demand is the bigger problem (at least for the UK).
The strongest counterargument is that firms do not need mature AI deployment to stop hiring. Freezing graduate intake is instant, cheap and reversible in a way that redundancy and reorganisation are not. Yet firms are cutting entry-level hiring while reporting little to no productivity gain, doing little reorganisation, and spending ten dollars per head per month on AI. They also appear to be holding on to staff, whilst reducing new hiring. This doesnt suggest a graduate being replaced by AI. It suggests that the traditional pathways for graduates may be saturated and those firms are not able to expand their client bases in the same way to accommodate expansion. Sounds like a demand problem to me.
None of this means labour markets will not change. They will, substantially, and for some occupations brutally. But those changes cannot be read off the technology alone. They cannot be separated from demography and the economic cycle. And they cannot be estimated without a view of a market that has not yet found its price.
Technology tells us what is possible. Markets tell us what is probable. Demography and demand tell us what is needed. Most current analysis does the first and reports it as the third.