Jevons and Baumol Walk into a Pub

AI may produce more software, art and economic value than ever while requiring fewer people to produce it. Jevons explains the abundance. Baumol explains why human time becomes expensive. Ownership decides who receives the saving.

So two economists walk into a pub.

One is William Stanley Jevons. The other is William Baumol.

Jevons buys the drinks, carries them across the room and hands one to Baumol.

This has the opening structure of a joke. What is less clear is who the joke is about, and who will be left laughing.

What Jevons actually said

William Stanley Jevons was writing about coal, not artificial intelligence. In The Coal Question, published in 1865, he challenged the idea that more efficient steam engines would reduce Britain's consumption of coal. Greater efficiency made coal cheaper to use for each unit of useful work. That encouraged more applications, more investment and more total consumption.

The idea is now known as the Jevons paradox, although the word 'paradox' is sometimes applied too casually. A rebound occurs whenever some of the expected saving is absorbed by additional demand. The full paradox requires demand to increase far enough that total resource use rises despite the efficiency improvement.

AI appears to be creating a substantial rebound in cognitive and creative production. When code becomes cheaper to produce, organisations do not necessarily commission the same amount of code and pocket the saving. They attempt projects that would previously have been rejected as too expensive. When an image takes seconds rather than days, we do not generate the same number of images. We generate hundreds, select a handful and start placing them on surfaces that previously had no illustration at all.

The same applies to reports, campaigns, product variations, customer messages, translations, presentations, videos and music. Lowering the cost expands the possible uses.

Jevons helps explain why AI is likely to produce more work rather than simply complete the existing volume more quickly.

He does not tell us how many people will be required to do it.

Baumol and the string quartet

A century later, William Baumol and William Bowen examined a different problem. A string quartet performing a twenty-minute piece still requires four musicians and twenty minutes. The musicians may become more skilled. The instruments may improve. The hall may provide better sound, and the performance may reach a larger audience through recording and broadcast. The live performance itself remains stubbornly resistant to labour productivity.

That creates what became known as Baumol's cost disease. Wages in more productive sectors rise. The quartet must compete for people who also need to pay the rising costs of housing, food and everything else. Its costs therefore rise even though its measured output per working hour does not.

The phrase makes it sound as though the quartet is ill. It is not. The economic problem is that some forms of value require human time, and human time has not become more plentiful.

This was the point at which Jack Conte struggled to continue during his recent talk on creativity in a world with AI. At around thirty-two minutes, he explained that there was no way to squeeze efficiency from the musicians performing the quartet.

He stopped for seventeen seconds.

The audience applauded. He thanked them, apologised and tried to return to his notes. He had to apologise again.

Conte had already described a career repeatedly enabled and then destabilised by technology: home recording, YouTube, algorithms, streaming, membership platforms and the demand for continuous short-form output. He was not mourning a world in which technology had never helped him. He was confronting the economic logic that can look at four skilled people creating something together and see an inefficiency waiting to be removed.

That moment led me back to a book I published sixteen months ago, and to an argument I had not fully developed when I wrote it.

More output does not require more people

The employment relationship can be expressed simply enough:

People required ≈ total output ÷ output per person.

Jevons pushes the first number upwards. Lower costs and greater capability expand demand.

AI pushes the second number upwards. Each person, or each small AI-enabled team, can produce considerably more.

If an industry produces three times as much while output per worker rises fourfold, the industry has grown dramatically while requiring 25 per cent fewer people. There is no contradiction. Production, revenue and customer activity can all increase while employment contracts.

Baumol adds pressure at the point where human productivity cannot be increased without changing the activity. An organisation can pay the rising relative cost of the human version, reduce the amount of human time within it, or substitute a different product generated through a more scalable process.

AI does not make the quartet more productive. It gives the purchaser another way to obtain music.

Together, these forces create the possibility of job-light abundance: more of almost everything, produced by relatively fewer people.

I have seen the denominator move

I work with a small engineering team. We recently rebuilt a platform that had accumulated over five years, originally created by a team of six, in seven months with three people.

The new platform is not a compromised imitation. It is more enterprise-ready, with serverless infrastructure, autoscaling, tighter security and deployment managed through GitHub. We were able to draw on established services rather than build every capability ourselves, and to use Kiro for a breadth of architectural, infrastructure and implementation knowledge that no small team could previously have held continuously.

The unit of work changed. We increasingly built epics rather than moving carefully through stories.

This is one small example. It does not establish what is happening across the whole technology industry. The earlier team also built in a different technical environment, without the same mature cloud services, development tooling or accumulated product knowledge. It would be dishonest to assign the entire improvement to generative AI.

But the experience removes one uncertainty. The productivity change is technically real. A much smaller group can now hold and deliver a much larger scope, while producing a more operationally mature result.

No one gave us half the day back.

Greater capability increased the ambition of what the team was expected to accomplish. Once an organisation has seen three people deliver at that level, six people doing the previous amount of work no longer looks normal. An exceptional result becomes evidence for the next staffing model.

This is how augmentation becomes a ratchet.

The jobs that disappear before they exist

Most public discussion looks for displacement in redundancy announcements. That misses the quieter mechanism.

AI does not need to remove the junior employee. It can remove the vacancy through which that person would have entered.

In August 2026, researchers at the Stanford Digital Economy Lab updated their analysis of payroll records. They found no evidence of widespread, economy-wide AI displacement. They did find that employment among workers aged 22 to 25 in highly AI-exposed occupations was approximately 19 per cent below where it would have been if it had kept pace with employment in less-exposed occupations. The adjustment appeared primarily through reduced hiring rather than increased dismissal.

The researchers are explicit that this is descriptive evidence, not proof that generative AI caused the entire divergence. Technology hiring has also been affected by post-pandemic overexpansion, higher interest rates, economic uncertainty and ordinary business cycles.

The pattern is nevertheless difficult to dismiss. It is concentrated among younger workers, in exposed roles, at the point of hiring rather than separation. The same Stanford work also found that the decline was concentrated in occupations where AI is used to automate tasks; employment was flat or rising where it was used more as a complement.

A UK government assessment found a similarly uncomfortable signal. Between 2022 and 2025, job adverts fell by 38 per cent in highly AI-exposed occupations, compared with 21 per cent in less-exposed work. It also noted that the number of 16 to 24-year-olds employed in computer programming fell by 44 per cent during 2024, while correctly warning that the fall cannot be attributed to AI alone.

Entry-level work has never been merely cheap production. It is where people learn to operate inside real systems, encounter exceptions, receive correction and turn codified knowledge into judgement. If experienced engineers can use AI to perform work previously allocated to graduates, the immediate efficiency may be genuine. Five years later, the organisation may discover that it has removed the route through which experienced engineers are made.

The alternative is not always that the junior role vanishes. Sometimes it remains, but quietly asks for a senior person. PwC's 2026 analysis found that AI-exposed junior job adverts were seven times more likely than less-exposed junior roles to demand traditionally senior skills such as strategic thinking and leadership. It also found that these 'seniorised' entry roles had grown, even while other entry-level roles declined. The door is still marked entry. The person expected to walk through it has already acquired experience somewhere else.

At the other end of working life, the evidence is less clean. Stanford found no comparable employment gap among experienced workers. That should restrain any claim that AI is already removing older workers at scale. A June 2026 research brief from the Center for Retirement Research did find that US workers aged 55 and above in highly exposed occupations had become somewhat more likely to leave work and move into unemployment since ChatGPT's release. It describes an emerging risk rather than a settled causal result.

The practical concern may sit partly between those datasets. Experienced incumbents retain organisational knowledge, relationships and authority. An experienced person trying to re-enter the market has none of that protection. They arrive with a higher expected salary, a career history that may not match the latest toolchain, and the usual burden of age discrimination. AI does not have to create those disadvantages to intensify them.

Graduates struggle to get onto the ladder. Older workers struggle to get back onto it. The people securely positioned in the middle are asked to cover more rungs.

India and the weakening link between growth and employment

The campus recruitment posters have not disappeared.

Tata Consultancy Services onboarded more than 44,000 freshers during its 2026 financial year. It remained one of India's largest graduate recruiters. Yet TCS ended that year with 584,519 employees, a net reduction of 23,460. In June, its chairman told shareholders that the company would continue hiring, but that the rate of employee additions would not be what it had been and the ability to recruit large numbers would cease to be a useful HR metric.

That is not the closure of the graduate route. It is something more complicated: tens of thousands of people can still enter a company whose total workforce contracts over the year. TCS subsequently added 9,279 employees in the quarter to June, further evidence that this is not a clean line from AI adoption to fewer jobs.

For decades, India's technology growth model was visibly tied to people. More client demand meant more delivery centres, more graduates and more billable employees. Headcount was not an unfortunate cost attached to growth. It was how growth happened.

That relationship is weakening.

In its 2026 strategic review, Nasscom projected that India's technology industry would reach $315 billion in revenue for the financial year, an increase of 6.1 per cent. It expected direct employment to reach approximately 5.95 million, an increase of around 135,000 jobs, or 2.3 per cent. Even on that projection, the industry was still creating jobs. It was creating them much more slowly than revenue.

This is not evidence of collapse. It may be more important as evidence of direction.

Revenue growth is running at almost three times employment growth. Global capability centres and AI specialists are expanding while large service firms become more selective about hiring, retrain existing employees and increasingly sell outcomes rather than volumes of labour.

The causes are not exclusively AI. Client demand, pricing, utilisation, offshoring patterns, economic cycles and the changing mixture of work all affect the ratio. AI is arriving inside a business model that was already trying to loosen the link between revenue and headcount.

Jevons appears in the continued growth. Baumol appears in the pressure to remove expensive human time from each unit delivered.

India suggests that the AI economy does not need to stop growing in order to become less inclusive. It can become larger, more capable and more profitable while offering relatively fewer people access to each increment of growth.

The campus still opens. The escalator above it is carrying fewer people.

The music survives. The musicians may not.

Conte's hope is real, but it operates at the level of humanity rather than employment. People will continue to make beautiful and meaningful things. AI will help some of them make work that would otherwise have remained beyond their technical ability, available time or budget.

I know this because I have done it. I have used AI to make music, build a highly bespoke website and turn ideas into things I could not reasonably have commissioned. I do not regard those experiences as fraudulent. The tools allowed particular human ideas to survive the journey into production.

It does not follow that a similar number of people will earn a living making those things.

The early evidence from creative work is mixed. Research on a large freelance platform found that workers in more AI-exposed occupations experienced a 2 per cent reduction in contracts and a 5 per cent fall in earnings after the arrival of generative AI tools. The effects were greater among some of the more experienced and highly rated freelancers, which complicates the comforting assumption that quality alone provides protection.

Broader research covering artistic occupations has found no general short-term collapse in artists' earnings. Employment results are less stable, and working hours appear to be changing more clearly than pay. That ambiguity makes sense. Creative work is not one labour market. A recognised performer with a committed audience occupies a different economy from a freelance illustrator producing early concepts, a copywriter drafting product descriptions or a junior composer creating inexpensive background music.

Upwork's own 2026 data provides an almost suspiciously neat Jevons example. Generative AI and creative-production work on the platform experienced 90 per cent year-on-year growth in contract starts while earnings per contract fell by 13 per cent. Platform data are not a census of creative employment, and Upwork has a commercial interest in presenting freelance markets as adaptive. Even with that caution, the combination is revealing. The number of transactions grew rapidly. The return from each transaction did not.

AI may expand both ends of the market. More people will create recreationally because the barriers are lower. A smaller number of established creators may use the tools to produce ambitious work with very small teams. The professional middle can contract between them.

Human creativity survives. Creative livelihoods become harder to sustain.

The same pattern can reach software, journalism, advertising and game development. The project continues. The team becomes smaller. The artefact may even improve.

Looking back, Countdown to 2100 anticipated part of this. I wrote that mass-produced art and music would increasingly be generated by AI while rare human talent became a premium experience, cherished partly because it was becoming scarce. What I did not consider carefully enough was how much more creative material might be made, or how easily abundance could conceal declining participation.

We may get more music than any previous society and fewer working musicians.

Where the dividend goes

There is credible evidence for a more optimistic interpretation. PwC's 2026 AI Jobs Barometer found that the most AI-exposed companies had increased headcount by 52 per cent relative to a 2018 baseline, compared with 36 per cent among the least-exposed companies. Wage growth was also stronger. These are associations across different kinds of firms rather than proof that AI caused the additional employment, but the data do not support a simple claim that AI exposure automatically destroys jobs.

The same report found strong productivity growth among the firms best able to integrate AI at scale. That is good news if the gain produces lower prices, better services, higher wages, new employment, shorter working time or investment that benefits a wider population.

There is no automatic mechanism requiring any of those outcomes.

A worker who completes a task 50 per cent faster does not normally receive the other half of the day. The organisation recalculates capacity. The gain can become more output, reduced headcount, lower prices, higher margins or some combination of them. Competition may eventually transfer part of the saving to customers. Owners of scarce infrastructure, platforms, data, distribution and capital remain particularly well placed to capture it.

For decades, the Turing Test provided our cultural image of AI maturity: a machine sitting obediently in an examination room, trying to persuade a human that it was one of us. The economically important systems are no longer waiting to be examined. They are being given access to tools and asked to perform valuable actions. Displacement does not require an artificial person. It requires a useful substitute for enough of a person's work.

At almost the same moment, the scale at the other end of the economy became difficult to ignore. SpaceX's flotation in June 2026 made Elon Musk the world's first trillionaire on the prevailing market valuation. The company's prospectus described Space, Connectivity and AI as its three foundational competitive advantages, and AI compute as its next trillion-dollar market.

We should not pretend that a missing graduate salary travelled directly into Musk's shareholding. His wealth spans several companies, SpaceX's flotation was the immediate valuation event, and a market valuation is not cash in a bank account. The comparison is not a causal proof. It is a picture of the different scales at which the same transition can be experienced.

For the worker, AI appears as a missing vacancy, an intensified workload or another demand to retrain.

For the owner of scalable infrastructure, it appears as asset appreciation and a claim upon future productivity.

The acceleration reaches everyone. The benefit does not.

What Jevons and Baumol allow us to see

Jevons would recognise the expansion of use. Software, analysis, imagery, music and organisational communication become cheaper, so we produce and consume more of them. He would not tell us that demand must expand enough to preserve employment. That depends upon how quickly output grows relative to productivity.

Baumol would recognise the rising relative cost of work that still requires human presence. He would also remind us that replacing a quartet with generated music has not increased the quartet's productivity. It has changed the product being purchased.

Neither theory says that technological progress is undesirable. Greater efficiency can create genuine abundance. Synthetic production can coexist with valuable human work. New industries and professions can emerge.

Neither theory decides who owns the capability, who receives the saving, how much human participation society chooses to preserve, or whether productivity becomes leisure for the many or wealth for the few.

Those are political, organisational and moral choices. We often disguise them as the inevitable consequences of technology because inevitability is more comfortable than ownership.

Sixteen months into seventy-five years

Countdown to 2100 made strong claims about the eventual displacement of human labour. I argued that more than half of existing jobs could be performed by AI and robotics within thirty years, and that the new roles created would not replace them at anything close to the same scale. At one point I suggested one new AI-related job for every thousand replaced.

On the evidence available today, that ratio was too categorical. We are not seeing economy-wide mass unemployment caused by AI. Experienced workers in exposed occupations have often remained resilient. AI-skilled roles command higher wages, and some highly exposed firms are hiring faster than less-exposed ones.

The more difficult correction is that the book treated replacement too much like an event. A job existed, a machine arrived, and the job disappeared. Some displacement will look like that. The earlier mechanism is often less visible.

The graduate vacancy is never opened. A team of three takes on work that would previously have supported six. An Indian technology industry grows revenue almost three times as quickly as employment. A creative marketplace records far more AI-enabled contracts while average earnings per contract fall. Those who remain employed are expected to deliver at the new speed.

The book anticipated displacement. It underweighted job-light growth.

It anticipated concentration. It did not imagine that the first trillionaire would arrive quite so quickly, through a company presenting AI as one of its three foundational advantages.

It anticipated rare human creativity becoming a luxury. It did not sufficiently consider the strange possibility that human creativity could flourish in aggregate while creative lives became less economically possible.

Sixteen months cannot validate a forecast covering seventy-five years. It can reveal whether the mechanisms beneath it are beginning to appear.

They are.

Not everywhere. Not cleanly. Not with AI as the only cause.

But clearly enough that waiting for an economy-wide unemployment spike would be a poor way to decide when the problem has started.

Last orders

The dream of productivity was that we might accomplish the necessary work more quickly and reclaim some portion of our lives.

The emerging reality is that the saved time belongs first to the organisation. The worker is expected to produce more, the company is expected to grow, and the owner holds the asset through which the gain compounds.

At six o'clock, the barman called last orders.

It seemed early, but the evening had met every productivity target. Everyone had consumed three or four times as much, with fewer people required to serve them.

Jevons looked at the empty glasses.

Baumol looked at the empty chairs.

The shareholders looked at the till.

I work with organisations navigating this shift, fractionally, as an adviser, or as a trusted collaborator. See how I work →