The Economy the Models Cannot See

“How are your economic models with thirty years of proprietary data doing?”

It is usually presented as a mark of strength. An investment firm has decades of data, a model refined across several cycles and an apparently scientific way of converting history into a forecast. But what if the difficulty is not the quantity of data? What if the relationships embedded in those data belong to an economy that is changing beneath us?

Models work by identifying relationships that held in the past. That does not make them useless. It makes them conditional. They work while the regime that produced those relationships remains sufficiently intact. When technology changes the way output is produced, a longer history can create false confidence rather than greater clarity.

Generative artificial intelligence may produce exactly this problem. Its effects will not necessarily arrive first as millions of announced redundancies. They may appear in projects completed without employees, vacancies that are never opened, smaller graduate classes, higher revenue per worker and rapidly rising demand for electricity and compute.

The economy could change materially before conventional indicators provide an unambiguous signal.

The Model Is Not the Economy

An economic model is a simplified description of relationships within a system. It can estimate what usually happens to employment when growth slows, to inflation when labour markets tighten or to investment when interest rates rise.

But models struggle when the structure generating the data changes. A relationship estimated during an economy in which cognitive output required human labour may not survive an economy in which machines can perform a growing portion of that work.

The traditional shorthand was:

Capital + Human Labour + Time → Output

The emerging production function adds a different group of inputs:

Capital + Energy + Compute + Models + Tokens + Human Judgement → Output

If output can grow without the historical increase in hours worked, several familiar ratios become less stable. Revenue may rise without proportional headcount. A new company may reach scale without building the organisation that an older model expects. Wage income may fail to capture the value of projects made possible by AI. Capital expenditure on data centres and electricity networks may become a leading indicator of productive capacity.

The problem is not merely that existing forecasts will contain the wrong coefficient. They may be looking at the wrong variables.

The Jobs That Never Existed

Public debate tends to ask how many jobs AI will destroy. That question directs attention towards redundancies: an identifiable company, a known number of employees and a visible date on which jobs disappear.

The first large employment effect may be quieter. A company may keep its existing workforce but reduce its graduate intake. A consultancy may decline to replace departing analysts. A technology company may allow a team of five to perform work previously assigned to ten. A founder may launch a business without hiring the developers, designers and administrators once needed to make it viable.

No employee is dismissed in any of these examples. Yet labour demand has fallen relative to the path it would otherwise have followed.

This is easy to miss in aggregate statistics. A company that dismisses 1,000 employees creates a visible event. Ten thousand companies each deciding not to recruit one junior employee create no comparable headline, yet together they prevent 10,000 jobs from coming into existence — ten times the visible loss. The quieter effect may also be more persistent, because those missing entry-level positions would have provided the training ground for the economy's future senior workers.

Unemployment data are designed principally to measure people who are seeking work and do not have it. They cannot directly count a counterfactual vacancy: the position a company would have opened in a world without AI.

Vacancies, hiring rates, graduate intakes, average team sizes, revenue per employee, spending on outsourced services and the employment content of new-business formation may therefore become more informative than redundancies alone.

The Apprenticeship Problem

AI creates an uncomfortable paradox. Companies need experienced people to supervise models, detect errors and exercise judgement. Yet experience has historically been produced by allowing inexperienced people to perform simpler work.

Junior analysts learn by gathering data, drafting routine material and watching how senior colleagues correct it. Programmers learn by writing and debugging relatively straightforward code. Accountants, lawyers and researchers acquire judgement through repeated exposure to cases that initially require close supervision.

Those are precisely the tasks that AI can absorb first.

The early productivity evidence suggests that AI can transfer some of the patterns used by high performers to less experienced workers. In a study of 5,179 customer-support agents, an AI assistant raised productivity by 14 per cent on average and by 34 per cent for novice and lower-performing workers.^[1] That is encouraging: AI can help beginners perform closer to the level of experienced colleagues.

But there is a difference between borrowing an expert pattern and becoming an expert. If entry-level work is removed altogether, the economy may enjoy an immediate productivity gain while weakening the mechanism that creates future supervisors.

This could eventually produce a barbell labour market: a small number of highly experienced people directing systems, together with lower-paid service work that remains difficult to automate, but fewer roles through which graduates traditionally moved from one group to the other.

The problem may not appear in senior-worker shortages for years. By the time it does, the missing cohort will already have passed through the labour market.

Exposure Is Not the Same as Unemployment

The International Labour Organization (ILO) estimates that one in four workers globally is employed in an occupation with some exposure to generative AI. It emphasises that transformation of work is currently more likely than the complete disappearance of occupations.^[2]

The International Monetary Fund (IMF) estimates that exposure is higher in advanced economies — about 60 per cent of employment — than in emerging markets, at approximately 40 per cent, or low-income countries, at 26 per cent.^[3] Rich economies employ more people in administrative, professional and cognitive occupations whose tasks are accessible to generative systems.

These figures do not mean that 60 per cent of advanced-economy jobs will disappear. A job is a collection of tasks. AI may automate some, complement others and leave the occupation intact. Falling production costs may also generate new demand and create work that cannot yet be anticipated.

The aggregate outcome therefore depends on three competing forces: substitution, where AI performs tasks for which businesses previously paid employees or outside contractors; complementarity, where AI raises the output and value of a worker; and demand creation, where lower costs expand consumption enough to generate new activity.

The historical analogy is not straightforward. Automation in manufacturing replaced physical tasks but also created large new service sectors. AI reaches into the service sector itself and can reproduce portions of language, analysis and code at very low cost. That makes the direction of redistribution easier to identify than the final number of jobs: returns are likely to move towards capital, compute, proprietary data, distribution and people whose judgement is difficult to reproduce.

Productivity May Appear in the Wrong Place

Official productivity is generally calculated by comparing measured output with measured labour input. AI complicates both sides.

Consider an individual who uses AI to build a research database that would otherwise never have existed. Capability and potential output have increased, but there may be no market price for the database and no avoided wage bill because no developer would have been hired. The gain is real, yet the national accounts may record little or nothing.

Alternatively, a company may use AI to produce the same output with fewer new hires. The early effect appears in profit margins and revenue per employee rather than in aggregate GDP. If the saved labour is not quickly redeployed, measured output may not initially rise at all.

Even when AI raises the quality of a service, statisticians must distinguish improvement from inflation. A faster, more accurate research product may be sold at the same price; a piece of software may become dramatically more capable without becoming more expensive. Quality-adjustment problems that already exist in the digital economy become more severe when the output is generated dynamically and customised for each user.

Controlled experiments show why the measurement issue matters. In one study of 453 professionals, generative AI reduced completion time for writing tasks by 40 per cent while increasing assessed quality by 18 per cent.^[4] A conventional measure focused only on the number of documents produced could miss part of the gain; one focused only on hours could misinterpret the quality change.

The economy may therefore experience a gap between technological capability and measured productivity. That gap should not be confused with evidence that nothing is happening.

Energy, Compute and Capital Expenditure Become Macro Data

AI-generated output consumes less human time, but more of other inputs: electricity, semiconductors, data-centre capacity, cooling, networks and capital.

The International Energy Agency (IEA) estimates that data centres consumed approximately 415 terawatt-hours of electricity in 2024, around 1.5 per cent of global electricity use.^[5] The aggregate share looks modest, but local effects can be large because data centres cluster around suitable grids and fibre networks.

This changes the macroeconomic transmission mechanism. AI adoption may be constrained not by the availability of labour, but by the time required to secure high-voltage transformers and grid connections, obtain planning approval, source advanced chips and finance the necessary infrastructure. Regions with reliable and comparatively inexpensive power may acquire a structural advantage. Regions with congested grids may face higher electricity prices or delayed investment.

The buildout is also capital intensive. It creates demand before the eventual productivity benefit is certain. Data centres, generation and transmission can support growth, but they can also create overcapacity if expected AI demand fails to materialise. The Federal Reserve has noted the growing importance of the AI buildout in US investment and the difficulty of measuring its eventual economic effects in real time.^[6]

An economic model concerned only with software adoption and labour substitution will therefore miss part of the cycle. The relevant dashboard increasingly includes semiconductor investment, data-centre construction, electricity demand, grid queues and the financing conditions surrounding infrastructure.

Inflation Could Move in Both Directions

AI is often assumed to be disinflationary because it lowers the cost of producing cognitive work. That may eventually be true for many services. Research, coding, administration and customer support can become cheaper per unit.

But the path need not be smoothly disinflationary.

The investment phase increases demand for chips, power equipment, construction, land and skilled technical labour. Grid constraints can raise local electricity prices. Capital expenditure may be financed during a period in which the productive return remains uncertain. Meanwhile, businesses may retain part of the labour saving as higher margins rather than immediately passing it to customers.

There is also a rebound effect. When a useful input becomes cheaper, total demand for it can rise. The economy may consume vastly more software, analysis and automated decision-making because each unit costs less. Falling unit cost does not guarantee falling aggregate expenditure.

The inflationary result depends on timing. AI can create a near-term investment boom, sector-specific bottlenecks and asset-price effects before producing broad service-sector disinflation. A static model may capture one side and miss the sequence.

What an AI-Era Dashboard Should Watch

No single new statistic will solve the problem. The practical response is to widen the set of indicators and focus on the margins where change first appears.

For labour, that means monitoring hiring rather than redundancies alone: entry-level vacancies, graduate intakes, recruitment by occupation, hours worked, contractor spending and the rate at which departing employees are replaced.

For companies, it means revenue and profit per employee, headcount growth relative to sales, compute expenditure, the size of teams at newly formed businesses and the changing mix between payroll and technology costs.

For physical capacity, it means data-centre construction, semiconductor supply, electricity demand, power prices, grid-connection queues and investment in generation and transmission.

For distribution, it means asking who captures the gain. Rising productivity can coexist with stagnant wages if the return accrues to model owners, infrastructure suppliers or shareholders. Aggregate output may look healthy while the political economy deteriorates.

Finally, models must leave room for newly feasible output. Economic value is created not only when an existing task becomes cheaper, but when a person or small business can attempt something that previously could not have been attempted at all.

History Is Evidence, Not Destiny

Thirty years of proprietary data remain useful. They show how relationships behaved under the production system that generated them. They do not establish that those relationships are laws of nature.

For investors, the danger is to wait for mass unemployment or a decisive break in productivity statistics before accepting that the structure has changed. By then, the transformation may already be visible in smaller graduate intakes, unusually high revenue per employee, rapid infrastructure investment and companies reaching scale without building traditional workforces.

For policymakers, the danger is to celebrate stable unemployment while the entry points into professional life are quietly narrowing. A labour market can look robust in aggregate even as it weakens the mechanism through which the next generation acquires experience.

For economists, the danger is to confuse what cannot yet be measured with what does not exist.

The best models of the AI economy will not necessarily be those with the longest histories or the greatest apparent precision. They will be those capable of recognising when the production function has changed, of observing employment that was never created and of combining traditional economic data with the new physical and organisational indicators of machine intelligence.

The model must serve judgement. It cannot replace it.