When Human Time Stops Being the Constraint

For most of economic history, output was tied to human time. A factory could equip a worker with better machinery and an office could give an analyst better information, but somebody still had to perform each step. Research, writing, coding, administration and coordination moved at approximately human speed.

Artificial intelligence is beginning to weaken that relationship.

This is often described as another productivity improvement. That is true, but incomplete. AI does not merely help people perform familiar tasks faster. In some circumstances it allows a person to undertake work that he or she could not previously have performed at all. It can turn an analyst into the effective manager of a research, editorial and technical production system. It can allow a small company to build capabilities that once required a much larger organisation.

That is a change not only in productivity, but in the production function itself.

From Human-Speed Production to Machine-Speed Execution

The old economy did not literally run on labour alone. It always required capital, energy, machinery and knowledge. But human time remained the binding constraint in most cognitive work. If an organisation wanted more analysis, more software or more client communication, it generally needed more people—or more hours from the people it already employed.

A simplified version of that production function was:

Capital + Human Labour + Time → Output

The emerging AI production function looks more like:

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

"Tokens" here do not mean cryptoassets. They are the units of information processed by an AI model: the metered raw material of inference. Agents and automation sit between the inputs and the result, executing and coordinating sequences of work at machine speed.

Human labour does not disappear. It changes position. Less human time may be required to produce every intermediate step. More value migrates towards defining the objective, providing context, checking the result and accepting responsibility for the final decision.

The decisive shift is therefore not from humans to machines. It is from paying primarily for human time to paying for machine capability supervised by human judgement.

My Own Small Experiment

I still feel slightly embarrassed telling people that I use AI to help write articles. If a writer is asked whether an article was drafted with AI, the question often sounds accusatory. The implication is that the work is somehow less legitimate because fewer human hours were spent producing it.

That reaction shows how strongly we continue to associate authorship with labour input. We accept financial terminals, databases, search engines, spelling checkers, research assistants and human editors. AI feels different because it participates directly in research, structure and expression—the activities we associate with authorship and intelligence.

In my own process, AI has increased productivity in the labour-intensive work of researching and writing an article by approximately three or four times. It helps investigate a question, challenge an argument, organise evidence and turn a collection of ideas into a coherent draft. It has not supplied my investment experience, selected what I believe or accepted responsibility for what is published. The judgement remains mine; much of the mechanical labour no longer has to be.

The effect becomes more profound when I move beyond activities I could previously perform myself. I am not a software developer. Without AI, building a database or developing a website would require programmers, database specialists, designers and perhaps a project manager. The cost and coordination might prevent the project from happening at all.

It is tempting to describe the resulting productivity increase as infinite. Strictly speaking, a percentage gain cannot be calculated from a starting capability of zero. The more important point is that AI has not merely accelerated an existing activity. It has given me productive capabilities I did not previously possess.

Economists distinguish between changes at the intensive and extensive margins. At the intensive margin, AI allows a writer, analyst or developer to produce more in a given period. At the extensive margin, it allows a non-programmer to build software, a small business to create much of its own operating infrastructure—the databases, research workflows, client-management systems and automated processes through which it functions—and an individual to attempt a project that previously required an organisation.

The second effect may eventually prove more important. It is also much harder to measure. Statistics can observe that an employee resolves more cases per hour. They are less capable of valuing a database that was never going to be built or a business that could not previously have been launched.

The Evidence: Large Gains, Unevenly Distributed

The early empirical evidence is meaningful. In an experiment involving 453 college-educated professionals, access to generative AI reduced the time required for mid-level writing tasks by 40 per cent while increasing assessed quality by 18 per cent.^[1] A study of 5,179 customer-support agents found that an AI assistant increased issues resolved per hour by 14 per cent on average. The gain was 34 per cent for novice and lower-skilled workers, with much less benefit for the most experienced.^[2]

These results suggest that AI can compress experience. It can capture patterns used by high performers and distribute them to people who have not yet acquired the same knowledge. That can be democratising: more people gain access to capabilities that were previously scarce.

But productivity is not automatic. AI performs best when a task is clearly specified, when the result can be checked and when errors are relatively easy to detect. Its apparent speed can become illusory when the user must reconstruct extensive context or inspect every detail of an unreliable answer.

A 2025 randomised trial by METR (Model Evaluation & Threat Research) examined experienced open-source developers working in mature codebases they knew well. Contrary to their own expectations, access to early-2025 AI tools increased completion time by 19 per cent.^[3] Complex context and the cost of reviewing imperfect output outweighed the speed of generation.

The lesson is not that AI has failed. It is that the value of AI depends upon the relationship between generation and verification. Where verification is cheap, machine speed is transformative. Where errors are subtle and consequential, human knowledge may remain the limiting factor.

The winners will not simply be those who use the most AI. They will be those who know when to trust it, when to constrain it and when to reject it.

When Code Is No Longer Scarce

Coding is the leading edge of this transition because software is abundant, structured and testable. A model can generate code and receive relatively rapid feedback on whether it works.

Anthropic's analysis of 500,000 coding interactions in 2025 classified 79 per cent of conversations using its coding agent as automation rather than augmentation. Web-development languages and user-interface work were particularly prominent.^[4] These data come from one provider and should not be mistaken for a census of the economy. But they confirm that coding is one of the areas in which users are most willing to delegate complete tasks to machines.

For decades, coding was a scarce and expensive skill. Building a software product required trained developers, capital and time. That scarcity supported software consultancies, outsourcing businesses and thousands of software-as-a-service companies. A company could charge a substantial recurring price for a narrow feature partly because reproducing it was costly and inconvenient.

AI is dismantling some of that scarcity. It can write, explain, translate, test and repair code. It enables experienced developers to handle more work and non-developers to create functional tools. As the cost of producing software falls, some companies will discover that their technological moat was not proprietary technology at all. It was the cost and inconvenience of reproducing their product.

AI is not killing software. It is commoditising the production of large categories of software.

That distinction matters. Businesses with proprietary data, embedded workflows, trusted brands, strong distribution, regulatory permission or network effects may become more valuable. Cybersecurity, data infrastructure and the systems supplying AI itself can flourish. But a company selling a simple, replicable feature at a premium multiple faces a different future when its customer can generate an adequate substitute or obtain the same capability inside a broader AI platform.

The margin structure changes as well. Traditional software could be written once and sold repeatedly at negligible marginal cost. Generative software moves part of the cost into continuous inference: electricity, chips, data-centre capacity and tokens. Economic rent may migrate away from easily replicated applications towards compute infrastructure, model providers, energy producers and businesses controlling proprietary data or customer access.

The Firm Becomes a System of Judgement and Agents

Companies exist partly because coordinating complex work through repeated market transactions is expensive. They hire people, establish processes and build managerial hierarchies to reduce those coordination costs. AI agents could reduce the cost of coordination itself.

A single model response is useful. An agent that can inspect a database, conduct research, write code, test the result, update a website and report exceptions begins to resemble a workflow. Multiple specialised agents connected to company systems and supervised by a small number of people begin to resemble parts of an organisation.

The autonomous company is not imminent. Reliability, security, legal accountability and access control remain formidable constraints. An agent capable of acting is also capable of making a consequential mistake at machine speed. Yet the direction matters. A founder may reach a scale that previously required dozens of staff. An established company may expand without adding headcount in the historical proportion.

AI-native firms may therefore look unusual when compared with the companies in traditional datasets: high revenue, low headcount, substantial compute expenditure and extreme dependence on a small number of judgement-intensive roles.

The effect is not necessarily a world without firms. It may be a world of smaller firms with greater productive reach.

The Digital Economy Becomes Physical Again

Machine intelligence substitutes for human cognitive time, but it is not weightless. It consumes electricity, semiconductors, cooling, networks, land and capital.

The International Energy Agency (IEA) estimates that data centres consumed about 415 terawatt-hours of electricity in 2024, approximately 1.5 per cent of global electricity use.^[5] The global percentage disguises severe local constraints. Data centres cluster around reliable grids, fibre connections, water, planning permission and available power. A country may have adequate total generation while a particular region cannot connect another hyperscale facility.

Software was once described as economically weightless. AI makes the digital economy physically intensive again. Behind each apparently effortless response sits a chain of mines, semiconductor fabrication plants, transformers, turbines, substations, servers and cooling systems.

The investment implications extend far beyond application companies. Beneficiaries may include semiconductors, networking, electrical equipment, grid infrastructure, generation and owners of scarce energy resources. The constraint on AI growth may arrive not through a shortage of ideas, but through permitting, grid queues, financing or electricity prices.

Energy and compute are no longer merely costs inside a technology model. They are becoming inputs into the productive capacity of the economy.

Human Judgement Is the Residual—and the Moat

AI will complement some people, substitute for some tasks, destroy some business models and create others. It will reduce the value of activities that can be generated cheaply and increase the value of scarce inputs surrounding them: energy, compute, proprietary data, distribution, trust and judgement.

For investors, the central question changes. Which companies own genuinely scarce assets, and which merely sell expensive versions of work that AI can reproduce? Which businesses control customer relationships and proprietary context? Which regions can supply the infrastructure required for compute? Which companies can expand margins with agents without destroying the knowledge on which future judgement depends?

For workers, the safest position may not be the person who produces the greatest quantity of routine output. It may be the person who defines the objective, understands the domain, verifies the result and accepts responsibility when the machine is wrong.

Using AI does not eliminate authorship or management. It changes what they mean. The author increasingly directs a production system. The manager increasingly allocates work between people and machines. The expert increasingly earns a return not for performing every step, but for knowing which steps matter and whether the result can be trusted.

The economy is moving from human-speed production towards machine-speed execution. Output is beginning to separate from hours worked and software from coding labour. In a world of increasingly abundant machine-generated work, human judgement may become the most valuable input of all.