AI Can Grow the Economy.
Who Captures the Gains?

Artificial intelligence promises to make work more productive. But an economy that produces more does not guarantee that the median worker receives a proportionate share of the additional wealth.

A company can deliver the same service with fewer working hours. It can use the capacity released to sell more. It can lower prices, raise wages or retain the savings as profit. Technology opens up these possibilities; competition, the organisation of work and bargaining power help determine which one prevails.

Our thesis: AI can increase global GDP while its benefits remain concentrated. For investors, identifying that concentration matters as much as anticipating adoption.

What the evidence supports

Reported — productivity within a task. In a study published in the Quarterly Journal of Economics in 2025, introducing an AI assistant to 5,172 customer-support agents increased issues resolved per hour by an average of 15%. The effects varied across workers, with particularly meaningful benefits for less experienced employees. This is evidence from a specific setting, not an estimate for the entire economy. [1]

Reported — employment exposure. An ILO–NASK analysis published in May 2025 estimated that one in four workers worldwide held an occupation with some exposure to generative AI. Only 3.3% of global employment fell into the highest exposure category. Its central conclusion points towards the transformation of roles more than their wholesale replacement. Exposure measures the potential to change tasks; it does not measure future redundancies. [2]

Reported — the limits of early evidence. A 2025 study of Denmark found task reorganisation but no significant effects on earnings or recorded working hours during the first two years after ChatGPT's launch. This does not establish that future effects will be absent. It shows that adoption, productivity and pay can move at different speeds. [3]

Modelled scenario — global GDP. An IMF working paper published in April 2025 estimated a level of global GDP between 1.3% and almost 4% higher over the following decade than under a baseline without those AI gains, depending on productivity assumptions. These figures do not mean 1.3% to 4% of additional growth every year, and they are not a guaranteed forecast. [4]

From time savings to economic growth

Companies must account for the cost of tools, integration, human review and errors. Minutes saved on a task do not automatically become additional sales or sustainable cost reductions.

There are also two distinct stages. Building data centres and installing equipment mobilises investment. Using that capacity to produce more, at higher quality and lower total cost, requires adoption and organisational change. The first can advance before there is sufficient evidence of the second.

Tracking AI investment alone therefore leaves half the analysis unfinished. We also need to examine the returns generated by its users.

Who can capture the gains?

Workers, when AI complements valuable skills and they have the bargaining power to translate productivity into higher pay, better conditions or shorter hours.

Consumers, when competition forces companies to lower prices or improve their service.

Shareholders, when a company retains the savings, expands demand and prevents technology costs from absorbing the benefit.

Infrastructure and technology suppliers, when they control resources that are difficult to replace and can charge for access to them.

These outcomes can coexist. The question is their relative weight. A tool available to everyone can improve productivity across an industry while competition simultaneously compresses margins within that industry.

The same reasoning applies across countries. The IMF model points to uneven benefits, depending on sectoral exposure, readiness to adopt AI and access to technology. That result depends on the model's assumptions; it is not an inevitable outcome. [4]

The less visible risk in the labour market

The debate often focuses on the number of jobs eliminated. An earlier change may appear in hiring: slower team expansion, fewer entry-level roles and higher expectations for people starting their careers.

MITUXA hypothesis: if basic tasks are automated without replacing their role in learning, some companies could weaken the development of the professionals they will need in the future.

There is a countervailing force. Lower costs can expand demand and make previously unaffordable services viable. New tasks may also emerge. The analysis needs to track substitution, complementarity and demand expansion without assuming in advance which will dominate.

Three possible paths

Broad diffusion. Adoption spreads across businesses of different sizes, training keeps pace with change and gains reach wages and prices. We would look for sustained improvements in productivity, real pay and service quality across a broad range of companies.

Concentrated gains. A small group of companies retains most of the benefit while hiring and wages lag. We would look for concentrated profits and a persistent divergence between productivity and pay. This is the central tension we want to monitor, not a quantified forecast.

Insufficient returns. Spending grows faster than economic usefulness. We would look for utilisation below expectations, high review costs and deteriorating cash flow after investment. The technology can continue to improve even as individual investments fail.

What this means for investors

The presence of “AI” in a presentation says little about the quality of an investment. The useful questions are economic:

• Does the company solve a problem customers will pay to address?
• Does it retain the savings, or pass them to customers through lower prices?
• Does it have a durable advantage in distribution, usable data, integration or trust?
• Do the returns justify computing, implementation and asset replacement costs?
• Does the share price already assume near-perfect execution?

Revenue per employee is insufficient: it can rise after headcount reductions without a sustainable improvement in the business. We prefer to assess it alongside customer retention, margins, return on invested capital and free cash flow per share after investment needs.

It also matters to distinguish temporary scarcity from structural advantage. A supplier can benefit from limited capacity today and lose pricing power as supply expands.

What would change our view?

Our concern about concentration would diminish if sustained real-wage gains, productive adoption among small businesses and new entry-level opportunities emerged.

It would strengthen if gains accumulated in the profits of a few companies alongside prolonged weakness in hiring and pay.

The growth thesis itself would weaken if adoption advanced without lasting productivity improvements after accounting for implementation, oversight and investment costs.

Beyond the Obvious

AI can make information cheaper to produce. If it does, value may shift towards the ability to verify, decide, distribute and take responsibility for the outcome.

The decisive question is who can turn technological capability into economic value — and retain a share of that value after competition and investment.

Higher GDP does not guarantee shared prosperity. And a successful technology does not guarantee a profitable investment at any price.