By Markus H.-P. Müller - 12 August 2026

Most of us are now using AI every day. But we still have no clear idea about what it means for our economies, societies or the environment. Vigorous debate continues about AI’s likely future impact on the labour force. We pay less attention to how the future economic rewards from AI should be distributed. My argument in this post is that we need a new accounting framework if we are to align AI’s costs with its potential benefits; this needs to consider not just AI-related revenue and expenditure but also a balance sheet for AI’s impact on “capital” in all its forms – produced, natural, human and institutional. As I discuss below, some aspects of this balance sheet are currently better developed than others: developing a coherent “capital” assessment will require work by both international bodies and corporates. I think that the priority is to develop a balance sheet that can highlight the key risks and opportunities around AI, rather than necessarily provide a comprehensive assessment.

If history can teach us one thing, it is that it is very difficult to predict the future. Societies at the beginning of the industrial revolution had little idea what it would eventually bring – although they could see the immediate pain.

Economics offers less guidance than we might expect. Because we live in an age of economic forecasts, we often see economics as forward-looking – a predictive social science. But this is not often the case: economics is essentially backward-looking.  Understanding and explaining an economic order is done by analysing its current and past operation:  Adam Smith was the original master of this. Even those economists who might be regarded as more prophetic – Karl Marx or J M Keynes, for example – have based their predictions on a broad assessment of why current economic systems do not work well or are not sustainable.

Over the last half century, economists have certainly not proved good at predicting changes to the economic regime. Consider three dates: 1973 (the first oil shock), 2008 (the start of the global financial crisis) and 2022 (the European energy crisis). The triggers for these crises were all different, but all (in retrospect) predictable.

In 1973, an exogenous impulse (higher oil prices) struck a functioning, if stressed, economic order. This caused a crisis within the economic order: Keynesian demand management failed against a phenomenon it had not theoretically anticipated. Policy responses were eventually found, e.g. monetary targeting, central bank independence, but only after a delay of five to ten years.

2008 was an endogenous failure. Securitisation was a legal, regulatorily sanctioned innovation; the received wisdom was that it was risk-reducing, not risk-creating. Some aspects of government policy (e.g. quantitative easing) were implanted quickly but note that international banking governance (in the form of Basel III) was not updated until 2010 – long after the crisis started.

The European energy crisis in 2022 was different again. The risk that Russia could curtail energy supplies to Europe was always evident: nobody in Paris, Berlin or Brussels was unaware in 2021 of where most of Europe’s gas came from. But Europe had decided to go for the advantages or cheap energy, without also booking in the strategic dependency risk: this was a problem hiding “in plain sight”.

AI contains elements of all three of these three crises. The rapid technological development of AI is delivering an exogenous shock to the economic system, but its implementation is also revealing endogenous pressures. And accompanying AI risks – to employment, to the environment – have been evident for some time to those who cared to look.

Finding an effective policy response to AI may, however, prove even more difficult than to those economic regime changes of 1973, 2008 and 2022. The basic problem is that policymakers can only govern what already has a form and technology only acquires a form once it has stabilised. A brief history of the European Union’s AI Act, from its first draft in 2021 to adoption in 2024, reveals continuous legislative changes to reflect the evolution of artificial intelligence itself – towards a system where, perplexingly for those drafting the legislation, the defining feature was its lack of fixed purpose. The legislator was always in “catch-up” mode.

One worry therefore is that AI legislation (in Europe at least) will be permanently disconnected with reality. A legislative procedure in the European Union takes three to five years. AI model generation takes, depending on how you count, twelve to eighteen months. So, in the time it takes European policymakers to formulate a rule, the object of that rule changes shape two to four times. That is no longer a delay. That is a structural non-synchronicity.

If our legislative process cannot keep up with the development of AI, does it have a hope of keeping up with its implications? Let us first consider the question which currently dominates the AI implications debate: will AI take our jobs away? (There is also a second question of growing importance: can we deal with the environmental consequences of AI? I will return to this second question later.)

I would suggest that we start by setting this question of jobs to one side. Nearly three hundred years after the start of the industrial revolution, we already know an answer to it: possibly yes, but in the long-run aggregate, employment re-equilibrates.  This answer has four qualifications in one sentence. The “long run” could imply pain for individual workers across generations; by looking at the “aggregate” we also ignore the groups of society that lose out; “employment” implies a focus on the quantity of jobs, not their quality. “Re-equilibriation” also gives little detail on the final outcome. But, although wholly inadequate, this is probably the best answer that anyone can give. 

For me, the more interesting question is as follows: what happens to prosperity when productivity gains are no longer distributed through wages?

In industrial society, the wage was the central channel of distribution. If productivity rose, wages rose with it. This was not necessarily a smooth process – as the existence of unions, collective agreements and labour law testifies. But it was a structural precondition for growth: as Henry Ford understood, the workers who built the cars also needed to be able to afford to buy one.

There has been considerable debate about an apparent recent gap in the US between wages and productivity growth: this appears to be widening, although by an uncertain amount. There is also rather clearer evidence that the share of labour income in national income has declined in nearly all industrial industrialised economies since the 1980s, for multiple possible reasons - i.e. long before the advent of AI.  The worry therefore is that the impact of AI could further undermined an already unstable distributive system.  

This question of distribution has multiple dimensions. It goes well beyond changes to the aggregate share of labour income share of national income or, indeed, the distribution of income within individual labour markets (e.g. across income levels or regions). A discussion of many broader distributive issues (e.g. resource costs, taxation, local vs. central governments) is also important if we are to align the costs of AI with its potential benefits.

I mentioned above the growing questioning of whether we can afford the environmental and resource costs of AI, for example its energy and water demands. It is interesting how much the debate around this is being conducted at a local or regional level, for example local US moratoriums on data centre developments. There are understandable reasons for this. The traditional factory was often a distribution mechanism (for wages, local supply chains, training etc.) not merely a site of production. By contrast, the data centre creates value in one place and distributes it elsewhere. Employee numbers tend to be small, limiting the amount of purchasing power going back into local communities, and local supply chains minimal – while resource demands are high. Any tax paid may also go to national rather than local governments.

The corporate economics of AI also demand a focus on distribution. These appear to favour large, dominant players. This market concentration is encouraged by very high fixed costs, the need for network and data feedback to develop models, high required levels of computing capacity and increasing vertical integration in the industry (the same firms controlling chips, clouds, models and application).  These sorts of concentration escape existing antitrust legislation (focused on market shares or sector pricing impact). 

For an appropriate policy response, past economists are worth revisiting. German economist Walter Eucken (1891-1950) , for example, provides a distinctive view on the role of the state in managing corporate activity. Not an enthusiast for day-to-day state intervention, he did however think that the state should have rules-based policy to limit private concentrations of power and make sure that competition survived. (As he realised, the successful competitor eventually often has no competitor left.)  His observations seem particularly appropriate in the global tech sector today.

I remain at heart an optimist and think that it will be possible to reconcile and even perhaps benefit from some of AI’s contradictions (e.g. between its resource use and its ability to improve resource efficiency, as I have discussed in a previous post). I also think that AI will be a development where ethics, democratic values and social literacy will be useful skills for individuals. But, when we look at this question of distribution, I think there is an urgent need also for a new accounting framework, to enable us to weigh up AI’s benefits and failings, and to determine who should get what.

At this point, an interesting overlap emerges between our responses to AI and those to environmental issues: this is the relationship between capital productivity and natural capital. Our data centre is essentially a converter. It converts electricity into computation and computation into heat. It needs energy, it needs cooling, and cooling in many designs needs water. It stands where energy is cheap and energy is often cheap where water is scarce.

We have to weigh up the potential future returns from the computation (e.g. better climate models, more efficient grids, more precise agriculture, faster materials research) against the costs (e.g. energy, water, land, materials). The gains are at some point in the future and are hard to attribute with any precision; the costs are immediate, and measurable.

And our current accounting systems cannot handle this. The consumption appears on balance sheets as a cost item. The return appears not at all, because it belongs to nobody. As long as natural capital does not appear on balance sheets, every claim about AI’s net contribution to sustainability is simply an unsubstantiated assertion.

So, if we are to account (and respond to) the costs and returns of the AI transformation - and how they are distributed across society – we need a better balance sheet. 

At present, our dominant measures of prosperity are poorly equipped to provide this. GDP remains a flow measure. It records what passes through the economy in a given period, but it tells us little about the condition of the underlying stocks of capital (produced, human, natural and institutional) on which future prosperity depends.

Existing accounting systems are already good at measuring produced capital (e.g. machines, buildings, infrastructure): we have numerous well-established systems for developing these systems, for example the International Financial Reporting Standards (IFRS). We have started the process to calculate it for natural capital (soils, water, forests, oceans, biodiversity) with both national approaches, for example from the UK’s Office for National Statistics, complementing international frameworks such as the System of Environmental-Economic Accounting was well as OECD and other guidance. We are possibly less advanced in developing a unified system for human capital (skills, education, experience) despite this being a commonly used concept in economics for over half a century (Gary Becker’s “Human Capital Theories”, the most well-known foundation work, was published back in 1964).  Current methods include cost, lifetime income and indicators-based approaches, with the UN establishing a task force on measuring human capital. As regards institutional capital (legal certainty, functioning administration, the reliability of contracts, a shared factual basis) analytical frameworks are currently proposed in academic studies but largely without multilateral institution backing (despite this form of capital underpinning the whole social edifice).  Creating a coherent new balance sheet assessment, incorporating all these forms of capital, will be difficult. Previously, for productive and natural capital, development of measurement methods has been done by a mixture of international, corporate and national government bodies. All will be needed here, but I think we should try to avoid getting bogged down in a very lengthy process: the key point is to identify and anticipate the key potential issues around AI.

AI development may itself force the balance sheet construction issue by making existing failures in capital measurement ever more apparent. If a country builds more data centres, consumes more electricity, expands grid infrastructure and generates more corporate revenue, GDP may rise. But if this process also increases pressure on water systems, accelerates land-use conflicts, weakens local communities, concentrates corporate power or erodes institutional trust, the country may not have become wealthier in any meaningful sense.

Distributional changes resulting from AI will be important for all these sorts of capital. We are already well aware that AI will change the relative value of different sorts of human capital, with implications for wages - the traditional distribution mechanism of industrial society.  In the case of natural capital, many aspects of AI will have resource costs (e.g. data centres, discussed above) but its results may also ultimately have resource benefits).  In terms of institutional capital, AI can strengthen it (e.g. through facilitating better public services) and weaken it (e.g. misinformation, privacy erosion).

So, while we do not need to predict exactly what AI will do in twenty years, we do need to ask what it is likely to do to the overall balance sheet of society. How are productivity gains being distributed? What will be the net impact of AI on natural resources? Is AI contributing or detracting from how our institutions function?

A new balance sheet approach would help us understand what is happening and how the returns and costs from AI should be distributed. We cannot predict the future, but we can at least start discussing ways to deal with its possible implications.

Markus H.-P. Müller’s research focuses mainly on the structural transformation of economies and societies and on sustainability. He has authored several books and articles on the transformation of society and economies and is a columnist for the Global Policy Journal at Durham University. Markus is Global Head of the Chief Investment Office of Deutsche Bank Private Bank and in June 2022 he also took on the role of Chief Investment Officer Sustainability.

Photo by beytlik from Pexels