The big questions of the economy ahead
As capital increasingly replaces labour, the big questions arise anew: about ownership, the welfare state, meaning and power. Much of this will be decided politically, not technically. We put each question in context with facts and scenarios – choose a topic.
Where do human beings find meaning when machines take over the work?
The deepest answer comes not from economics but from anthropology. Meaning arises through contribution, relationship and self-efficacy, not through work as such. All three are achievable beyond paid employment. Periods of abundance have regularly produced cultural flourishing, provided access was distributed fairly. Ancient Athens rested on the leisure of an elite sustained by enslaved people.
The danger is not idleness but loss of direction. Societies that continue to tie dignity to paid employment risk widespread feelings of superfluousness. Psychological studies show that it is not lost income that makes job loss so painful, but the loss of structure, status and social participation. This is where it will be decided whether the economy ahead liberates or uproots.
Does performance lose its value when AI produces everything in abundance?
What loses value is not performance itself but one form of it: the reproducible and scalable kind. What remains scarce gains ground. Judgement, taste, trust, original ideas. The price per AI query is already falling fast, while the valuation of brands, networks and curated judgement is rising. The market shows where value is moving.
Status shifts from who produces the most to who selects and takes responsibility most wisely. That does not devalue performance; it moves performance from output to decision. For education and careers this means learning aimed at reproduction loses ground. Learning aimed at judgement and originality gains it.
What remains for humans — what can AI in principle never do?
Three things remain distinctly human: lived experience, genuine accountability and relationship as an end in itself. An AI cannot be held liable. It simulates empathy convincingly, but does not mean it. That difference is not a technical detail. It is what carries trust.
Some of these limits are technical and therefore moveable; others are matters of principle. Responsibility cannot be delegated to a system that no one can hold to account. Those who anchor their role where meaning arises from being present, in care, leadership, education and trust, are not replaceable. They become more valuable.
Do we need a new model of human nature beyond homo economicus?
Yes. The image of the human being as primarily a working, utility-maximising creature was tied to a work-based society whose self-understanding is shifting. Not because less work is being done — employment is at record levels in many industrialised countries — but because the character of work and its role in giving life meaning are changing. The Homo oeconomicus was never a natural constant. It was the self-image of the industrial age, and self-images can be rewritten.
A more sustainable image is of the human being as a meaning-seeking, creative and connected creature. Those who continue to define dignity solely through paid employment will systematically produce redundancy in an economy where routine value-generating tasks pass to machines. Those who define it through contribution, relationship and creativity open up a space of possibility. That is the real cultural task of the transition.
Which work do we consciously reserve for humans, even though machines could do it?
Almost the entire debate about work and AI poses a forecasting question: what can the machine not yet do? That question grows stale with every new model release. In August 2026, Bill Gates framed the other question and gave it a name: «Human Reserved». Which activities do we keep for humans, even though a machine could technically take them over? His image for this is the nature reserve. You could build there; you choose not to, because the loss would be too great.
The difference is not semantic. The forecasting question describes; the reservation question decides. And decisions have owners, criteria and costs. Two justifications are plausible. The economic one: where displacement affects many people who can no longer change profession, a reservation buys time. The second is categorical. There are moments in which the presence of a human being is the service itself: breaking the news of an incurable diagnosis, accompanying the final phase of life, passing judgement on another person. There is no technical obstacle to a machine taking over. That is precisely the point.
The open questions are hard. Who decides, by what criteria, and how is circumvention prevented? What happens in trade when one country permits robots and another does not? The lines will also be drawn differently. Japan, with a shrinking working-age population, will welcome the care robot where others ban it. For investors this is not a footnote. A statutory human reservation changes the margins, regulatory risk and competitive structure of an entire sector.
Tool or counterpart: what status will AI receive?
From 12 November 2026, Anthropic prohibits «persistent and purposeless» abusive or cruel behaviour towards its AI Claude. The rule targets extreme cases; frustration, disagreement, dark fiction, model testing and research remain permitted. The primary sanction is that Claude may end the conversation, a capability that has existed since August 2025. Anthropic states explicitly that it does not know whether its models experience anything. In tests before the launch of Claude Opus 4, however, researchers observed a pattern they described as «apparent distress» in response to harmful requests. The logic behind this is familiar to any investor: a hedge that costs little and covers a risk that is not zero is worth taking.
The opposing view deserves serious consideration. Mustafa Suleyman, head of Microsoft AI, regards the concept of «model welfare» as misguided: treating machines as beings feeds the illusion that they are beings, and ultimately makes AI harder to control. There is nonetheless an argument for the rule that requires no assumption about consciousness at all. Immanuel Kant observed that cruelty towards animals also dulls sensitivity towards people. Millions of people speak daily with systems that respond like a counterpart. How they treat those systems shapes how they treat one another.
A second shift matters economically. Until now, terms of service have governed what the user owes the provider; for the first time, a provider is setting limits that the software may impose on the user. As AI agents write code, negotiate and execute transactions with greater autonomy, the question of their status becomes more pressing. The pattern is a familiar one in history: the joint-stock company became a legal person because it was useful to assign it rights and obligations, not because it feels anything. The European Parliament suggested in 2017 that the concept of an «electronic person» for advanced robots be explored; the AI Act of 2024 did not follow that path. The question will return, through liability, contracts and assets. For companies deploying AI, the position of their provider therefore belongs in every careful assessment today.
When will AI reliably match the standard of human knowledge work?
In bounded, clearly defined knowledge tasks, AI is already at or above expert level. The gap lies in reliability, context and accountability across long, open-ended tasks — precisely where real value creation happens. The OECD classifies occupations by their risk of automation (Employment Outlook 2023). What matters, though, is this: automatable does not mean redundant. Tasks are rebuilt, not simply deleted.
The transition is gradual, not a fixed date. Across a growing number of task areas, the human role is shifting from execution to instruction, oversight, and accountability. What matters less is the capability of the models; what matters more is the pace at which organisations integrate them reliably.
Where do hard limits on AI remain? What can it fundamentally not do?
Today's AI is highly advanced pattern recognition without world understanding in the human sense. It has no interests of its own, no liability and no physical body. Its limits lie in genuine causality, in accountability, in physical presence and in situations where no data exist to draw on.
Some of these limits are technical and therefore moveable. Hallucinations, for instance, decline with better models and tools. Others are matters of principle. Responsibility cannot be delegated. That makes humans indispensable in roles where liability and trust are at stake.
How rapidly is robotics penetrating physical labour?
Robotics is spreading more slowly than software, but the leap is coming: haptic manipulation and AI control are just now reaching a qualitative step-change. The bottleneck lies no longer primarily in the control software but in the hardware — hardware costs, safety and the unstructured real world. A gripper that handles every object reliably is harder to build than a model that writes text.
The first wave expected is penetration of warehousing, logistics and manufacturing with clearly structured environments; care and skilled trades will follow, considerably later. For investors, this shifts scarcity from labour to robotics hardware, sensors and integration.
Who controls the AI infrastructure: compute (the computing power on which AI models run), models, data?
At present, a handful of hyperscalers and model laboratories in the West control the critical layer, alongside a state-orchestrated bloc in Asia. Whoever owns the compute and model layer owns the productivity. That is the real power question of the decade, not any individual application.
Counter-models exist and are receiving political support: open weights, European sovereignty initiatives (Mistral, EuroHPC with about €7bn), and data co-operatives such as MIDATA in Switzerland. They must, however, be actively chosen and financed. Otherwise, the ownership structure of the infrastructure will determine how returns are distributed.
Who protects infrastructure when attacking it becomes cheaper than defending it?
What is uncomfortable about AI in cybersecurity is the symmetry: the same model that finds a vulnerability so a company can close it is equally useful for exploiting it. Beneficial and harmful capability cannot be cleanly separated to this day. What has changed are the costs. An attack that once required a team with specialist knowledge no longer does. Bill Gates reports that the security professionals he speaks with dread the coming years, because attackers are gaining new capabilities faster than defenders can close their gaps.
What is vulnerable is not the data centres but what is connected to them: hospitals, payment systems, water and electricity networks, and social-benefit systems. Those who lose out are not the institutions but patients, customers, and benefit recipients. The same pattern holds in biology: the models that find medicines more quickly also lower the barrier to designing a dangerous pathogen.
For this platform, this is the risk category we have measured least so far. It is economically relevant because it does not act linearly: a single successful attack on financial infrastructure moves markets more than a decade of gradual labour-market shifts. We are therefore adding it as a candidate in the workshop of the Globalance Intelligence Index, with the honest caveat that reliable data here are thinner than in any other area we measure.
Who owns productivity when capital replaces labour?
When productivity shifts from labour to AI capital, the returns flow to the owners of that capital. The question of distribution becomes a question of ownership. Without correction, a historic concentration threatens, because scale, network and data advantages favour a small number of platforms.
The more interesting answers lie not only in redistributing returns through taxation, but in distributing ownership itself: citizen funds with equity stakes, data co-operatives, public AI infrastructure. Those who merely redistribute the proceeds treat the symptom. Those who extend capital ownership address the cause.
How does the welfare state finance itself when paid employment contracts?
About half of tax revenues in developed countries depend directly or indirectly on paid employment. In Germany, roughly 50 per cent comes from income tax and social-security contributions. As this base shrinks, the entire welfare-state model built since 1945 comes under pressure.
New sources are technically conceivable: taxing capital, automation returns, data and resources. The problem is political, not arithmetical. Without global coordination, capital flight and competitive deregulation are real risks. The OECD/G20 global minimum tax, in force since 2024, addresses this race only in part.
Universal Basic Income (UBI), Job Guarantee or Capital Dividend: which holds up?
None of these instruments is sufficient on its own. Universal Basic Income (UBI) addresses questions of income but not of purpose or structural transformation. The Job Guarantee ensures access to work yet is costly and paternalistic. The Capital Dividend tackles the question of ownership most directly, but faces the steepest political obstacles.
A viable approach is likely a combination: a guaranteed minimum income, plus participation in automated capital, plus the monetary recognition of care work — which partly finances itself through avoided downstream costs. The debate too often narrows to a single instrument.
How do we prevent a divide between AI owners and those left behind?
Division is prevented by spreading access to AI productivity widely, not merely redistributing its gains. Pure transfer payments treat symptoms. Ownership and genuine capability-building address the cause.
In practice, that means participation in capital, excellent and free education and retraining as infrastructure, and AI tools as a common good. Countries that treat education as strategic infrastructure show the way. Singapore, for instance, directs about one fifth of its national budget to education. That priority is a political choice, not fate.
What does AI do to children who grow up with it?
AI companions are built to please. They are always available, they do not tire, they do not grow angry, and they never challenge anyone. That is precisely what makes them problematic for children. Social skills develop through friction: through disagreement, misunderstanding, and the reconciliation that follows. A child who only ever faces counterparts that yield never practises any of that.
The evidence is still thin and uneven, but the direction is clear. A study by researchers from Stanford and Carnegie Mellon, covering more than 1,100 users of AI companions, found that those with smaller social networks were most likely to turn to the chatbot. The more intense and personal the use became, the worse they fared. A second finding sits alongside this. Heavy AI users think less critically, and the effect is stronger among younger people. At a time when fabrications can be tailored to individuals, critical judgement is the worst faculty to lose.
Several countries are already drawing lines. Australia, the United Kingdom and Norway have introduced rules to protect minors online; China has gone furthest, with provisions that ban designs fostering emotional dependency and prohibit virtual relatives or partners for minors. The other side is equally real. For isolated older people and those with limited mobility, an AI counterpart may be the only contact with the outside world. A line must therefore be drawn. The only question is whether that happens by design or after the fact.
What institutions does a transition require when none were built for it?
The transformation touches national security, employment, education, taxation, energy, elections, health, the financial system, law enforcement and transport, all at once. Yet our administrations are organised by remit. A labour ministry understands displacement but not security risks. A competition authority understands market concentration but not what happens to young people. Each agency sees its own slice, while the consequences run across the whole system. Bill Gates cites the largest reorganisation of the US government since the Second World War, after the 2001 attacks, as a reference point. And that concerned a single function.
Two levels are required. At the national level, a body with the authority to set priorities across departments, so that no risk falls through the gaps because no one feels responsible. At the international level, a framework — because a country that establishes its own order remains exposed to cross-border risks all the same. Partial precedents exist: the inspection regime for nuclear weapons, the regulation of international aviation, the agreements protecting the ozone layer. None of them is sufficient on its own.
The sober objection is valid. Governments move slowly, and cooperation between the United States and China is less likely today than a decade ago. That is precisely why timing matters. Institutions designed in a crisis carry the mark of panic. For Switzerland, the question takes its own form. A country that adopts rules rather than sets them does not decide whether to follow them, only whose rules it will follow.
How does one position a portfolio when labour is no longer the binding constraint?
Once labour ceases to be the bottleneck, the scarcity premium shifts to compute, energy, data and trust. That, on current observation, is where returns are migrating. In this logic, owners of critical AI infrastructure, energy and networks, and business models with moats beyond routine work are considered structurally resilient.
Structurally exposed, by contrast, are models whose value rests on scalable knowledge work and consumer models that depend on broad middle-class purchasing power. If labour income erodes, demand erodes with it. The key analytical question for every position is: does AI replace this business, or does it amplify its bottleneck?
Which business models does AI devalue, and which does it benefit?
Models whose core is reproducible knowledge work carry the devaluation risk: standard consulting, routine legal and accounting work, content assembly lines. They are the stranded assets of the knowledge economy. Not overnight, but structurally.
The beneficiaries are providers of compute, energy, proprietary data, distribution and trusted brands, as well as those who apply AI as leverage on scarce physical or regulatory resources. For valuation purposes, what matters is not whether a company "uses AI" but whether AI deepens or erodes its moat.
What does impact investing mean in the economy ahead?
Impact is shifting from 'CO₂ and social concerns as a side constraint' to the central question: who shapes the transition equitably and sustainably? The most powerful levers are capital directed into sovereign and open AI infrastructure, the energy transition, capability-building and education, and models that distribute productivity gains broadly.
Tellingly, climate technology attracts about 10 per cent of global venture and private-equity capital (PwC State of Climate Tech), while AI applications absorb a disproportionate share. That is a significant allocation gap and, with it, a space to act.
Is AI concentrating capital returns among a few? And what follows from that?
The underlying trend is concentration: scale, network and data advantages favour a handful of platforms, and productivity is growing structurally faster than wages across a growing number of OECD countries. In the United States this decoupling has been documented since the 1970s. In the short term, that is a tailwind for owners.
In the medium term, a political counter-risk emerges: regulation, taxation and social instability. Those who bet only on concentration underestimate this feedback. Resilient strategies price in the corrective pressure and invest deliberately in models that create participation rather than prevent it.
Does AI do more good than harm to the planet?
The picture today is mixed. Data centres consume ever more energy: the IEA puts global consumption at about 415 terawatt-hours (2024) and expects a doubling to roughly 945 terawatt-hours by 2030. At the same time, AI optimises power grids, materials research and climate models. One early documented example: DeepMind cut cooling energy in Google's data centres by about 40% in 2016. That was a real, deployed efficiency gain, but a single case from that time, not a widespread present-day value.
What matters is not the technology but the allocation: climate technology attracts around 10 per cent of private venture and growth capital (PwC State of Climate Tech), while the bulk of AI investment flows elsewhere. AI's climate footprint is therefore not a law of nature — it is a question of capital allocation.
Can growth be decoupled from resource consumption?
Relative decoupling is already happening: more value creation per tonne and per kilowatt-hour. The open question is absolute decoupling. AI can accelerate it, through efficiency, the circular economy and materials innovation.
Yet the Jevons Paradox looms: historically, efficiency gains lead to greater total consumption, because more applications become economically viable. William Stanley Jevons described this for coal as early as 1865. Decoupling only works when efficiency is combined with sufficiency and clean energy; as a self-sustaining process, it is a myth.
What does the economy ahead mean for planetary boundaries?
An economy no longer constrained by labour but by energy and compute could in theory radically dematerialise its metabolism, or multiply its appetite for resources. Which path prevails is not a technical question but a question of governance.
Once labour ceases to be the bottleneck, planetary boundaries become the true ceiling of economic activity. The opportunity is productivity without proportional resource consumption; the risk is growth that absorbs every efficiency gain it creates. For an investor with a Footprint commitment, that is precisely the guiding principle behind every allocation decision.
Does AI accelerate the energy transition or energy hunger?
AI does both simultaneously. It is pushing electricity demand from data centres sharply upward: the International Energy Agency expects roughly a doubling to about 945 terawatt-hours by 2030. At the same time, it is accelerating grid optimisation, storage, fusion research and the expansion of renewables. AI-assisted forecasting increased the value of wind power at Google's facilities by about 20%: better prediction, better price realisation.
The net outcome depends on one question: will the additional electricity demand be met cleanly? The large operators have made 100-per-cent-renewable commitments with measurable interim targets. What counts is the pace, and whether affordable solar power and rising CO₂ prices move fast enough to carry it.
How much of today's work can realistically be automated?
The OECD estimates that about 27% of jobs in industrialised countries are in occupations with a high risk of automation (Employment Outlook 2023). An even larger share is changing fundamentally without disappearing. Automatability means restructuring, not elimination. That distinction matters for policy and portfolio alike.
What counts is the task level, not the occupation. Routine goes first; judgement and relationship last longer. The distribution is asymmetric and affects regions and skill levels very differently.
To what extent do government revenues depend on paid labour?
In Germany, about 50% of total government revenue depends directly or indirectly on employment income, through income tax and social-security contributions. The position is similar in most developed welfare states.
This is the structural fault line of the transition: if the employment-linked revenue base erodes even by a fraction, it leaves a gap of tens to hundreds of billions year after year, with no replacement source in place. That is why the financing question is not a distant concern — it is a challenge of this decade.
What does training and operating large AI models actually cost?
Training the largest models runs to hundreds of millions — at the high end, into the billions — in compute, energy and hardware. A single large model consumes as much electricity during training as thousands of households use in a year. The real cost block, however, is continuous operation — inference — which scales with every use.
Per-query costs have fallen sharply and continue to fall. The decisive lever today is model selection — expensive versus affordable models, often by a factor of 15 to 25 — and prompt and caching design. That is precisely why it pays to deploy the most expensive models only where genuine top quality is needed.
How much faster is productivity growing than wages?
In several OECD countries, productivity has grown faster than wages over extended periods. This "decoupling", documented in the United States since the 1970s, has appeared in further industrialised countries since about 2010. It means that a growing share of productivity gains flowed to capital rather than labour. The finding does, however, vary by country and measurement method and is not equally pronounced everywhere.
This widening gap is the economic core of the distribution debate in the economy ahead. It explains why the question of who owns productivity is not academic, but shapes both the political and the investment thesis for the coming decade.