Artificial Intelligence, Labour Income and Effective Demand: A Theoretical Framework for the Distributional Absorption Threshold of AI-Induced Productivity

Abstract

Artificial intelligence (AI) may raise productivity by automating tasks, augmenting human work and reducing information-processing costs. Yet productivity gains are not necessarily converted into broadly shared purchasing capacity or fully absorbed output. This conceptual review develops the notion of the Distributional Absorption Threshold of AI-Induced Productivity, defined as the point at which AI-related productivity growth outpaces the growth of broadly distributed real purchasing power and household consumption. The framework links AI-induced productivity to labour income, income distribution, prices, investment, fiscal redistribution, external demand and effective demand. It distinguishes a favourable transmission path, in which productivity gains support wages, disposable income, consumption and output absorption, from a critical path, in which weak distributive transmission may generate absorption tension. The review formulates conceptual propositions and preliminary operational indicators for future empirical research while treating the threshold as an analytical construct rather than a fixed empirical constant. Its contribution is theoretical: it reframes the AI productivity debate beyond both technological optimism and automation anxiety by connecting technological change, distribution and demand-side realisation.

1. Introduction

Artificial intelligence has become one of the central technologies through which contemporary economies seek to increase productivity, reorganise work and strengthen competitiveness. Recent advances in machine learning and generative AI have intensified this debate because AI systems are no longer confined to narrowly defined routine processes. They increasingly affect prediction, language, coding, design, customer interaction, professional services and knowledge-intensive tasks. For this reason, AI is often discussed as a potentially general-purpose technology whose economic effects depend on complementary innovation, organisational adaptation, skills, data infrastructure and institutional conditions [1,2,3].

The productivity potential of AI is increasingly documented at the level of specific tasks, workers and firms. Experimental and field evidence suggests that generative AI can reduce task completion time, improve output quality and support productivity in activities such as writing, customer support and professional consulting [4,5,6]. At the same time, these effects appear to be heterogeneous. They depend on task characteristics, worker experience, organisational context, capability boundaries and the ability of firms to redesign workflows around AI tools. AI should therefore be approached as a technology with significant productivity potential, but not as a mechanical source of immediate and uniform macroeconomic productivity growth.

A large part of the public debate on AI has focused on the possibility of job displacement. This concern is legitimate, but it is not sufficient for understanding the macroeconomic implications of AI-induced productivity growth. The task-based literature has shown that technological change may substitute some tasks, complement others and create or reinstate new forms of work [7,8,9,10]. AI should therefore not be interpreted only through the lens of technological unemployment. Employment displacement is one possible channel, but productivity gains may also affect labour income through slower wage growth, reduced hours worked, changing bargaining power, labour-market polarisation, declining labour share or the concentration of gains in profits and capital income.

This article starts from a broader macroeconomic question: even if AI increases productivity, under what conditions are the resulting gains transformed into broadly distributed real purchasing power and effective demand?

The question matters because productivity growth expands the capacity to produce, but it does not automatically ensure that additional output will be absorbed by the market. The economic realisation of productivity gains depends on demand channels, including household consumption, investment, public expenditure and external demand. It also depends on the distribution of income between labour and capital, the extent of redistribution, price dynamics and the marginal propensity to consume across household groups.

To capture this problem, this conceptual review develops the notion of the Distributional Absorption Threshold of AI-Induced Productivity. The concept refers to the point beyond which productivity gains associated with the adoption or use of AI are no longer accompanied by proportionate increases in broadly distributed real purchasing power and household consumption. In simplified terms, the threshold becomes relevant when the rate of AI-induced productivity growth persistently exceeds the rate of growth in broadly distributed real purchasing power. The concept was initially formulated as a definitional contribution [11], and the present review develops it into a wider theoretical framework.

The contribution of the review is threefold. First, it reframes the debate on AI and productivity by shifting attention from technological efficiency alone to the distributive and demand-side realisation of productivity gains. This contribution does not replace existing theories of effective demand, underconsumption, wage-led growth or distribution-sensitive growth. Rather, it applies and extends their central insight to the specific case of AI-induced productivity by asking how technology-related productivity gains are transmitted into labour income, household purchasing capacity and demand-generating expenditure. In this sense, the novelty of the framework lies not in proposing a new standalone productivity-pay indicator, but in specifying an AI-specific absorption mechanism that links productivity gains, distributive transmission and demand-side realisation. Second, the review distinguishes between a favourable transmission path, in which AI-induced productivity gains support wages, purchasing power, consumption and output absorption, and a critical transmission path, in which productivity gains are weakly transmitted to household income and effective demand. Third, it proposes a conceptual basis for future empirical operationalisation through indicators such as the productivity-real labour income gap and the absorption tension indicator, while recognising that such indicators are preliminary and require further empirical validation.

This review is theoretical and conceptual in nature. It follows a conceptual review approach rather than a systematic review protocol. The literature is used to develop, position and delimit a theoretical framework, not to provide an exhaustive mapping, bibliometric analysis or quantitative synthesis of all studies on AI, productivity and labour-market outcomes. The review draws selectively on bodies of literature directly relevant to the proposed construct, including research on general-purpose technologies and AI-related productivity, task-based approaches to technological change, labour income and functional income distribution, and Keynesian and Kaleckian perspectives on effective demand and distribution-sensitive growth. Empirical studies are considered where they clarify mechanisms relevant to the framework, including task-level productivity effects, occupational exposure to AI, wage and employment channels, and possible indicators for future empirical operationalisation.

The literature was selected purposively according to its conceptual relevance to the framework, rather than through an exhaustive database-screening procedure. Sources were included when they contributed directly to one or more of the following analytical components: AI as a productivity-enhancing technology, task-based technological change and AI exposure, labour income and functional income distribution, household consumption and effective demand, or the possible operationalisation of AI adoption, productivity, income and demand indicators. Priority was given to peer-reviewed academic contributions, established theoretical works and institutional sources that provide comparable indicators or policy-relevant evidence. Studies were not included merely because they addressed AI in general, but because they clarified mechanisms linking AI-related productivity to labour income, purchasing power, consumption or demand-side absorption. The literature was analysed interpretively and synthetically in order to define concepts, identify transmission channels, formulate propositions and delimit possible empirical indicators, rather than to produce a bibliometric mapping, frequency count or meta-analytic estimate.

Accordingly, the manuscript does not claim to offer a systematic literature review, a meta-analysis or an empirical test of the proposed threshold. For this reason, no PRISMA flow diagram, formal search protocol or inclusion-exclusion procedure is reported. This is a limitation of scope rather than a claim of exhaustive coverage. The article does not claim that a distributional absorption threshold has already been reached in any specific economy, nor does it assume that AI necessarily produces mass unemployment or demand weakness. Its purpose is to clarify concepts, integrate mechanisms and formulate propositions that future empirical research may examine.

The review is structured as follows. Section 2 discusses AI as a productivity-enhancing technology and distinguishes between productive capacity and economic realisation. Section 3 examines AI, labour tasks and income distribution. Section 4 develops the link between labour income, household consumption and effective demand. Section 5 defines the Distributional Absorption Threshold of AI-Induced Productivity and presents the favourable and critical transmission paths. Section 6 formulates the conceptual propositions of the article. Section 7 discusses possible indicators and research designs for future empirical operationalisation. Section 8 presents the theoretical and policy implications of the framework, and Section 9 concludes by identifying future research directions.

2. Artificial Intelligence and Productivity: From Technological Capacity to Economic Realisation

2.1. Artificial Intelligence as a General-Purpose and Productivity-Enhancing Technology

In this review, artificial intelligence technologies are understood broadly as machine-learning and data-driven systems that can perform or support tasks involving prediction, classification, optimisation, pattern recognition, natural-language processing, content generation, decision support and human–machine interaction. This includes, but is not limited to, generative AI systems, predictive analytics, machine-learning applications, recommendation systems, computer vision and AI-assisted decision tools. The term does not refer to a single homogeneous technology but to a family of digital technologies whose productivity effects depend on the tasks affected, the sector of application, the quality of implementation and the complementary organisational and human capabilities available.

From this perspective, AI can be analysed as a general-purpose and productivity-enhancing technology whose economic significance depends on wide applicability, continuous improvement and complementarities with other innovations [1,2,3]. AI is therefore not merely another digital tool. It may become a productivity-enhancing infrastructure that changes how firms process information, allocate resources, organise labour, interact with consumers and generate value.

The productivity potential of AI operates through several channels. AI can automate routine, repetitive or codifiable tasks, thereby reducing the labour time required for specific activities. It can also augment human work by supporting drafting, programming, diagnosis, translation, customer interaction, managerial analysis and knowledge-intensive decision-making. In addition, AI can reduce prediction and information-processing costs, which may improve the speed and quality of decisions in sectors where uncertainty and data complexity are central. Finally, AI may enable new products, services and business models that were not feasible under previous technological conditions. These mechanisms suggest that AI may influence productivity both by saving labour in some tasks and by complementing labour in others.

However, the relationship between AI and productivity should not be interpreted mechanically. The history of general-purpose technologies shows that productivity gains are rarely immediate, uniform or automatically visible in aggregate indicators. They usually require complementary investments in intangible capital, human skills, organisational redesign, managerial capabilities, data infrastructure and institutional adaptation [2]. This is particularly relevant for AI because its value depends not only on access to algorithms but also on the capacity of firms and institutions to integrate AI into routines, workflows and decision processes. A firm may adopt AI tools without being able to reorganise production around them. Conversely, firms with stronger data systems, complementary skills and better organisational capabilities may obtain larger productivity gains from similar technologies.

For this reason, it is necessary to distinguish between AI as a source of technological potential and AI as a source of realised productivity growth. Adoption alone is not sufficient. Productivity effects depend on the scale of use, the nature of the tasks affected, the quality of implementation, the capacity to redesign work and the broader institutional environment. This distinction is important for the article’s theoretical argument because the distributional absorption problem cannot be understood if AI-induced productivity is treated as an automatic macroeconomic outcome.

2.2. Evidence on AI-Related Productivity Gains and Their Heterogeneity

Recent evidence on generative AI confirms that AI can improve performance in specific tasks while also showing that its effects are heterogeneous across workers, tasks and organisational contexts. Noy and Zhang [4] show experimentally that generative AI can increase productivity in professional writing tasks by reducing completion time and improving output quality. Brynjolfsson et al. [5] find that a generative AI assistant increased productivity among customer-support agents, with stronger effects for less experienced workers. Dell’Acqua et al. [6] show that AI can improve productivity and quality in knowledge-intensive consulting tasks, while also emphasising that performance depends on whether tasks fall within or outside the technology’s effective capability frontier.

These findings provide concrete evidence that AI may raise task-level productivity. They do not, however, justify the conclusion that aggregate productivity growth will automatically accelerate across the whole economy. Task-level improvements do not necessarily translate into proportional macroeconomic gains. Aggregation depends on the diffusion of AI, the share of affected tasks in total production, complementarities with human labour, the capacity of firms to reorganise production and the distribution of AI adoption across sectors. AI may generate large productivity gains in some firms and limited effects in others, producing an uneven productivity landscape rather than a uniform technological acceleration.

This heterogeneity has also been recognised in broader policy-oriented analyses. Filippucci et al. [12] argue that AI may support productivity growth while also emphasising uneven adoption, concentration of AI capabilities, complementarities with skills and possible distributional implications. Acemoglu [13] similarly cautions against assuming very large aggregate productivity effects without examining task exposure, cost savings, automation, complementarities and diffusion. These contributions support a prudent interpretation: AI has substantial productivity potential, but the magnitude, timing and distribution of its macroeconomic effects remain conditional.

In this framework, sectoral and occupational differentiation is therefore treated as a central source of heterogeneity, rather than as a directly estimated empirical effect. AI-related productivity effects are expected to differ across industries and professions because tasks, data availability, regulatory constraints, complementarities with human labour and organisational capabilities differ substantially. The framework does not assume that AI produces a uniform productivity effect across the economy. Instead, it treats AI-induced productivity as a conditional and uneven process whose macroeconomic relevance depends on the sectors and tasks affected, the scale of adoption and the extent to which productivity gains are transmitted into income and demand.

The implication for the present article is that AI-induced productivity should be treated as a conditional process rather than as a guaranteed outcome. This is not a limitation of the theoretical framework, but one of its starting premises. If productivity gains are heterogeneous and institutionally mediated, then their distribution becomes central. A productivity gain captured by a small number of firms, capital owners or high-income groups has different macroeconomic implications from a productivity gain transmitted broadly through wages, prices, employment income or public revenues.

2.3. From Productive Capacity to Economic Realisation

The key distinction for this article is the distinction between productive capacity and economic realisation. AI may increase the capacity of firms and economies to produce more output with the same or fewer inputs. Yet a rise in productive capacity does not automatically imply that the additional output will be realised through market demand. In macroeconomic terms, the capacity to produce and the capacity of the market to absorb what is produced are related, but they are not identical.

The term “AI-induced productivity” is used in this article in a conceptual rather than directly measured sense. It refers to productivity gains that are plausibly associated with the adoption, exposure or organisational integration of AI technologies, including both direct effects on task performance and indirect effects through workflow redesign, decision quality, cost reduction, innovation and new business models. The article recognises, however, that isolating the specific contribution of AI from other sources of productivity growth is empirically difficult. Measured productivity may also reflect capital deepening, non-AI digitalisation, managerial change, market conditions, sectoral composition, regulation and macroeconomic cycles. For this reason, the framework treats AI-induced productivity as an analytical category that future empirical studies would need to identify through careful research design, rather than as an effect that can be inferred mechanically from aggregate productivity indicators.

Much of the current debate on AI remains strongly supply-side. AI is often examined through its expected effects on productivity, efficiency, automation, innovation, costs and output per worker. These dimensions are essential, but they do not exhaust the macroeconomic question. Even when productivity rises, the resulting output can be fully realised only if there is sufficient effective demand. Demand may come from household consumption, private investment, public expenditure or external markets. If these channels do not expand in line with productive capacity, part of the productivity potential may remain weakly absorbed.

The demand-side dimension becomes especially important when productivity gains are unevenly distributed. Productivity growth can support economic realisation when it is transmitted into higher real wages, employment income, household disposable income, lower consumer prices or public revenues that finance transfers and services. In these cases, the increase in productive capacity is accompanied by an increase in purchasing power or demand-generating expenditure. By contrast, if productivity gains accrue mainly to profits, capital income or high-saving groups, while real labour income stagnates or grows more slowly, the link between productivity and household demand may weaken. This does not mean that profits or investment are unimportant. It means that the absorption of additional output depends on the composition of income flows and on the channels through which they return to demand.

This article therefore approaches AI-induced productivity as a two-stage problem. The first stage concerns whether AI increases productive efficiency. The second concerns whether the resulting gains are transmitted into the income and expenditure flows required for effective demand. The proposed concept of the Distributional Absorption Threshold of AI-Induced Productivity starts from this second stage. It asks whether, and under what conditions, the economy’s capacity to produce may expand faster than the broadly distributed purchasing power required to absorb additional output.

This formulation does not imply that AI is economically harmful by itself. Rather, it indicates that the macroeconomic significance of AI depends on the institutional and distributive context in which productivity gains are generated, allocated and transformed into demand. If compensating demand channels are sufficiently strong, weak labour income transmission may not become a macroeconomic constraint. However, if these channels are insufficient and if household purchasing capacity grows persistently more slowly than AI-induced productivity, a distributional absorption problem may emerge.

The contribution of this section is therefore to shift the analytical focus from technological capacity to economic realisation. AI may expand what the economy can produce, but the realisation of that potential depends on whether productivity gains return to the economy as effective demand. This bridge prepares the argument developed in the following sections: AI affects tasks and labour income, labour income shapes household purchasing power, and purchasing power is central to the absorption of additional output.

3. AI, Labour Tasks and Income Distribution

3.1. From Occupations to Tasks: A More Precise View of AI Exposure

A careful analysis of artificial intelligence and labour income should begin with the distinction between occupations and tasks. Occupations are bundles of tasks, and technological change rarely affects all tasks within an occupation in the same way. This distinction is central to the task-based literature, which shows that digital technologies may substitute for some activities, complement others and reshape the composition of work rather than simply eliminate entire occupations [7,8,9]. From this perspective, the relevant question is not whether an occupation is replaced by AI, but which tasks within that occupation become automatable, augmentable or reorganised.

This distinction is particularly important in the case of AI because AI systems do not map neatly onto traditional categories of manual or routine work. Earlier waves of computerisation and automation were often associated with routine cognitive and manual tasks. By contrast, recent AI systems, especially machine learning and generative AI, may affect prediction, text production, image recognition, translation, coding, customer interaction, diagnosis, legal drafting, teaching support and managerial analysis. AI exposure may therefore be significant in some non-routine cognitive occupations that were previously considered less exposed to automation [14,15,16].

Exposure, however, should not be confused with displacement. An occupation may be highly exposed to AI because many of its tasks overlap with AI capabilities, but this does not mean that the occupation will necessarily disappear. Exposure may lead to substitution, but it may also lead to augmentation, task reallocation, changes in skill requirements or new forms of human-AI complementarity. The distinction is essential for avoiding an overly deterministic interpretation of AI and work. AI exposure identifies where technological pressure or opportunity may arise; it does not, by itself, determine the labour-market outcome.

Empirical and institutional studies support this more nuanced interpretation. Felten et al. [15] develop an AI occupational exposure measure that can be applied across occupations, industries and geographical areas. Georgieff and Hyee [17] find no simple cross-country relationship between AI exposure and employment growth, although their results suggest differentiated effects depending on computer use and occupational conditions. OECD (Organisation for Economic Co-operation and Development) evidence also indicates that current AI use at work can be associated with positive outcomes for some workers, including performance and job satisfaction, while also raising concerns about job loss, work intensity, privacy and bias [18,19]. Taken together, these findings support a conditional interpretation of AI’s labour-market effects.

The task-based approach is useful for the present article because it separates three analytically distinct dimensions: technological capability, labour-market exposure and distributive outcome. Technological capability refers to what AI systems can technically perform. Labour-market exposure refers to the extent to which workers’ tasks overlap with those capabilities. The distributive outcome refers to how the gains, costs and risks associated with that exposure are allocated between workers, firms, consumers and the state. The distributional absorption framework developed in this article depends mainly on the third dimension. AI may raise productivity through exposed tasks, but the macroeconomic relevance of this increase depends on whether the resulting gains support labour income and effective demand.

3.2. Substitution, Complementarity and Task Reinstatement

AI may affect labour income through several mechanisms. The first is task substitution. When AI performs tasks previously carried out by workers, firms may reduce labour demand for those activities, lower the number of hours required or reorganise work around fewer employees. The second is task complementarity. AI may increase workers’ productivity by helping them perform tasks faster or better, potentially supporting higher wages, better job quality or new forms of professional capability if the gains are shared. The third is task reinstatement. New technologies can create new tasks, occupations and organisational functions, including roles related to AI supervision, data governance, model evaluation, human-AI coordination and ethical compliance [10].

The balance between substitution, complementarity and reinstatement is crucial for the distribution of income. If AI primarily replaces labour in existing tasks and if new labour-intensive tasks do not emerge sufficiently, labour income may weaken. If AI mainly complements workers and raises their productivity, labour income may increase, especially where workers have bargaining power, scarce skills or institutional protection. If productivity gains accrue mainly to firms, platforms, capital owners or a limited group of highly skilled workers, the distributional effects may be unequal even without large aggregate employment losses.

Unemployment should therefore not be treated as the only relevant channel. A narrow focus on job loss may obscure other forms of distributive weakening. AI-related productivity gains may fail to support broad purchasing power even if employment remains formally stable. This may occur through slower wage growth, reduced hours, weaker bargaining power, task fragmentation, increased work intensity, labour-market polarisation or a declining labour share. Conversely, AI may support labour income if productivity gains are translated into wage growth, better job design, new occupations or inclusive forms of skill upgrading.

These conclusions are consistent with the broader automation literature. Acemoglu and Restrepo [20] show that industrial robots can have negative effects on employment and wages in local labour markets, while their framework remains useful because it distinguishes displacement, productivity and reinstatement effects. Evidence on demographics and automation further suggests that technology adoption may respond to labour scarcity and institutional conditions, not only to technological opportunity [21]. Recent analyses of generative AI similarly point to heterogeneous exposure involving both substitution and complementarity [16], while OECD evidence reports both worker-level benefits and concerns related to job loss, work intensity, privacy and bias [18,19]. Taken together, these studies support analysing AI through mechanisms and institutional conditions rather than through a single predetermined employment scenario.

3.3. Labour Income, Wage Share and the Distribution of AI-Related Gains

For the theoretical framework developed in this article, the key issue is not only whether AI changes tasks but also how the productivity gains associated with those task changes are distributed. A productivity-enhancing technology can generate different income trajectories depending on bargaining power, ownership structures, wage-setting institutions, skill distribution, market concentration and public policy. AI may raise output per worker, but the income generated by this increase may be transmitted to labour, capital, consumers, the state or external markets in different proportions.

In a favourable trajectory, AI increases productivity and part of the gains is transmitted to workers and households through higher wages, employment income, reduced prices, better job quality or social and fiscal channels. In this case, productivity growth contributes to broadly distributed purchasing power and may support household consumption. In a more critical trajectory, productivity increases but real labour income grows more slowly, working time declines, wage gains are concentrated among a limited group of workers, or profits capture a disproportionate share of the gains. This second trajectory is particularly relevant for the distributional absorption framework because it creates the possibility that productive capacity expands faster than broadly distributed household purchasing power.

This argument does not imply that profits, capital accumulation or firm-level productivity gains are undesirable. They may finance investment, innovation and future growth. The problem arises when the circulation of productivity gains back into demand becomes too weak or too concentrated. If the income generated by AI-induced productivity gains is saved, retained, distributed to high-saving groups or invested in ways that do not sufficiently support demand, household consumption may not expand in proportion to productive capacity. In this sense, labour income and functional income distribution are not secondary to the productivity debate. They are part of the conditions under which productivity gains become economically realised.

This is where the AI productivity debate should connect more explicitly with the literature on income distribution and effective demand. A productivity gain matters macroeconomically not only because it increases output per unit of input, but also because it changes the flow of income in the economy. If that income flow supports household purchasing capacity on a sufficiently broad basis, productivity gains may be absorbed through consumption and other components of demand. If it does not, the economy may face a gap between technological capacity and market absorption.

The next section therefore develops the demand-side foundation of the article. It connects labour income, household disposable income, marginal propensity to consume and final household consumption. This step is necessary because the proposed distributional absorption threshold cannot be understood only from the perspective of productivity or labour-market exposure. It must be understood as a relationship between productivity growth, income distribution and the capacity of demand to absorb additional output.

4. Labour Income, Household Consumption and Effective Demand

4.1. From Labour Income to Household Disposable Income

The previous section showed that AI-related productivity gains may be distributed through different labour-market and income channels. The present section develops the demand-side implication of that argument. A productivity gain becomes macroeconomically relevant not only because it increases output per unit of input, but also because it modifies the income flows through which households, firms and governments finance expenditure. For this reason, the connection between labour income, household disposable income and consumption is central to the proposed distributional absorption framework.

Labour income is particularly important because, in most economies, it represents the main source of income for a large share of households. Wages, salaries, self-employment income and employment-related earnings affect the capacity of households to finance consumption without relying excessively on credit or asset liquidation. When productivity gains are transmitted into labour income, they can strengthen household purchasing power and support the absorption of additional output. When this transmission is weak, the macroeconomic effects of productivity growth may depend more heavily on other channels, such as investment, public expenditure, social transfers, lower prices or external demand.

This does not mean that labour income is the only relevant component of demand. Profits may finance investment, public revenues may finance transfers and services, and external demand may absorb part of domestic production. However, labour income deserves particular attention because it links the production side of the economy to household consumption more directly than many other income flows. If productivity growth raises output potential but does not sufficiently support the income of households with a relatively high propensity to consume, the demand-side realisation of that productivity may become more fragile.

The functional distribution of income is therefore relevant to the analysis. The division of value added between labour and capital affects the composition of income flows and, consequently, the way productivity gains return to demand. Classical and Keynesian-inspired approaches have long recognised that wages, profits and investment interact in shaping demand and growth [22,23,24]. More recent research on the decline of the labour share has also shown that the distribution between labour and capital is not constant and may vary across countries, sectors and time [25,26]. This matters for the present article because AI-induced productivity gains may have different macroeconomic effects depending on whether they strengthen labour income, capital income, prices, investment or public revenues.

Household disposable income provides a broader bridge between production and consumption because it includes not only labour income but also taxes, transfers, social benefits, property income and other redistributive mechanisms. This is important because weak labour income transmission may, in some contexts, be partly offset by fiscal redistribution, social protection or public transfers. Conversely, if labour income weakens and redistribution is limited, the capacity of households to absorb additional output through consumption may be reduced. The relevant question is therefore not only whether AI raises productivity, but whether the income generated by that productivity reaches households in forms that support real purchasing power.

This distinction matters because labour income and household disposable income are not merely social outcomes of technological change. They are macroeconomic channels through which productivity gains are transmitted to, or disconnected from, effective demand.

4.2. Consumption Heterogeneity and the Marginal Propensity to Consume

The relationship between income distribution and demand depends crucially on the heterogeneity of household consumption behaviour. If all households spent the same proportion of additional income, the distribution of productivity gains would matter less for aggregate consumption. In practice, however, households differ substantially in their marginal propensity to consume. Lower-income, liquidity-constrained or low-wealth households tend to spend a larger share of additional income, while wealthier households generally save a larger share [27,28]. This insight is central to the distributional absorption framework.

The marginal propensity to consume links distributional outcomes to macroeconomic demand. A productivity gain transmitted to liquidity-constrained or low-wealth households is more likely to become consumption expenditure than an equivalent gain accruing to households or entities with a higher propensity to save. The same reasoning applies to the distinction between labour income and capital income. If AI-induced productivity gains are concentrated in profits, dividends, retained earnings or capital gains, their immediate effect on household consumption may be weaker than if part of those gains is transmitted through wages, employment income or transfers. Mian et al. [29] develop a related argument by showing how differences between borrowers and savers can shape aggregate demand.

Accordingly, the issue is not a simple contrast between wages and profits, but the macroeconomic circulation of income. Profits may support demand through investment, taxation or redistribution, and wages may support demand differently depending on the households that receive them. The proposed threshold concerns the point at which the circulation of productivity gains becomes too weak or too concentrated to sustain the demand required for the absorption of additional output.

Consumption heterogeneity also implies that aggregate indicators may conceal important mechanisms. An economy may show rising average income while the income of high-consumption households stagnates. It may show productivity gains while household consumption grows slowly because gains accrue to groups with lower marginal propensities to consume. It may also maintain consumption temporarily through credit, even when labour income growth is weak. Such patterns are analytically important because they suggest that output absorption can be supported, delayed or distorted through different income and financial channels.

For this reason, a theoretical framework on AI-induced productivity should not stop at productivity or wage indicators alone. It must examine how productivity gains affect the distribution of real purchasing power. The relevant concept is not simply average income growth, but broadly distributed real purchasing power. This broader measure is intentionally wider than wages. It includes the purchasing capacity of households after accounting for prices, taxes, transfers, working time, employment income and other income sources. Such a framing is necessary because AI may affect not only wages, but also prices, profits, public revenues, job structures and the need for redistribution.

4.3. Effective Demand and the Absorption of Productivity Gains

The concept of effective demand provides the macroeconomic foundation for the proposed distributional absorption framework. In Keynesian and Kaleckian traditions, output is not determined only by productive capacity but also by expenditure decisions and by the demand that validates production [22,30]. This perspective is particularly useful for analysing AI because a technology that raises productive efficiency may still require demand-side conditions for its output effects to be realised. AI may make it possible to produce more, faster or at lower cost, but the economic realisation of this potential depends on whether households, firms, governments or foreign buyers are willing and able to purchase the additional output.

Distribution-sensitive models of demand provide a further layer of analysis. Bhaduri and Marglin [24] show that the effects of changes in income distribution depend on the relative responses of consumption, investment and net exports. Later wage-led growth approaches developed this insight by examining whether a higher wage share supports or constrains demand under different institutional and external conditions [31,32]. These approaches are not directly about AI, but they offer a theoretical basis for analysing whether productivity gains are more likely to be absorbed when they are transmitted into wages and household income.

The relevance of this literature for AI lies in its treatment of distribution as a demand condition. If productivity gains reduce costs and prices, households may benefit through higher real purchasing power even without proportional wage increases. If productivity gains support investment, firms may absorb part of the additional output through capital formation. If public revenues rise, governments may support demand through expenditure, transfers or services. If external demand is strong, foreign markets may absorb part of domestic production. These are all possible channels of absorption. The proposed framework does not deny them. It argues more cautiously that when these channels are insufficient and when broadly distributed purchasing power lags behind productivity growth, an absorption tension may emerge.

This formulation also distinguishes the proposed threshold from a simple underconsumption argument. The relevant issue is not whether household consumption alone immediately matches productivity growth, but whether the overall demand structure, including consumption, investment, public expenditure and external demand, can realise the productive capacity enabled by AI.

A distributional absorption problem may therefore be understood as a mismatch between the pace of productivity expansion and the pace at which purchasing power and demand-generating expenditure expand. If AI-induced productivity growth is accompanied by higher real wages, lower prices, stronger investment, adequate redistribution or external demand, the additional productive capacity may be absorbed. If, however, productivity gains are weakly transmitted to households and are not sufficiently compensated by other demand channels, the economy may become capable of producing more than the market can absorb under existing distributive conditions.

This reasoning prepares the conceptual definition developed in the next section. The Distributional Absorption Threshold of AI-Induced Productivity does not describe a collapse of demand, nor does it assume that AI necessarily reduces employment. It refers to a condition in which the distributional transmission of productivity gains becomes insufficient to sustain proportionate growth in broadly distributed real purchasing power and household consumption. The concept therefore integrates three elements that are often discussed separately: technological productivity, income distribution and effective demand.

This integration is the main theoretical step required for a more balanced discussion of AI and macroeconomic transformation. A purely technological view may overstate the automatic benefits of productivity growth. A purely labour-displacement view may overstate the risk of unemployment while neglecting other distributive channels. A demand-side distributional view allows a more cautious interpretation: AI may generate productivity gains, but the absorption of those gains depends on how the income generated by productivity is distributed, spent, invested, taxed and transferred. This is the analytical space in which the proposed threshold becomes meaningful.

5. The Distributional Absorption Threshold of AI-Induced Productivity

5.1. Conceptual Definition and Theoretical Positioning

The concept of the Distributional Absorption Threshold of AI-Induced Productivity [11] is proposed to capture a specific macroeconomic and distributional condition. It refers to the point beyond which productivity gains associated with the adoption or use of artificial intelligence are no longer accompanied by proportionate increases in broadly distributed real purchasing power and household consumption. Building on this initial definitional formulation, the present review develops the concept into a broader theoretical framework linking AI-induced productivity, labour income, income distribution, household demand and output absorption.

The threshold does not describe a technological limit of artificial intelligence. It does not refer to the maximum capacity of AI systems to automate tasks, process information, support prediction or increase efficiency. Rather, it refers to the macroeconomic conditions under which the productive capacity enabled by AI may fail to be fully realised through demand. In this sense, the threshold is not located inside the technology itself, but in the relationship between technological productivity, distributive transmission and effective demand.

This distinction matters because a technology can expand productive capacity without automatically generating the purchasing power required to absorb additional output. AI may allow firms to produce more efficiently, reduce costs, improve decision-making and reorganise tasks. However, if the gains generated by these improvements are not sufficiently transmitted into household income, lower prices, investment, public expenditure or external demand, the resulting productive capacity may remain only partially realised. The threshold therefore refers to a possible mismatch between the expansion of supply-side capacity and the distributive foundations of demand.

Figure 1 summarises the general conceptual sequence proposed in this article. It presents the threshold framework as a single chain linking AI adoption or exposure, productivity growth, distributive transmission, broadly distributed purchasing power, household consumption and effective demand, and the possible outcome of output absorption or absorption tension. Its purpose is to show the overall logic of the framework, not to distinguish between alternative transmission paths.

The concept builds on, but is not identical to, existing demand-side and distribution-sensitive interpretations of technological progress. Theories of effective demand emphasise that productive capacity must be validated by expenditure. Underconsumption arguments highlight the possibility that insufficient purchasing power may constrain the realisation of output. Wage-led and distribution-sensitive growth approaches show that the functional distribution of income can shape consumption, investment and net exports [22,24,30,31,32]. These studies provide the theoretical foundation of the present framework, but they do not specifically address the transmission of AI-induced productivity gains through labour income, household purchasing capacity and output absorption.

The specific contribution of the Distributional Absorption Threshold of AI-Induced Productivity does not lie in identifying, separately, the productivity–pay gap, effective demand constraints or distributional risks. These elements are already present in the literature on labour share, effective demand, wage-led growth, demand-constrained technological change and AI-related distributional risks [22,24,25,30,31,32,33,34]. The added value of the concept lies in combining these elements into an AI-specific absorption mechanism. While Gries and Naudé [33,34] show that aggregate demand matters for AI, automation and growth, the proposed concept gives a more explicit analytical form to the threshold-like mismatch that may emerge when AI-related productivity gains outpace the growth of broadly distributed real purchasing power, household consumption and compensating demand channels. It also moves beyond a general discussion of AI-related inequality by treating distributional effects not only as equity outcomes but also as part of a macroeconomic realisation mechanism linking AI-induced productivity, labour income, purchasing power and effective demand. DPI (Productivity–real labour income gap) and TA (Absorption tension indicator) are therefore not presented as new standalone indicators but as preliminary empirical signals that may help future research identify when this absorption mechanism becomes weak or incomplete.

The relationship with the author’s earlier definitional contribution [11] should therefore be understood more specifically. The earlier formulation introduced the Distributional Absorption Threshold of AI-Induced Productivity as a concise definitional construct. The present review adds four elements: it positions the concept in relation to effective demand, labour-share and wage-led growth in the literature; distinguishes favourable and critical transmission paths; formulates conceptual propositions; and proposes preliminary operational indicators and data sources for future empirical research. The added value is thus the transformation of the initial definition into an analytical framework for studying the transmission of AI-induced productivity into labour income, purchasing power and effective demand.

The threshold should therefore be understood as a context-dependent analytical construct rather than as a fixed empirical constant. Its relevance may vary across economies, sectors and institutional settings, depending on labour-market institutions, fiscal redistribution, price dynamics, investment behaviour, external demand and the sectoral structure of AI adoption.

5.2. What the Threshold Is and What It Is Not

A careful definition requires a clear distinction between what the distributional absorption threshold is and what it is not. The concept does not imply that AI is inherently harmful, nor does it deny the potential benefits of AI for efficiency, innovation, service quality and productive capacity. Its purpose is more specific: to ask whether the gains generated by AI-related productivity improvements are distributed and recirculated in ways that support effective demand.

Nor is it a synonym for technological unemployment. Employment displacement may weaken income transmission, but it is only one possible channel. Absorption tension may also arise through slower wage growth, reduced working hours, declining labour share, profit concentration or limited disposable income growth, even when aggregate employment remains relatively stable. Similarly, the threshold is not determined by the rate of AI adoption alone. High AI adoption may be compatible with strong absorption if productivity gains support wages, lower prices, investment, public revenues or external demand.

Finally, it should not be interpreted as a standard productivity slowdown or as a universal crisis point. It refers to a possible mismatch between AI-enabled productive capacity and the demand-side conditions required for its realisation. Whether such a mismatch emerges depends on institutions, distributional mechanisms and compensating demand channels.

Table 1. Conceptual clarification of the distributional absorption threshold.

Dimension	The Threshold Is	The Threshold Is Not
Technological meaning	A condition concerning the economic realisation of AI-enabled productive capacity	A technical limit of AI systems
Labour-market meaning	A broader distributive mechanism involving wages, hours worked, labour share and household income	A synonym for mass unemployment
Measurement meaning	A relationship between productivity growth and broadly distributed real purchasing power	A simple AI adoption rate
Macroeconomic meaning	A possible mismatch between productive capacity and effective demand	A standard productivity slowdown
Institutional meaning	A context-dependent condition shaped by labour-market institutions, fiscal redistribution, price dynamics and demand conditions	A universal threshold identical across countries

summarises these distinctions.

5.3. The Favourable Transmission Path

The favourable transmission path describes the conditions under which AI-induced productivity gains can be economically absorbed. In this trajectory, AI raises productivity and part of the resulting gain is transmitted through channels that support demand: higher real wages, stronger employment income, lower consumer prices, increased household disposable income, investment, public revenues or external demand. Under these conditions, the productive capacity enabled by AI is more likely to become demand-realised output.

The favourable path should therefore be understood as a broad demand-supporting configuration, not as a wage-only mechanism. Labour income remains central, but lower prices, investment, public expenditure, transfers and external demand may also contribute to the absorption of AI-induced productivity gains.

5.4. The Critical Transmission Path

The critical transmission path describes the conditions under which AI-induced productivity gains may become weakly absorbed. In this trajectory, productivity increases, but the distributive transmission of the gains is limited. Real labour income may grow slowly, wage gains may be concentrated, working hours may decline, employment income may weaken, or profits may capture a disproportionate share of the productivity gain. If these developments are not offset by lower prices, investment, redistribution, public expenditure or external demand, household consumption may fail to expand in proportion to productive capacity.

This interpretation does not imply that profits or capital income are economically undesirable, nor does it require a collapse of employment. The problem is more specific: absorption tension may arise when the income generated by productivity growth does not return to demand with sufficient strength or breadth.

Figure 2 complements Figure 1 by separating the general sequence into two ideal-type transmission paths. The favourable path (A) describes a configuration in which AI-induced productivity gains strengthen broadly distributed real purchasing power and effective demand. The critical path (B) describes a configuration in which productivity gains are weakly transmitted to labour income and household purchasing power, thereby increasing the risk of partial output absorption or absorption tension.

5.5. Simplified Representation and Analytical Interpretation

In simplified form, the distributional absorption threshold may be expressed as follows:

Rate of AI-induced productivity growth > Rate of growth in broadly distributed real purchasing power

(1)

This inequality should be interpreted carefully. It does not mean that every productivity gain must be matched immediately by household consumption growth. Nor does it imply that household consumption is the only possible demand channel. It indicates a condition in which the supply-side capacity generated by AI expands faster than the demand-side purchasing power broadly available to households. If the difference is temporary or compensated by other demand channels, it may not become a macroeconomic problem. If it is persistent and uncompensated, it may signal a distributional absorption tension.

The term “broadly distributed real purchasing power” is intentionally wider than real wages. It includes household purchasing capacity after taking into account employment income, working hours, prices, taxes, social transfers, disposable income and other relevant income flows. This broader formulation is necessary because AI-induced productivity gains may affect households through multiple channels. A narrow wage measure may miss the effect of lower prices, transfers, taxes, working time or changes in employment income.

The analytical value of the threshold lies in its capacity to connect three debates that are often separated. The first concerns AI and productivity: can AI raise output per worker or reduce production costs? The second concerns distribution: who receives the income generated by productivity gains? The third concerns demand: is there enough purchasing power and expenditure to absorb the additional output? The threshold concept integrates these questions into a single framework.

The Distributional Absorption Threshold of AI-Induced Productivity is therefore best understood as a theoretical lens. It does not predict a single outcome for all economies. It asks whether the productivity gains associated with AI are transmitted into the demand conditions required for their economic realisation. This is the point at which AI productivity becomes not only a technological issue but also a question of income distribution, purchasing power and macroeconomic absorption.

6. Conceptual Propositions

The theoretical framework developed in the previous sections suggests that AI-induced productivity growth should be analysed not only as a technological or labour-market phenomenon but also as a distributive and demand-side process. For this reason, the present article formulates conceptual propositions rather than empirical hypotheses. The distinction is important. Hypotheses are normally designed for direct empirical testing within a specified dataset and model. Conceptual propositions, by contrast, organise the theoretical logic of the article and identify relationships that future empirical research may examine.

The propositions below are not intended as deterministic claims. They do not assume that AI necessarily reduces employment, weakens wages or creates demand constraints. They identify the conditions under which AI-induced productivity gains may either support output absorption or generate distributional absorption tension. In this sense, the propositions translate the threshold concept into a more explicit analytical structure.

6.1. Productivity Gains and Broadly Distributed Purchasing Power

P1. 

AI-induced productivity gains are more likely to support demand-realised output when they are transmitted into broadly distributed real purchasing power.

The first proposition links AI-induced productivity gains to their demand-side realisation. Productivity growth contributes to output absorption when part of the resulting gains strengthens household purchasing capacity through wages, employment income, prices, transfers or other demand-supporting channels.

This proposition does not imply that wages are the only relevant channel, nor that household consumption is the only form of demand. Rather, it emphasises that the macroeconomic relevance of AI-induced productivity depends on the circulation of productivity gains through income and expenditure flows that can sustain demand. This interpretation is consistent with the broader literature on effective demand and distribution-sensitive growth, which emphasises that the composition of income flows matters for aggregate demand [22,24,30,31].

6.2. Productivity–Income Decoupling and Absorption Tension

P2. 

Distributional absorption tension becomes more plausible when AI-related productivity growth persistently outpaces real labour income, household disposable income and household consumption growth.

The second proposition identifies the conditions under which the proposed threshold becomes analytically relevant. A productivity–income gap alone is not sufficient to indicate distributional absorption tension. Such a gap may be temporary, may reflect non-AI productivity drivers, or may be offset by lower prices, investment, public expenditure, redistribution or external demand. The threshold interpretation becomes more plausible only when three elements are jointly observed: significant AI adoption or exposure, persistent productivity growth exceeding the growth of real labour income and household disposable income, and weak household consumption growth that is not sufficiently compensated by other demand channels.

Conversely, evidence against the threshold interpretation would include cases in which AI adoption or exposure is high and productivity growth is strong, while real labour income, household disposable income and household consumption also grow broadly in line with productivity. It would also include cases in which weak consumption dynamics are better explained by non-AI shocks, demographic change, inflation, fiscal consolidation, household deleveraging or external conditions. In this sense, the proposition is conditional rather than unfalsifiable. The threshold is not inferred from any productivity–income gap but from the joint observation of AI-related productivity dynamics, weak distributive transmission and insufficient demand compensation.

6.3. Employment Displacement as One Channel Among Several

P3. 

Employment displacement is only one possible channel through which AI-induced productivity gains may become weakly absorbed; wage compression, reduced hours, declining labour share, profit concentration and unequal access to AI complementarities may also weaken demand transmission.

The third proposition clarifies that weak absorption should not be reduced to technological unemployment. As the task-based literature suggests, AI may operate through substitution, complementarity and task reinstatement [7,8,9,10]. The relevant issue for the present framework is whether the gains generated through these channels are broadly transmitted to labour income and household purchasing power.

6.4. Compensating Demand Channels

P4. 

Redistribution, lower prices, public expenditure, investment and external demand may mitigate or delay the emergence of a distributional absorption threshold.

The fourth proposition clarifies the role of compensating demand channels. A productivity–income gap does not mechanically imply that the threshold has been reached, because output may also be absorbed through investment, public expenditure, redistribution, lower prices and external demand. The threshold becomes more plausible only when these channels are insufficient.

6.5. Institutional and Cross-Country Heterogeneity

P5. 

The distributional absorption threshold is likely to vary across economies depending on labour-market institutions, welfare systems, wage-setting mechanisms, market structure, fiscal capacity, household indebtedness and exposure to external demand.

The fifth proposition emphasises that the threshold should not be interpreted as universal. The same AI-induced productivity gain may have different macroeconomic effects across countries, sectors and institutional environments. In economies with strong wage bargaining, inclusive social protection and effective fiscal redistribution, productivity gains may be more likely to support household income and demand. In economies with weak labour institutions, high inequality, concentrated market power or limited redistribution, the same gains may be more weakly transmitted to broadly distributed purchasing power.

This proposition is especially relevant for future comparative research in the European Union. EU economies differ in wage-setting institutions, welfare systems, industrial structures, household consumption patterns, digital capacity and exposure to external demand. These differences may shape both the distribution of AI-induced productivity gains and the demand-side absorption of those gains. Consequently, the threshold should be investigated as an institutionally mediated condition rather than as a single numerical point applicable to all economies.

This institutional heterogeneity strengthens the usefulness of the concept. It allows researchers to examine not only whether AI raises productivity but also why the macroeconomic realisation of productivity may differ across countries. The concept can therefore support future comparative studies on the relationship between AI adoption, productivity growth, income distribution and effective demand.

6.6. Synthesis of the Conceptual Propositions

Taken together, the five propositions clarify the theoretical contribution of the article. The Distributional Absorption Threshold of AI-Induced Productivity is not a prediction of inevitable demand weakness. It is a framework for analysing the conditions under which productivity gains may become weakly absorbed because their distributive transmission into real purchasing power is insufficient.

More specifically, the propositions show that weak absorption may emerge through several labour income and demand-transmission channels. Job displacement matters, but the more general issue is whether AI-induced productivity gains circulate through the economy in ways that sustain demand. When this circulation is broad, AI-induced productivity growth may support output absorption; when it is weak or concentrated, absorption tension may become more likely.

Table 2. Conceptual propositions and underlying mechanisms.

Proposition	Core Mechanism	Expected Theoretical Implication
P1. AI-induced productivity gains support demand-realised output when transmitted into broadly distributed real purchasing power.	Productivity gains strengthen household purchasing power and demand-generating expenditure.	Output absorption becomes more likely.
P2. Distributional absorption tension becomes more plausible when AI-related productivity growth persistently outpaces real labour income, household disposable income and household consumption growth.	Productive capacity expands faster than household demand capacity.	Distributional absorption tension becomes more likely.
P3. Employment displacement is only one possible channel of weak absorption.	Wage compression, reduced hours, declining labour share and profit concentration may weaken demand transmission.	AI effects should be analysed beyond unemployment alone.
P4. Redistribution, lower prices, public expenditure, investment and external demand may mitigate the threshold.	Alternative demand channels compensate for weak labour income transmission.	The threshold is conditional, not automatic.
P5. The threshold varies across institutional and national contexts.	Labour institutions, welfare systems, fiscal capacity and external demand shape transmission.	Comparative analysis is needed to identify country-specific risks.

summarises the five conceptual propositions, their core mechanisms and their expected theoretical implications.

7. Operationalising the Concept for Future Empirical Research

7.1. From Theoretical Construct to Empirical Framework

The Distributional Absorption Threshold of AI-Induced Productivity has been developed in this review as a theoretical construct. Nevertheless, the concept is designed to be operationalised in future empirical research. This is important because the value of the framework lies not only in its conceptual clarity but also in its capacity to guide measurable analysis of the relationship between AI, productivity, income distribution and effective demand.

Operationalising the concept requires a careful distinction between three levels of analysis. The first concerns AI adoption or AI exposure. The second concerns productivity and the distribution of productivity gains. The third concerns the demand-side absorption of those gains through household purchasing power and other components of effective demand. Future empirical research should avoid collapsing these levels into a single indicator. A high level of AI adoption does not necessarily imply high productivity growth, and high productivity growth does not necessarily imply strong output absorption.

Although the framework is general and not limited to the European Union, this section gives particular attention to EU data sources because they provide comparatively harmonised indicators on AI adoption, labour markets, productivity, income and household consumption across countries. The EU is therefore used mainly as an illustrative and empirically tractable context for future comparative operationalisation, not as the exclusive geographical scope of the theory. The same framework could be applied to other economies or regions where comparable data on AI exposure, productivity, income distribution and demand channels are available.

For non-EU applications, Eurostat and AMECO (Annual Macro-Economic Database of the European Commission’s Directorate-General for Economic and Financial Affairs) variables may be replaced by functionally equivalent national accounts, labour market, firm adoption or household income sources, provided that definitions, deflators and sectoral classifications are made comparable. The analytical requirement is not the use of a specific European database, but the consistent combination of AI adoption or exposure, productivity, labour income, disposable income and consumption indicators.

A useful empirical strategy would therefore treat the threshold as a relationship rather than as a single variable. The relevant question is not whether a country or sector uses AI, but whether AI-associated productivity growth is accompanied by sufficient growth in broadly distributed real purchasing power. This requires combining indicators of AI adoption or exposure with indicators of productivity, labour income, household disposable income and real household consumption. The concept is thus operationalised through the interaction between technological change, distributive transmission and demand-side realisation.

7.2. Measuring AI Adoption and AI Exposure

The first measurement issue concerns the distinction between AI adoption and AI exposure. AI adoption refers to the actual use of AI technologies by firms, organisations or households. AI exposure refers to the extent to which occupations, tasks or sectors could be affected by AI capabilities, regardless of whether AI is already widely used. These two concepts are related, but they are not equivalent.

For European Union economies, Eurostat provides a direct source for firm-level AI adoption through indicators on enterprises using artificial intelligence technologies, including the dataset on artificial intelligence by size class of enterprise [35,36]. Such data are useful because they provide comparable information across EU Member States and enterprise-size categories. However, they also have limitations. The time series is still relatively short, and the data measure reported use of AI technologies rather than the intensity, quality or productivity impact of that use.

Complementary evidence can be obtained from firm-level and policy-oriented sources. The OECD report on AI adoption in firms, prepared in collaboration with Boston Consulting Group (BCG) and INSEAD (Institut Européen d’Administration des Affaires), provides survey-based evidence on adoption barriers, firm capabilities and organisational conditions [37]. The Stanford AI Index offers broader contextual information on AI development, investment, technical progress and diffusion [38]. For the European policy context, the European Commission’s AI Continent Action Plan is relevant because it frames AI adoption as a strategic objective for competitiveness, infrastructure, skills and industrial transformation [39].

A further complementary source can be firm-level AI disclosure in annual reports or regulatory filings. Dogan and Teklie [40], using a decade of S&P 500 10-K filings, show that disclosure-based measures can trace the diffusion and salience of AI in corporate reporting. Such indicators are useful because they may capture whether AI has become strategically or risk-relevantly visible to firms. However, they should be used with caution. A binary variable indicating whether a firm mentions AI may become less informative as AI disclosure becomes widespread, and the volume of AI-related text should not be treated as a direct measure of AI engagement, implementation quality or productivity impact. Disclosure-based indicators therefore complement, but do not replace, adoption and exposure measures. They are particularly useful for triangulating AI salience at the firm level, but they do not by themselves identify AI-induced productivity gains or the distributive transmission of those gains into labour income and demand.

AI exposure should be measured differently. Occupational exposure indices, such as those developed by Felten et al. [15], Pizzinelli et al. [41] and Gmyrek et al. [42], are useful because they capture the degree to which tasks or occupations overlap with AI capabilities. These measures are especially valuable when actual adoption data are unavailable or too recent. However, exposure indicators should be interpreted cautiously. They measure potential exposure, not actual adoption, job loss or productivity effects. They are best used as indicators of where AI-related productivity and distributive pressures may arise, not as direct measures of realised economic outcomes.

7.3. Productivity, Labour Income and Demand-Side Indicators

The second measurement level concerns productivity and the distribution of productivity gains. Labour productivity can be measured using output per worker, output per hour worked or value added per unit of labour input. For macroeconomic analysis in EU economies, AMECO provides annual data on output, employment, compensation, wage share and related macroeconomic aggregates [43]. For sector-level analysis, EU KLEMS (European Union database for growth and productivity analysis based on capital, labour, energy, materials and services inputs) offers industry-level data on productivity, employment, capital formation and technological change [44]. For broader international comparisons, the Penn World Table provides information on output, input and productivity across countries and over time [45,46].

Labour income transmission can be captured through several indicators, including real compensation of employees, real wages, employment income, hours worked, adjusted wage share and labour share. No single indicator is sufficient. Real wages may capture purchasing power per worker, but may miss changes in employment and hours worked. Compensation of employees may capture aggregate labour income, but may not fully reflect distribution across households. Wage share may capture functional distribution, but may not identify which households receive the income. Future empirical research should therefore use a combination of labour income indicators rather than relying on a single measure.

Household disposable income is essential because it connects production and distribution to consumption. It incorporates labour income, taxes, transfers, social benefits, property income and other redistributive mechanisms. It is therefore better suited than wages alone for analysing demand-side absorption. Real household final consumption expenditure is also central because it captures the extent to which household purchasing power is converted into demand. However, consumption must be interpreted carefully. It may be temporarily sustained by credit, wealth effects or savings, even when labour income growth is weak.

Inequality indicators may also be relevant. The World Inequality Database and the World Income Inequality Database provide sources for analysing income and wealth distribution across countries and over time [47,48]. These sources can help future research examine whether productivity gains are broadly distributed or concentrated. The inclusion of inequality measures is important because the proposed threshold concerns not only average purchasing power, but also broadly distributed real purchasing power.

7.4. Possible Operational Indicators: DPI and TA

The indicators proposed below should be interpreted as preliminary operational tools for future empirical research, not as definitive measures of a causal threshold. They are intended to identify patterns that may be consistent with distributional absorption tension, but they cannot by themselves prove that AI has caused such tension. This distinction is especially important because observed productivity growth reflects many determinants besides AI, including capital deepening, non-AI digitalisation, managerial change, sectoral composition, regulation and macroeconomic conditions. Consequently, DPI and TA should be combined with independent measures of AI adoption, AI exposure or AI disclosure, and with empirical strategies capable of distinguishing AI-related effects from broader productivity dynamics.

The first proposed operational indicator is the productivity-real labour income gap:

DPI_it = g(PROD_it) − g(RW_it)

(2)

where DPI_it denotes the productivity-real labour income gap in country or sector i at time t, g(PROD_it) denotes the growth rate of labour productivity, and g(RW_it) denotes the growth rate of real labour income or real wages.

A positive DPI value indicates that labour productivity is growing faster than real labour income. This indicator is closely related to the established productivity-pay gap and labour-share literature [25] and should not be interpreted as a new measure in itself. Its role in the present framework is more specific: it can be used as one component of an AI-oriented analysis when combined with evidence on AI adoption or exposure. A positive DPI value does not automatically imply a distributional absorption problem, because the gap may be temporary, may reflect non-AI productivity drivers, or may be offset by lower prices, transfers, public expenditure, investment or external demand. However, if positive DPI values are persistent, coincide with significant AI adoption or exposure, and are accompanied by weak household consumption or disposable income growth, they may provide preliminary evidence consistent with weak transmission from productivity to demand.

The second proposed indicator is the absorption tension indicator:

TA_it = z[g(PROD_it)] − z[g(CONS_it)]

(3)

where TA_it denotes absorption tension, g(PROD_it) denotes productivity growth, g(CONS_it) denotes real household consumption growth, and z represents standardisation.

This indicator compares the standardised growth of productivity with the standardised growth of real household consumption. Because it uses z-scores, TA is sample-dependent and should not be interpreted as having a fixed universal zero or a stable cross-sample threshold. Its value depends on the countries, sectors, time period and variables included in the analysis. For this reason, TA should be used as an exploratory and comparative signal, not as a standalone measure of the distributional absorption threshold. A positive TA value may suggest that productivity growth is outpacing consumption growth within the analysed sample, but it does not identify causality and does not, by itself, show that AI is responsible for the divergence.

A simple hypothetical illustration may clarify the interpretation of the two indicators. Suppose that, in country i, labour productivity grows by 4% in year t, while real labour income grows by 2%. In this case, DPI_it would be 2 percentage points, indicating that productivity is growing faster than real labour income. If, in the same analytical sample, the standardised value of productivity growth is +1.20 and the standardised value of real household consumption growth is +0.30, TA_it would be +0.90. Such values would not prove the existence of a distributional absorption threshold, but they would identify a case requiring further examination, especially if AI adoption or exposure is significant and compensating demand channels are weak.

Future empirical research could also combine DPI and TA with indicators of AI adoption, AI exposure or AI disclosure. The most relevant cases would be those in which productivity grows faster than real labour income and also faster than household consumption, while AI adoption or exposure is significant and compensating demand channels are weak. Conversely, if productivity grows faster than wages but consumption remains strong because of lower prices, transfers, investment spillovers, public expenditure or external demand, the threshold may not be reached. This is why the concept should be measured through a set of indicators rather than through a single formula.

Evidence against the threshold interpretation would include cases in which AI adoption or exposure is high, productivity growth is strong, and yet real labour income, household disposable income and household consumption also grow proportionately. It would also include cases in which weak consumption dynamics are better explained by non-AI shocks, demographic change, fiscal consolidation, inflation, financial deleveraging or external conditions. In this sense, the framework is not designed to classify every productivity–income gap as an AI-related absorption problem. It requires the joint observation of AI-related productivity dynamics, weak distributive transmission and insufficient compensating demand channels.

7.5. Research Designs and Methodological Cautions

Several empirical designs could be used to investigate the proposed framework. Descriptive and comparative analyses could map countries or sectors according to AI adoption or exposure, productivity growth, labour income dynamics, disposable income growth and household consumption. Such analyses would not identify causal effects, but they could identify cases where absorption tension appears more or less plausible.

Sector-level and panel designs would be especially useful because AI adoption and exposure are uneven across industries. Sectoral analysis could combine productivity and compensation data, such as EU KLEMS [44], with AI exposure or adoption indicators to examine whether more exposed sectors show different patterns of productivity, labour compensation and value-added distribution. Panel models could then examine whether productivity growth is associated with household consumption growth conditional on labour income transmission, redistribution, prices, investment and external demand.

Future empirical work may also consider quasi-experimental or exposure-based strategies when data permit, for example by comparing sectors, occupations or firms with different degrees of AI exposure before and after major AI diffusion episodes. Threshold models, including panel threshold approaches [49], may be useful for testing whether the relationship between productivity growth and demand-side absorption changes beyond a certain level of distributive decoupling, such as DPI. However, such models require sufficiently long and comparable data series and should be treated as a future research possibility rather than as an immediate requirement.

The main methodological caution is that AI adoption, productivity growth, income distribution and demand are jointly determined. Firms may adopt AI because they are already productive; productive firms may pay higher wages or exercise stronger market power; and household consumption may respond to credit, wealth, expectations, inflation and public policy, not only to labour income. Future empirical studies should therefore avoid simplistic causal claims and should combine AI adoption, AI exposure or AI disclosure measures with research designs capable of separating AI-related effects from broader productivity, institutional and macroeconomic dynamics. In this sense, the distributional absorption threshold remains a conceptual and analytical tool for guiding empirical investigation, not a mechanically observable point.

Table 3. Possible indicators for future empirical operationalisation.

Analytical Dimension	Possible Indicators	Possible Sources	Main Limitation
AI adoption	Enterprises using AI technologies; AI use by firm size and sector; firm-level AI disclosures	Eurostat [35,36]; OECD/BCG/INSEAD [37]; AI Index [38]; corporate filings [40]	Short time series; adoption and disclosure do not measure intensity, implementation quality or productivity impact
AI exposure	Occupational AI exposure; generative AI exposure; complementarity/substitution indicators	Felten et al. [15]; Pizzinelli et al. [41]; ILO/Gmyrek et al. [42]	Measures potential exposure, not actual adoption, productivity effects or labour income outcomes
Productivity	Labour productivity; output per hour; value added per worker; total factor productivity	AMECO [43]; EU KLEMS [44]; Penn World Table—PWT [45,46]	Measures total productivity, not AI-induced productivity by itself; aggregation may hide firm and sector heterogeneity
Labour income transmission	Real wages; real compensation of employees; employment income; hours worked; adjusted wage share; labour share	AMECO [43]; Eurostat sector accounts [50]; EU KLEMS [44]	Single indicators may not capture household distribution, working-time changes or income concentration
Disposable income and consumption	Real household disposable income; real household final consumption expenditure	Eurostat sector accounts [50]; AMECO [43]; World Bank WDI (World Development Indicators) [51]	Consumption may be sustained by credit, wealth effects or savings
Distributional conditions	Income inequality; wealth inequality; income shares; labour share	WID (World Inequality Database) [47]; WIID (World Income Inequality Database) [48]; AMECO [43]; PWT [45,46]	Cross-country comparability and frequency may vary; average indicators may hide within-household and group-level differences
Absorption tension	DPI_it; TA_it; joint patterns of productivity, labour income, disposable income and consumption	Author’s proposed operationalisation based on the conceptual framework developed in this article	Exploratory indicators only; they do not identify causality and must be combined with AI adoption, exposure or disclosure measures

summarises possible indicators for future operationalisation, together with indicative sources and main limitations. These indicators should be read as complementary measurement dimensions, not as standalone evidence that AI has caused a distributional absorption threshold.

8. Discussion: Theoretical and Policy Implications

8.1. Reframing AI Productivity Beyond Automation Anxiety

The framework developed in this article contributes to the debate on artificial intelligence by shifting attention from automation anxiety alone to the distributive and demand-side realisation of productivity gains. Much of the public and academic discussion on AI has been organised around the question of whether AI will replace jobs. This question is important, but it is not sufficient. A broader macroeconomic interpretation should also ask how AI-induced productivity gains are distributed, whether they strengthen household purchasing power and whether they support the demand required to absorb additional output.

This reframing does not minimise the relevance of labour displacement. The task-based literature shows that automation can displace labour from some activities while also creating or reinstating other tasks [9,10]. The literature on job polarisation further shows that technological change can reshape the structure of employment and wages across occupations and skill groups [52,53]. However, the distributional absorption framework suggests that the central issue is broader than employment levels. Even if aggregate employment does not collapse, productivity gains may still be weakly absorbed if they are not transmitted into broadly distributed real purchasing power.

This perspective is particularly relevant for AI because AI may affect both routine and non-routine cognitive tasks. Its distributive consequences may therefore differ from those of earlier waves of automation. AI may complement some workers, substitute specific tasks performed by others, increase productivity in knowledge-intensive occupations and strengthen the position of firms that control data, models or digital infrastructure. In this context, the more relevant question is not whether AI is simply labour-saving or labour-augmenting, but whether the resulting gains circulate through the economy in ways that sustain effective demand.

The proposed threshold is therefore not a prediction of weak demand, but a way of identifying when the macroeconomic benefits of AI-induced productivity may remain incomplete because distributive transmission is weak.

8.2. Distributional Transmission as a Condition of Inclusive Productivity

The concept of inclusive productivity helps clarify the contribution of the proposed framework. Productivity growth is usually treated as a desirable economic outcome because it can increase output, reduce costs and support higher living standards. Yet its social and macroeconomic effects are stronger when it is connected to broad gains in income, purchasing power and welfare. If productivity gains are highly concentrated, their effect on household consumption and effective demand may be weaker.

This does not mean that all productivity gains must be distributed immediately through wages. Such an interpretation would be too narrow. Productivity gains may support demand through several channels, including lower consumer prices, investment, public revenues, transfers, social services and external demand. The relevant issue is whether these channels are strong enough to transform productive capacity into realised output. If they are weak, the economy may experience a form of distributive under-realisation, in which technological capacity expands more rapidly than the purchasing power required for its absorption.

The literature on AI and economic growth recognises that AI may raise productivity and accelerate automation, but also that the distribution of its gains may be uneven [54,55]. Korinek and Stiglitz [56] similarly highlight the distributional challenges associated with AI and worker-replacing technological progress. The present article connects this concern to effective demand. Inequality is not only a social or ethical concern in the context of AI. It may also become a macroeconomic issue if it weakens the demand-side absorption of productivity gains.

This interpretation is consistent with distribution-sensitive approaches to growth: the demand effect of productivity gains depends on who receives the associated income and how it returns to expenditure. The framework therefore focuses on the circulation of income through consumption, investment, redistribution, public expenditure and exports.

8.3. Institutional Mediation of the Threshold

The distributional absorption threshold should not be understood as a universal point detached from institutions. The same AI-induced productivity gain may have different effects depending on wage-setting systems, collective bargaining, social protection, competition policy, fiscal capacity, public investment, household balance sheets and the sectoral structure of AI adoption. For this reason, the threshold is better understood as institutionally mediated rather than technologically predetermined.

This institutional mediation should not be interpreted as treating institutions as direct effects of AI. In the framework proposed here, institutions and public policies are conditioning factors that influence how AI-related productivity gains are distributed, absorbed or compensated. Future empirical research should therefore distinguish between AI adoption or exposure, productivity and income outcomes, and the institutional environment through which these outcomes are transmitted to households and demand.

Relevant policy channels include wage-setting mechanisms, training and reskilling programmes, employment protection, social insurance, tax-benefit systems, public investment, competition policy and digital market regulation. These channels may influence whether AI-related productivity gains are transmitted into wages, disposable income, lower prices, public revenues, investment or broader effective demand. This institutional perspective is especially relevant for the European Union, where differences in welfare systems, labour-market institutions, fiscal capacity, digital readiness and exposure to external demand may lead to different forms of absorption or absorption tension across countries.

8.4. Policy Implications for AI-Induced Productivity

The policy implications of the framework should be formulated cautiously. The article does not argue that AI should be slowed down or discouraged. Nor does it suggest that productivity gains are undesirable. On the contrary, AI-induced productivity may support innovation, competitiveness and welfare. The relevant policy issue is how to strengthen the channels through which productivity gains are translated into broad purchasing power and effective demand.

From this perspective, policy relevance concerns several interconnected channels. Wage-setting and labour-market institutions may influence whether productivity gains reach households through labour income. Redistribution and social protection may support disposable income when labour income transmission is weak. Competition and market regulation may affect the pass-through of productivity gains into prices and the degree of profit concentration. Public investment may help transform productivity gains into demand-generating expenditure and future productive capacity. External demand may also absorb part of the additional output, although it cannot serve as a universal solution for all economies simultaneously.

Table 4. Policy-relevant channels for reducing distributional absorption tension.

Policy Channel	Analytical Role	Relevance for Absorption
Wage-setting and labour institutions	Influence the transmission of productivity gains to labour income	Strengthen household purchasing power
Skills and mobility policies	Help workers move towards complementary tasks	Reduce displacement and support income continuity
Redistribution and social protection	Convert part of productivity gains into household demand	Stabilise consumption and disposable income
Competition and market regulation	Affect price pass-through and profit concentration	Support consumer purchasing power and limit excessive concentration
Public investment	Absorb output and raise future productive capacity	Transform productivity gains into demand-generating expenditure
External demand	Provides markets for additional output	May compensate for domestic demand weakness, but cannot be universal

summarises the main policy-relevant channels through which distributional absorption tension may be reduced or mitigated.

8.5. Limits of the Theoretical Argument

The framework proposed in this article has several limits that should be made explicit. First, it does not provide empirical proof that a distributional absorption threshold has already been reached in any economy. The concept is theoretical and analytical. It identifies a possible condition and offers a way to study it, but it does not claim that the condition is already present.

Second, the framework does not isolate AI from other forms of technological change. In practice, AI adoption interacts with digitalisation, automation, platformisation, global value chains, demographic change and macroeconomic policy. Future empirical research will need to distinguish AI-related effects from broader technological and institutional processes. This is methodologically difficult, especially because AI adoption data are still recent and often incomplete.

Third, the framework is not a pessimistic forecast and does not deny that AI can generate broad benefits through productivity growth, service improvements, new tasks, cost reductions and innovation. Its narrower claim is that the macroeconomic realisation of these benefits depends on distributive and demand-side conditions, including the strength of compensating channels such as investment, public expenditure and external demand.

The value of the proposed concept lies precisely in this caution. It allows researchers to move beyond both technological optimism and automation anxiety. It asks how productivity gains are transmitted, who receives the income they generate, and whether demand is sufficient to realise the additional output. These questions are likely to become more important as AI adoption expands and as economies seek to transform technological capacity into broad-based welfare gains.

9. Conclusions and Future Directions

9.1. Main Theoretical Conclusions

This review developed the concept of the Distributional Absorption Threshold of AI-Induced Productivity as a theoretical framework for analysing the relationship between artificial intelligence, productivity, labour income, income distribution, household consumption and effective demand. Its central argument is that AI-induced productivity growth should not be assessed only through technological efficiency, but also through the distributive channels that determine whether productivity gains are transformed into broadly distributed real purchasing power and demand-realised output.

The proposed threshold is not a technological limit of AI and does not imply that AI-induced productivity growth is undesirable. It refers to a macroeconomic and distributional condition in which productivity gains may begin to outpace the growth of household purchasing capacity and consumption, without sufficient compensation through other demand channels. In this sense, the framework shifts the debate from whether AI can raise productivity to whether the gains from AI-induced productivity can be economically realised under existing distributive and institutional conditions.

A central contribution of the review is the distinction between favourable and critical transmission paths. The favourable path describes a situation in which AI-induced productivity gains strengthen purchasing power, effective demand and output absorption. The critical path describes a situation in which productivity gains are weakly transmitted to labour income and household demand, increasing the risk of absorption tension. These paths are not predictions, but analytical representations whose relevance depends on wage-setting mechanisms, market structure, redistribution, investment behaviour, price dynamics and external demand.

The review also showed that weak absorption cannot be reduced to employment displacement alone. A distributional absorption problem may arise if wage growth is limited, hours worked decline, labour income becomes more concentrated, the labour share weakens, or productivity gains accrue mainly to capital income and high-saving groups. This broader interpretation allows the AI debate to move beyond employment-centred concerns and to focus on the distributive conditions under which productivity gains become demand-realised.

9.2. Future Research Directions

Future research should first develop empirical strategies for measuring the relationship between AI-induced productivity growth and household purchasing capacity across the income distribution. This requires combining indicators of AI adoption or AI exposure with indicators of labour productivity, real labour income, household disposable income, final household consumption, wage share, investment, public expenditure and external demand. Because AI adoption data are still relatively recent, early empirical work should remain exploratory and cautious.

A second direction concerns the distinction between AI adoption and AI exposure. Adoption indicators capture the actual use of AI technologies, but they often have short time series and may not measure the intensity or quality of implementation. Exposure indicators capture the potential relevance of AI capabilities for occupations or tasks, but they do not measure actual adoption or realised productivity effects. Future studies should therefore combine both types of indicators rather than treating either of them as sufficient.

A third direction concerns sector-level research. AI adoption and AI exposure are unlikely to be uniform across the economy. Some sectors may experience stronger productivity gains, while others may experience weaker effects or slower diffusion. Sectoral analysis could clarify whether the distributional absorption problem is more likely to emerge in industries where productivity gains are high but labour income transmission is weak.

A fourth direction concerns household-level data. The proposed framework emphasises broadly distributed real purchasing power, not only average income. Future research should therefore examine how productivity gains are transmitted across income groups, wealth groups and household types. This would allow researchers to analyse whether gains reach households with higher marginal propensities to consume or remain concentrated in groups with weaker immediate effects on consumption demand.

A fifth direction concerns institutional comparison. The threshold is unlikely to be identical across economies. Comparative research within the European Union could examine how wage bargaining, social protection, taxation, redistribution, competition policy, public investment and external demand shape the absorption of AI-induced productivity gains. Such research would help identify whether some institutional configurations are better able to transform technological productivity into broad-based welfare gains.

Finally, future research could examine whether the proposed operational indicators, such as the productivity-real labour income gap and the absorption tension indicator, can provide useful empirical signals. These indicators should not be treated as final measurements of the threshold. They are exploratory tools that require validation, sensitivity testing and careful interpretation. Their value lies in making the relationship between productivity, income distribution and demand more visible.

Overall, the article suggests that the economic significance of AI will depend not only on what AI can technically do, but also on how societies organise the distribution and circulation of the gains it generates. AI-induced productivity growth may support broad welfare improvements if it is connected to purchasing power and effective demand. If this connection weakens, productivity gains may be less fully realised. The concept of the Distributional Absorption Threshold of AI-Induced Productivity offers a theoretical lens for examining this possibility in a cautious, non-deterministic and policy-relevant manner.