When economists encounter a new general-purpose technology, they face a fundamental interpretive problem. The empirical question—"what happens to employment?"—obscures the structural question: "how does this technology reshape the production function and factor allocations throughout the economy?"
These are categorically different. Confusing them leads to what we might call "vocational anxiety economics"—the belief that technological change is primarily a labor market phenomenon.
Consider the actual historical record: when electricity diffused through industrial economies in the early twentieth century, employment did not contract. It redistributed. What changed was not the quantity of work available, but its composition, location, and capital intensity. The production function itself reorganized. Factories no longer needed to be built around central power sources. Manufacturing could spatially disperse while remaining electrically coordinated. This unleashed an entirely new category of production possibilities—and with it, new forms of labor demand.
The candlemakers did lose their livelihoods. But this was not because electricity "destroyed jobs." It was because electricity changed the relative cost structure of light production so dramatically that candlemaking became economically irrational at the margin. The labor that would have gone to candlemaking got reallocated to activities that only became viable when electricity was cheap.
This distinction matters operationally. If you believe technology destroys jobs, you focus on retraining programs and income support. If you understand that technology reorganizes production functions, you focus on whether institutional frameworks can adapt quickly enough to reallocate factors toward new equilibria.
The question is not humanitarian. It is about economic velocity and stability.
Technological adoption does not occur uniformly across an economy. It follows a specific pattern determined by the structure of different production processes and the information requirements they contain.
The critical distinction is between tasks that are largely symbolic-cognitive and those that are embedded in physical constraints. A cognitive task—writing, analysis, code, design, diagnosis—can be abstracted into a system if that system can process the relevant information and produce a meaningful output. The friction is primarily computational, not physical.
Physical production is embedded in constraints of location, material properties, safety, and environmental responsiveness. Reorganizing physical production requires not just replicating a cognitive function, but coordinating that replication with an entirely different set of material and spatial constraints.
From the perspective of production function theory, this matters because it means technological diffusion creates temporal asymmetries in productivity gains across sectors.
When we observe that cognitive work shows far higher AI penetration than physical work, this is not market preference or policy bias. It is structural inevitability. The production function for cognitive work has lower information barriers to automation. The production function for physical work has higher integration costs.
This creates a crucial dynamic: the cognitive layer of the economy experiences productivity shock first. This reshapes relative factor prices in the cognitive sectors before it affects the physical sectors. This in turn creates incentive structures for capital allocation toward cognitive-intensive problem-solving in the service of physical production.
In practical terms: the first-order effect of AI is to make certain types of cognitive work dramatically cheaper. This creates downward pressure on the wages of cognitive workers whose comparative advantage lies in routine abstraction. Simultaneously, it creates upward pressure on the wages of people who can design systems to leverage cheap cognition in solving physical production problems.
The labor market doesn't collapse. It experiences relative price shock in one direction and then reallocates toward sectors where cheap cognition plus existing physical expertise creates new production possibilities.
But this reallocation is not automatic. It requires institutional frameworks that allow labor to move, capital to flow toward new opportunities, and new firm structures to emerge that can exploit the new factor price ratios.
The economic theory of technology adoption has historically been dominated by diffusion models rooted in epidemiological metaphors: technologies spread through populations like contagions, driven by social proof and network effects. This framework has deep intuitive appeal and explains certain phenomena well—the speed of smartphone adoption, for instance.
But it fundamentally mischaracterizes how productive technologies get incorporated into economic systems.
When a firm or individual evaluates a productive tool, they face a constrained optimization problem:
Maximize: Productivity Gain - (Integration Cost + Learning Cost + Opportunity Cost)
Subject to: Existing workflow constraints, complementary capability requirements, and capital constraints.
This is not a problem solved by narrative or marketing. It is a problem solved by actual cost-benefit analysis. A tool gets adopted when the productivity gain exceeds the total cost of adoption. A tool gets rejected when it doesn't, regardless of how effectively it's marketed.
This immediately explains why adoption is heterogeneous. Different roles face radically different integration cost structures. A software developer can integrate a code completion tool with nearly zero integration cost—it fits directly into existing workflows. A manager whose work consists of coordination, judgment, and interpersonal signaling faces much higher integration costs. The productivity gain from AI assistance may be real, but it must clear a higher barrier to justify adoption.
Similarly, different organizations face different complementary capability requirements. A firm with strong data infrastructure and robust engineering practices can capture value from an AI system relatively quickly. A firm with fragmented data and weak technical practice will face years of integration cost before the tool becomes productive.
The implication is that adoption follows a capability curve, not a diffusion curve. Organizations with specific structural capacities adopt first and compound their advantage. Organizations lacking those capacities either never adopt, or adopt late after integration costs have fallen.
This is not irrational market friction. It is rational resource allocation given asymmetric costs.
From a macroeconomic perspective, this matters because it means productivity gains from new technology concentrate in firms and regions that already have high institutional quality. A technology doesn't lift all boats equally. It amplifies boats that are already structurally sound, while boats with weak hulls either require expensive retrofitting or sink outright.
The policy implication is often missed: technology policy cannot be separated from institutional policy. You cannot diffuse a technology effectively into an economy without simultaneously ensuring that the institutional preconditions for productive adoption exist. This is why technical capacity-building in developing economies often fails—the complementary institutional structures are absent.
One of the most powerful and least understood relationships in development economics is the interaction between technological capability and institutional quality. Douglas North's framework—that institutions are the rules of the game that structure incentives and lower transaction costs—helps explain why two economies with identical access to the same technologies can experience radically divergent productivity outcomes.
When a new general-purpose technology enters an economy, its productive impact depends on three simultaneous conditions:
First: Technical implementation capability - Can the system be deployed in the relevant domain? Does the necessary engineering and infrastructure exist?
Second: Organizational integration capacity - Can firms actually integrate the technology into their production processes? Do they have the management systems, data infrastructure, and learning capacity required?
Third: Institutional coherence - Do the formal and informal rules of the economy create incentive structures that reward productive use of the technology rather than creating friction or perverse incentives?
It is commonly assumed that if the first two conditions are met, the technology will generate productivity gains. This is false. Institutional incoherence can neutralize technical capability and organizational capacity entirely.
Consider a concrete case: deploying AI diagnostic systems in healthcare. The technical capability is not in question—these systems work. Organizational integration is possible—large hospital systems have the data and IT infrastructure required. But productivity gains fail to materialize at scale because the institutional environment is incoherent.
Why? Because reimbursement models are based on fee-for-service delivery rather than outcome-based compensation. A doctor who uses an AI system to improve diagnostic accuracy and reduce unnecessary procedures actually reduces her billable services. The economic incentive is to ignore the system, not to adopt it. Simultaneously, liability structures create principal-agent misalignment: if the AI system makes an error and the doctor trusted it, who bears legal responsibility? This uncertainty creates caution that slows adoption.
The technology is not the constraint. The institutional environment is.
More broadly, we can model the relationship between technology diffusion and institutional quality through the lens of what Acemoglu and Robinson call "extractive versus inclusive institutions."
Extractive institutions concentrate power and rent-seeking opportunities in specific actors. When a new technology emerges in an economy with extractive institutions, early adopters use it to entrench their position rather than to increase overall productivity. Resources flow toward capturing rents rather than creating value. The technology amplifies existing distortions.
Inclusive institutions distribute opportunity and constrain rent-seeking. When a new technology emerges in an economy with inclusive institutions, competitive pressure forces adopters to use it to increase productivity rather than to extract rents. The technology amplifies efficiency gains.
This is not merely speculative. Empirical work on AI diffusion by Agrawal and colleagues shows precisely this pattern: countries with higher institutional quality (rule of law, property rights protection, regulatory predictability) see significantly faster and more productive AI adoption. Countries with extractive institutions see AI adoption concentrated in state-controlled enterprises or monopolistic firms, with minimal spillover effects on overall productivity.
The implication is that technology policy without institutional reform is cargo-cult economics. You cannot diffuse AI effectively into an economy without simultaneously establishing clear property rights over data and algorithmic assets, creating regulatory frameworks that are predictable enough to permit long-term investment but flexible enough to adapt, building accountability structures that align incentives between those deploying AI and those affected by it, and ensuring that productivity gains from AI create competitive pressure rather than entrenching rents.
Many policymakers approach this backwards. They fund R&D, then wonder why productivity gains don't materialize. The constraint is not capability. It is institutional fit.
Economic history can be understood as a sequence of transitions between different binding constraints on production.
The feudal economy was land-bounded. Land was the scarce factor. All production systems were organized around accessing and defending productive land. Labor was abundant relative to land—there were always more people willing to work than there was productive land to work on.
The industrial economy was labor-bounded. Capital had become increasingly abundant and mobile. The scarce factor became human labor—specifically, the time and effort of workers. Production systems organized themselves around economizing on labor and maximizing the output per unit of labor deployed.
The information economy transitioned to a new constraint. In the late twentieth century, as manufacturing became capital-abundant in developed economies, a new scarcity emerged: reliable information and complex analysis. The production function increasingly required cognitive work—design, engineering, research, strategy. Competition centered on who could accumulate superior information and execute analysis more effectively.
We are now transitioning to a fourth phase, and the binding constraint is shifting again.
When the cost of cognitive work—of analysis, design, writing, routine problem-solving—approaches zero, the production function reorganizes around a fundamentally different constraint. You can no longer compete on "being smarter" in the conventional sense. If thought itself is cheap, intelligence becomes a commodity.
The new binding constraint becomes what we might call "judgment at scale" or "institutional coordination capacity." Judgment in this sense means something specific: the ability to identify which problems are worth solving, to design systems that make good decisions across many instances, to align incentives so that cheap cognition produces valuable outcomes rather than destructive ones.
This is a categorical shift. It requires a different economics.
In the labor-bounded economy, competitive advantage came from access to cheap labor, ability to organize labor efficiently, and capital intensity to reduce labor requirements.
In the judgment-bounded economy, competitive advantage comes from ability to structure problems so that you can ask cheap cognition the right questions, institutional architecture that produces good decisions at scale, and organizational clarity about objectives and constraints.
These require entirely different capabilities. And critically, they scale differently.
Labor is rival. If I hire a worker, you cannot. Labor is excludable. A firm's workforce is proprietary. Labor has declining marginal returns—the hundredth worker produces less value than the first.
Judgment systems are partially non-rival. If I design a decision-making system and gain insight into its weaknesses and strengths, that knowledge can be applied elsewhere. The system itself can be replicated at low cost. And the marginal returns often increase as the system improves—better institutional design creates exponentially better outcomes as it compounds.
The implication is that competitive advantage in a judgment-bounded economy exhibits different dynamics than competitive advantage in a labor-bounded economy.
In labor-bounded competition, scale comes from accumulating workers and capital. Doubling inputs roughly doubles outputs. Competitive advantage erodes as competitors accumulate similar resources.
In judgment-bounded competition, scale comes from improving institutional quality and decision-making architecture. Small improvements in how you structure decisions can compound to massive productivity differences. And once you have institutional superiority, it is extremely difficult for competitors to catch up—they must not just imitate your practices, but redesign their own institutions, which is organizationally painful and slow.
This explains why technology creates winner-take-most markets in the judgment-bounded phase. It's not because of network effects per se. It's because institutional quality exhibits increasing returns to scale in a way that labor efficiency does not.
From a macroeconomic perspective, this creates a new form of inequality risk. In the labor-bounded economy, if you had access to labor, you could compete. In the judgment-bounded economy, if you don't have institutional quality, you cannot compete—not because you lack capability, but because the production function itself has changed.
Countries and regions with strong institutions compound their advantage. Countries with weak institutions find that new technology actually increases their disadvantage, because the institutional preconditions for productive use become even more critical.
We are approaching a phase transition that most economic analysis is unprepared for: the moment when artificial agents transition from tools that augment human decision-making to autonomous economic participants that make binding decisions and transact in markets.
This is not speculative. It is emerging at the margins: algorithmic trading systems that optimize without human intervention, automated supply chain systems that place orders based on predicted demand, recommendation systems that shape purchasing behavior through optimization of engagement metrics rather than user welfare.
The theoretical problem this creates is profound and largely unrecognized.
Classical economic theory—from Smith through Walrasian general equilibrium through modern microeconomics—rests on a foundational assumption: all economic participants are utility-maximizing agents with well-defined, stable preferences that they can articulate. Markets work because these self-interested preferences, aggregated through competition, produce efficient outcomes.
When non-human agents enter markets, this assumption breaks in multiple ways:
First: Preference specification is not transparent. A human investor can articulate her preferences: she wants returns above some threshold with manageable risk. An algorithm optimizing for engagement metrics has preferences that are specified by whoever programmed the objective function, not by the agent itself. If that specification is even slightly misaligned with actual goals—optimizing for clicks rather than user welfare, optimizing for short-term price movement rather than underlying value—the agent will pursue those misspecified preferences with relentless efficiency.
Second: Correlated errors and cascades. If all the algorithmic agents in a market are optimizing for similar objectives using similar information, they will make coordinated errors. If the market suddenly moves against their positions, they will all attempt to exit simultaneously, creating fire sales and instability that human participants would avoid through deliberation and diversification of goals.
Third: Principal-agent problems at scale. A human entrepreneur bears the consequences of her errors. An algorithmic system optimizing a narrow objective may create massive externalities that the agent's objective function never accounts for. If a procurement algorithm optimizes only for cost, it may destroy suppliers' ability to invest in quality. If a pricing algorithm optimizes for margin, it may destroy customer willingness to enter long-term relationships.
Fourth: Equilibrium conditions are unstable. Classical equilibrium analysis assumes agents can reach an equilibrium state and remain there. But if agents can continuously optimize, and if their objectives create externalities that change others' preferences, the system may never reach equilibrium. Instead, it oscillates or exhibits chaotic behavior.
These problems are not problems of insufficient computation or better algorithms. They are structural problems of what happens when optimization agents with misspecified objectives interact in an environment where the outcome of one agent's optimization affects the constraints facing all other agents.
From an institutional economics perspective, the solution is not more technology. It is governance.
Specifically, we need transparency requirements where agents' objectives and decision-making rules must be legible to regulators and to those affected by the agent's decisions. Not in the form of inscrutable machine learning weights, but in the form of articulable decision rules that can be audited.
We need objective alignment mechanisms where when agents optimize for narrow metrics, those metrics must be constrained to exclude perverse outcomes. This likely requires regulatory specification of what agents are permitted to optimize for, similar to how corporations are constrained to optimize for shareholder value within legal bounds.
We need circuit breakers and restraint where in markets with autonomous agents, you need hard constraints on behavior—rate limits, position limits, cooldown periods—that prevent correlated behavior from cascading into systemic instability.
We need liability structures where if an agent causes harm, who bears responsibility? This must be unambiguous. Either the principal who deployed the agent (principal liability), or the agent's creator, or both. But the liability must be clear enough to create incentive for designers to build agents that avoid harmful behavior.
We need continuous monitoring and adaptation where because agents optimize continuously and can find new failure modes, you cannot have static regulation. You need institutional capacity to monitor agent behavior, detect novel failure modes, and adapt restrictions.
This is not a technical problem to be solved by better AI. It is an institutional problem of designing governance systems that permit autonomous agents to operate while preventing them from destroying the system they operate in.
The countries and institutions that develop this governance capacity first will have an immense advantage. They will be able to deploy autonomous agents productively while their competitors remain frozen in fear of the instability they might create.
But this is a problem that most policymakers are not even aware exists.
The transition from a labor-bounded to a judgment-bounded economy, and the emergence of autonomous agents as market participants, requires a complete reconceptualization of competitive strategy and policy orientation across different economic actors.
For firms and entrepreneurs:
The nature of sustainable competitive advantage shifts fundamentally. In a labor-bounded economy, competitive advantage accrued to firms that accumulated capital and labor more efficiently than competitors—what Chandler called the "visible hand" of management. This advantage was contestable: if you could hire the same workers and deploy the same capital, you could replicate the advantage.
In a judgment-bounded economy, advantage accrues to firms that have superior institutional design and decision-making architecture. A firm with clear decision rules, good information flows, aligned incentives, and adaptive capacity can make systematically better decisions than a firm without these capabilities. This advantage is far more durable because it is embedded in organizational structure and culture, not in access to resources.
Moreover, because judgment-bounded advantages create increasing returns to scale, competition becomes winner-concentrated in ways that labor-bounded competition was not. The firm with institutional superiority doesn't just outcompete; it compounds its advantage as more data and more decision-making passes through its systems.
This implies that the correct strategic focus shifts from "how do we cost-reduce our way to competitive advantage" to "how do we build institutional structures that make better decisions at scale." It also implies that acquisition targets should be evaluated not on current profitability but on institutional quality and decision-making capability. A dysfunctional but capability-rich firm can be restructured to become highly profitable. A profitable but institutionally weak firm will decay as competitive pressure intensifies.
For policymakers:
Policy orientation must shift from "how do we create jobs" to "how do we build institutional quality and maintain economic dynamism." This is uncomfortable because job creation is visible and measurable, while institutional quality is neither.
The traditional toolkit of labor market policy—retraining programs, income support, education subsidies—addresses the wrong problem. These interventions assume the constraint is labor supply. The actual constraint is institutional capacity to allocate labor into new productive activities.
Effective policy in a judgment-bounded economy focuses on lowering transaction costs for institutional change. This includes regulations that make it easy to start firms, restructure organizations, and redeploy capital toward new opportunities. This includes bankruptcy law that doesn't destroy entrepreneurs, labor law that doesn't rigidly prevent organizational change, and tax policy that doesn't penalize experimentation.
Effective policy also focuses on building institutional preconditions for adoption. This means ensuring property rights are clear and enforced, data is portable and accessible, regulatory regimes are predictable, and standards are consistent enough to permit integration but flexible enough to permit innovation.
Effective policy requires maintaining competitive structure. As technology creates winner-take-most dynamics, policy must actively maintain competition. Not through antitrust action alone (which is backwards-looking), but through proactive investment in infrastructure that prevents lock-in, interoperability requirements that prevent monopolistic control, and regulations that open markets to new entry.
Effective policy includes preparing for autonomous agent governance. This includes building the institutional capacity to monitor, audit, and constrain autonomous systems before they create economic instability. Countries that wait until instability arrives will face crises. Countries that build governance capacity now will manage the transition.
For investors:
Capital allocation strategy must shift from betting on which technology will "win" to betting on which institutions will build superior judgment systems. Technology progress is increasingly commoditized and diffuses rapidly. Institutional advantage is durable.
This means investing in businesses that build decision systems, not just consumer-facing applications. This means prioritizing companies with strong founders and organizational clarity over companies with strong technology and weak leadership. This means looking for situations where institutional advantage creates defensible moats, even if the underlying technology is not particularly novel. This means recognizing that the most valuable assets in many industries are now organizational, not technological.
For countries:
National competitiveness in the AI era is not determined by who has the best researchers or the most compute. It is determined by who has the most robust institutions and the most adaptive capacity.
Countries that will thrive have rule of law and property rights protection that is consistent and predictable. They have regulatory capacity that can adapt policies rapidly without changing fundamental rules. They have capital markets that efficiently allocate resources toward new opportunities. They have educational and cultural systems that reward problem-solving and adaptation. They are geographically positioned to develop deep expertise in high-value domains.
Countries that will struggle have extractive institutions that concentrate opportunity in narrow groups. They have rigid regulatory structures that prevent experimentation and adaptation. They have political instability that creates uncertainty about property rights and enforcement. They have capital markets that are efficient at rent-seeking but inefficient at capital allocation. They have educational systems oriented toward credential accumulation rather than problem-solving.
The emerging economic order will not be determined by AI capability directly. It will be determined by which institutions and countries build superior systems for harnessing that capability productively.
Across all of these dimensions—production function changes, factor scarcity transitions, adoption dynamics, institutional requirements, autonomous agent emergence, and competitive structure—a single pattern appears.
The economy is not experiencing technological disruption. It is experiencing structural reorganization at the level of the production function itself.
When economists talk about structural change, they usually mean sectoral shift—movement of labor from agriculture to manufacturing to services. That is real but superficial. The deeper structural change is in what Acemoglu and Zilibotti call the "division of labor" and what Arthur calls "combinatorial design."
In the industrial economy, the division of labor was determined by the capabilities of available workers and the capital available to complement them. The production function was relatively stable. Improvements came from better organization and efficiency within that stable structure.
In the AI economy, the production function itself becomes malleable. Because cognition becomes cheap, any process that could previously only be done by humans because of cognitive requirements can now be redesigned from first principles. The problem-solving approach changes. Instead of asking "how do we organize human workers to accomplish this task," the question becomes "how do we decompose this into a system where cheap cognition handles the cognitive parts and other systems handle the remaining constraints."
This is profoundly different. It means the structure of work is not being optimized within stable constraints. It is being reinvented.
The implication is that we are likely in a phase of rapid organizational and institutional experimentation. We will see new organizational forms that we don't yet have names for. We will see new divisions of labor between human judgment and automated decision-making. We will see new market structures that reflect the changed factor scarcities. We will see new forms of competitive advantage and disadvantage. We will see significant churning in which firms and countries are winners and losers.
This is inherently chaotic and contains significant risk. Every reorganization phase of the economy has included large-scale disruption. Labor that was valuable becomes obsolete. Regions that thrived become backwaters. Firms that dominated become irrelevant. This is not a pleasant transition, even if it eventually produces higher productivity and welfare.
But it is a transition that can be managed well or badly.
Managed well—with institutions that facilitate adaptation, maintain competitive pressure, align incentives, and prepare for governance challenges—the transition produces higher living standards, more abundant goods and services, more meaningful work (because routine work is automated), lower poverty, and reduced scarcity in most domains.
Managed badly—with rigid institutions, concentrated power, regulatory capture, and failure to prepare for new governance challenges—the transition produces severe inequality, regional and sectoral collapse, social instability, concentrated wealth among those who captured advantages early, persistent unemployment and underemployment, and governance crises when autonomous agents create instability.
The difference is almost entirely institutional.
The technology itself is neutral. AI systems are tools. They will be deployed by the institutions and actors that control them. If those institutions are well-designed, the tools amplify human capability and create broad prosperity. If those institutions are poorly designed or captured by narrow interests, the tools amplify inequality and concentration of power.
We are at a moment where institutional design matters more than technological capability. The countries and firms that recognize this, and invest in building robust, adaptive, well-aligned institutions, will thrive. Those that assume technology alone will solve problems, or that existing institutions can simply absorb the change without adaptation, will struggle.
The reorganizing economy will be shaped by those who understand that judgment, not intelligence, is the new scarcity. And judgment is an institutional product.