Delegation Is the Novelty

I. Introduction

The most reassuring argument against alarm about artificial intelligence is the historical one. It runs as follows. The textile mills of late-eighteenth-century England were predicted, by intelligent and informed observers, to abolish the labor of the spinner and the weaver. They did abolish those particular forms of labor, but they did not abolish labor; they generated, on a timescale of two or three generations, more employment than they destroyed, distributed across a larger and more productive economy. Electrification was predicted to dissolve the structure of the industrial firm, and indeed it eventually did so, but on a timescale of nearly four decades, during which the institutions of the firm, the union, the regulatory state, and the financial system reorganized themselves around the new technology. The mainframe computer of the 1950s was predicted to displace clerical labor en masse; it did displace specific clerical functions, but it generated a vast new domain of information work that absorbed and surpassed the displaced labor. In each case, the technology was absorbed by institutions that adapted, sometimes painfully but ultimately successfully, to its presence. The pattern is sufficiently regular that it has acquired the status of a historiographic commonplace: the predictions of dislocation are routinely correct in direction and routinely wrong in magnitude and timing.¹

If the pattern is regular, the inference is irresistible. Proactive AI — by which I mean throughout this paper systems that initiate action on their own, monitor state continuously, and operate on schedules or triggers without per-instance human authorization — is the latest entry in this sequence, and the prudent expectation is that it will be absorbed in roughly the same way. There will be dislocation. There will be institutional turbulence. But the deep structures of economic, political, and legal life will adapt, on a timescale measured in decades, and the outcome will look in retrospect like the outcomes of the preceding waves: more productivity, reorganized work, expanded economic activity, and institutions that look continuous with their predecessors. On this view, the contemporary alarm about AI rests on a failure of historical perspective. The predicted dislocations are real, but the predicted civilizational rupture is the latest instance of a recurring overestimation that we have ample reason, by induction, to discount.

This paper takes the historical-precedent counterargument seriously. It is the most powerful argument against the position developed in the broader research program of which this paper is a part — the position that proactive AI is an institutional actor rather than an instrument, and that its accommodation by existing legal and political institutions will be genuinely difficult rather than routinely absorptive. If the historical-precedent argument is correct, the broader thesis is wrong. The argument deserves the full weight of comparative-historical scholarship, not the easy dismissal of "this time is different" claims that, as historians of finance have shown, are reliably wrong (Reinhart and Rogoff 2009).

The paper argues that the historical-precedent counterargument nonetheless fails, but for a precise reason that does not generalize to other "this time is different" claims. The reason is that proactive AI is not the latest entry in the sequence of automation waves in the relevant sense. It is the first entry in a different sequence. Prior automation waves automated production (mechanization), power transmission (electrification), and information processing (computerization). They did not automate delegation — the transfer of authority to act on behalf of a principal. Proactive AI does automate delegation, and the institutional adaptations required for delegation-automating technology are not the adaptations the historical record gives us evidence about. The historical-adaptation literature is not wrong about the prior waves; it is wrong in its assumption that proactive AI belongs to the same category.

The argument turns on a definitional move that I want to motivate at the outset. The word "automation" is, in ordinary usage, a heterogeneous category. It covers the automation of physical motion (the loom, the assembly line, the robotic arm), the automation of energy flow (the steam engine, the electric motor, the power grid), the automation of symbolic operations (the abacus, the calculator, the spreadsheet), and what I will here distinguish as the automation of authorization — the granting of authority to act without per-instance invocation. The first three forms of automation share a structural property: the human principal must invoke the technology in order for the technology to do anything. A loom does not weave until thrown; a power tool does not drill until triggered; a spreadsheet does not compute until invoked. The technology is potential; the action is actual only when commanded. The fourth form of automation — the delegating form — operates on a different principle. The human principal grants a scope of authority, and the system acts within that scope without per-instance command. The relationship is not invocation; it is authorization. The system is, in the relevant sense, an agent rather than an instrument.

The historical record gives us extensive evidence about institutional adaptation to invocation-based automation. It gives us almost no evidence about institutional adaptation to authorization-based automation, because authorization-based automation has, until recently, been confined to narrow domains (the autopilot, the thermostat, the automated trading algorithm) operating under closely circumscribed scopes that resemble the engineering of safety-critical control systems more than they resemble general delegation. The contemporary moment is the first in which authorization-based automation is becoming a general technology, applicable across the breadth of activities that institutions delegate to human agents. The institutional adaptations required by this transition are not the institutional adaptations of the prior waves, and the comfortable inference from the prior waves' adaptive trajectories is therefore not available.

The paper proceeds in eight further Parts. Part II develops the definition of delegation and distinguishes it from automation with the precision the rest of the argument requires. Part III conducts a comparative-historical analysis of three prior automation waves — mechanization, electrification, computerization — examining each for delegation-relevance. Part IV treats the most important partial exception: the rise of the modern corporate hierarchy in response to the scale demands of the railroad, as analyzed in Chandler's Visible Hand (1977). Part V specifies what is genuinely new about proactive AI as a delegation-automating technology. Part VI engages with Acemoglu and Johnson's Power and Progress (2023), with attention both to the alignments and the disagreements. Part VII generates four falsifiable predictions. Part VIII addresses three counterarguments, including the most pressing one: the claim that delegation has always been done by humans-via-tools and that proactive AI merely shifts the tool–agent ratio without crossing a categorical boundary. Part IX concludes with the broader stakes for institutional design.

A note on scope. The paper addresses the question of categorical novelty, not the question of normative valence. Whether proactive AI is good or bad on net, and under what institutional conditions, is the subject of the companion papers in this research program. The argument here is the prior one: that the institutional question — what adaptations are required — cannot be answered by analogy to the prior waves, because the prior waves were not adapting to a delegation-automating technology. The normative argument can be developed only after the categorical question is settled.

II. What Is Delegation?

The argument depends on the definitional distinction between delegation and automation. The distinction is intuitive but, in the existing literature, surprisingly underdeveloped. The economics of automation has been concerned almost entirely with the substitution of capital for labor and the corresponding effects on productivity, wages, and distribution; it has had little reason to develop a vocabulary for the species of automation in which the substituted-for activity is itself the activity of acting on behalf of another (Autor 2015; Acemoglu and Restrepo 2018, 2022). The principal–agent literature, by contrast, has developed a rich vocabulary for delegation but has assumed that the agent is a human or, at most, a firm composed of humans, and has had little reason to address the case in which the agent is a non-human system (Holmström 1979; Holmström and Milgrom 1991; Tirole 1986). The intersection — automation as it applies to delegation, or delegation as it operates through non-human agents — is the space this paper occupies.

I will define the terms with the precision the argument requires.

A. Automation

By automation I mean the substitution of mechanical, electrical, or computational process for human action in the performance of a defined task. The substitution is task-level. The human principal identifies a task, configures the automating technology to perform that task, and invokes the technology to perform the task at a particular moment or in response to a particular trigger. The substitution does not extend beyond the task. The loom weaves; it does not decide whether to weave, when to weave, what to weave, or for whom to weave. Those questions are decided by the human principal and communicated to the loom through invocation. The loom is an instrument; the human is the principal.

This characterization holds across the prior waves of automation. The mechanical loom of the 1780s substituted machine motion for human muscle in the performance of weaving, but the decisions about whether to weave, what to weave, and for whom remained with the mill owner. The electric motor of the 1890s substituted electrical for mechanical power transmission in the performance of factory work, but the decisions about what to produce and how to organize production remained with the manager. The mainframe computer of the 1950s substituted symbolic calculation for human calculation in the performance of accounting, payroll, and inventory tasks, but the decisions about how to organize the firm and which calculations to perform remained with the controller. The technology, in each case, was an instrument of vast power. It did not exercise authority.

B. Delegation

By delegation I mean the transfer of authority to act on behalf of a principal within a specified scope. The transfer is not task-level but scope-level. The principal does not specify the particular actions to be taken; the principal specifies a scope of permitted action and grants authority to act within that scope. The agent, having received the grant, acts on the agent's own initiative within the scope, without per-instance invocation by the principal. The agent's actions are imputed to the principal as a matter of legal, organizational, and moral attribution, subject to the conditions the law of agency specifies (Restatement (Third) of Agency §§ 1.01, 2.01; see also Watts 2020).

The relationship is constituted by three features. First, scope: the principal demarcates a domain of permitted action, which may be defined by task, by sphere, by counterparty, by budget, by time, or by some combination. Second, initiative: the agent decides when to act within the scope, without consulting the principal at each instance. Third, attribution: the agent's actions are treated, by relevant institutional vocabularies (legal, organizational, moral), as actions of the principal, with appropriate limitations on the principal's liability where the agent has acted outside scope or in breach of duty. The presence of these three features distinguishes delegation from instruction. A manager who specifies each action a subordinate is to take is not delegating; the manager is using the subordinate as an instrument. A manager who specifies a scope and grants the subordinate authority to act within it is delegating. The difference is the locus of initiative.

The economics of organizations, the law of agency, and the political-theoretic literature on representation all converge on essentially this conception of delegation, with variations in vocabulary. The economics literature emphasizes the information-asymmetric and incentive-misalignment features of the relationship (Jensen and Meckling 1976; Tirole 1986). The legal literature emphasizes the attribution rules and the duties of loyalty and care (Restatement (Third) of Agency; DeMott 2006; Easterbrook and Fischel 1991). The political-theoretic literature emphasizes the legitimacy conditions under which the principal's authorization is binding on the agent and the agent's actions are binding on third parties (Pitkin 1967; Mansbridge 2003). The conceptions differ in emphasis and in their normative attachments, but they share the structural account: delegation is the transfer of authority, marked by scope, initiative, and attribution.

C. The Distinction

The distinction between automation and delegation can now be stated with precision. Automation is the substitution of mechanical or computational process for human action under instruction. Delegation is the transfer of authority to act on behalf of a principal with initiative. The two are not points on a continuum; they are different relations. Automation answers the question "who will perform this action?" Delegation answers the question "who is authorized to decide what action to perform?" An automation relation can exist without a delegation relation (a power tool used by its owner). A delegation relation can exist without an automation relation (a human agent acting on behalf of a human principal). The two relations can co-exist in a single configuration (a human employee using a power tool on the firm's behalf), but they remain conceptually distinct.

The argument of this paper is that proactive AI, in the strong sense of the term, is the first general technology that performs the delegation relation rather than the automation relation. The system is configured with a scope rather than an instruction. The system acts on its own initiative within the scope. The system's actions are attributed to the deploying principal under emerging configurations of contract, organizational responsibility, and (in some jurisdictions) statutory rule.² These are the features of delegation, not of automation. The technology is, in the precise sense, an agent — though, as Part V will develop, an agent of an unprecedented kind.

The distinction has the analytical work of the paper to do. If automation and delegation are different relations, then the historical record of institutional adaptation to automation does not, without further argument, license inferences about institutional adaptation to delegation. The further argument would have to show that the prior waves' adaptations were in fact delegation-relevant — that they included, alongside the adaptation to invocation-based technology, the development of new authority structures, new scope-granting practices, new attribution rules. The next two Parts conduct that examination. The honest finding is that the prior waves' adaptations were, with one important exception, not delegation-relevant in the relevant sense. The exception (the Chandler corporate hierarchy) is delegation-relevant but does not generalize in the way the historical-precedent argument requires.

D. Why the Distinction Matters

A skeptic might grant the distinction in principle but doubt that anything important hangs on it. Surely, the skeptic might say, every technology is on some continuum from pure instrument to pure agent, and proactive AI just sits further along the continuum than its predecessors; the categorical framing is a rhetorical move rather than a substantive one. The reply is that the institutional vocabularies appropriate to instruments and to agents are structurally different. Instruments are governed by the laws of property, products liability, and contract. Agents are governed by the laws of agency, fiduciary duty, and (where the principal is a state) public law. The vocabulary appropriate to an instrument addresses the relationship between the instrument's owner and the instrument's effects on third parties. The vocabulary appropriate to an agent addresses the relationship between the principal, the agent, and the third parties to whom the agent's actions are attributed. These vocabularies have different conceptual structures, different historical developments, and different conditions of application. When a technology shifts from instrument to agent, the applicable vocabulary changes, and the institutional adaptations required are the adaptations of the new vocabulary, not of the old.

This is not a rhetorical observation. It is a doctrinal one. The most fundamental disputes in the contemporary legal scholarship on AI — whether to treat AI systems under products liability or under agency, whether to recognize a fiduciary relation between AI deployers and the third parties they affect, whether to permit corporate forms that include AI agents — are disputes about which institutional vocabulary applies (Bryson, Diamantis, and Grant 2017; Chopra and White 2011; Lior 2020; Lemley and Casey 2019). The disputes are not, in the main, disputes about facts; they are disputes about categorization, and the right categorization determines which body of doctrine governs. If proactive AI is correctly categorized as an instrument, the products-liability framework, with its accumulated wisdom about defective products and inadequate warnings, is the right starting point. If proactive AI is correctly categorized as an agent, the agency framework, with its accumulated wisdom about scope of authority and breach of duty, is the right starting point. The two frameworks generate different liability rules, different remedial structures, different procedural conventions, and different institutional configurations. The categorization question is not merely a labeling question.

The argument of this paper is that the categorization question must be answered in favor of agency, and that the historical-adaptation literature has been applying the wrong category. The prior waves were correctly handled within the instrument framework, and the institutional adaptations to those waves were adaptations within the instrument framework. The current wave is being mis-handled by the application of the same framework. The reframing requires the comparative-historical work the next three Parts conduct.

III. Three Prior Waves

To assess the historical-precedent counterargument requires examining the prior waves with sufficient specificity to determine whether their institutional adaptations were, in fact, delegation-relevant. I take three waves: mechanization in the late eighteenth and early nineteenth centuries, electrification in the late nineteenth and early twentieth centuries, and computerization beginning in the 1950s. For each I ask four questions. What was automated, precisely? What institutional adaptations were required? Were any of those adaptations delegation-relevant in the sense developed above? How long did adaptation take?

A. Mechanization (c. 1770–1840)

The mechanization of textile production in late-eighteenth-century England is the canonical case of the first modern automation wave. The relevant inventions — Hargreaves's spinning jenny (1764), Arkwright's water frame (1769), Crompton's mule (1779), Cartwright's power loom (1785) — substituted mechanical motion for the human muscle of the spinner and the weaver, in conjunction with the centralization of work in the factory and the deployment of water and (later) steam power as the energy source for the mechanical motion. The substitution was, in the terms developed above, an instance of pure automation. The machines performed weaving and spinning under the instruction of the mill owner and the foreman; they did not exercise initiative. The mill owner decided what to produce, in what quantity, for what market, at what specification; the machines executed the production.

What institutional adaptations did mechanization require? The economic-history literature has documented a vast set: the formation of the factory as an organizational form distinct from the workshop, with implications for the structure of labor markets and the spatial organization of production (Mantoux 1928; Pollard 1965; Landes 1969); the development of factory discipline and time-discipline, with the corresponding transformation of work culture (Thompson 1967); the emergence of factory legislation regulating working hours and conditions, beginning with the Health and Morals of Apprentices Act of 1802 and extending through the Factory Acts of 1833, 1844, and after (Hutchins and Harrison 1903; Marvel 1977); the development of the trade union movement and the eventual legal recognition of collective bargaining (Webb and Webb 1894); the reconfiguration of family economic life and the eventual emergence of the male-breadwinner household as a normative model (Horrell and Humphries 1995); and, on a longer timescale, the development of the welfare state as a response to the dislocations of industrial society.

These adaptations were profound and consequential. They occupied the political and intellectual energy of British society for roughly a century. But they were not, in the precise sense, delegation-relevant. The adaptations were responses to the social facts that mechanization produced — concentration of workers in factories, separation of work from home, exposure of workers to industrial hazards, dependence of family income on the factory wage. They were not responses to a new authority relation between principal and agent. The mill owner's relation to the mill was the relation of an owner to property; the mill owner's relation to the workers was the relation of an employer to employees, with the workers as the agents and the machines as the instruments. The novel institutional vocabularies developed in response to mechanization — factory legislation, collective bargaining, the welfare state — addressed the social organization of labor in the new productive context. They did not address the question of how authority should be allocated between principal and non-human agent, because no such question arose: the machines were not agents.

The timeline of adaptation deserves emphasis. The first generation of textile mechanization is dated to the 1770s and 1780s; the institutional response is variously dated to the Factory Acts of the 1830s and 1840s (half a century later), the consolidation of trade-union rights in the 1870s and 1900s (a century later), and the welfare state of the post-1945 settlement (a century and a half later). The standard claim that automation waves are absorbed by institutions over decades is, in the case of mechanization, a substantial understatement: the absorption took generations, and the social conflicts during the absorption included the Luddite revolts, the Chartist movement, and the political mobilizations that produced the great extensions of the franchise. The institutions adapted. They did so on a timescale that, transposed to the contemporary case, would imply institutional absorption of proactive AI extending into the twenty-second century. Whether that timescale is acceptable, in light of the rate of contemporary capability development, is a question Part VII will address.

B. Electrification (c. 1880–1930)

The second automation wave, electrification, was for fifty years the paradigmatic example used by economic historians to argue that general-purpose technologies require long periods of complementary institutional and organizational change before their productivity effects are realized. David's 1990 paper, "The Dynamo and the Computer," is the canonical statement: the electric motor was introduced in industry in the 1880s, but the productivity gains attributable to electrification did not materialize until the 1920s, because the realization of those gains required the reorganization of factories from shaft-and-belt power distribution to unit-drive electric power distribution, the redesign of factory floor layouts to take advantage of decoupled power, and the development of new managerial techniques (Taylorism, scientific management, the early development of operations research) that exploited the flexibility electrification permitted (David 1990; see also Devine 1983; Du Boff 1979).

What was automated by electrification, precisely? The narrow answer is that electrification automated power transmission. The steam engine of the earlier industrial era produced concentrated mechanical power that had to be transmitted to the point of use through shafts, belts, and gears, with significant friction losses and severe constraints on factory layout. The electric motor permitted power to be generated centrally and transmitted as electricity to the point of use, where small electric motors could be deployed flexibly. The substitution was the substitution of an electrical for a mechanical transmission system. It was not, in itself, a substitution of machine action for human action — the machines being powered remained the same machines, performing the same operations under the same human instructions.

The broader answer is that electrification permitted a reorganization of production that, in conjunction with electrified specialized machinery and the new managerial techniques, substituted machine action for human action across a wider range of factory tasks. The assembly line is the paradigm: Ford's Highland Park plant of 1913 combined electrified power distribution, specialized single-purpose machinery, and the moving assembly line to substitute machine pacing for human pacing in automobile production (Hounshell 1984). The substitution was, again, instrument-substitution under human instruction. The assembly line did not decide what to produce, how fast to move, what specifications to meet. Those decisions were made by managers — Ford's managers, in this case, on the basis of Ford's strategic decisions — and communicated to the line through engineering specification and supervisory instruction.

What institutional adaptations did electrification require? The list is again extensive: the development of the modern industrial corporation as a vehicle for capital-intensive production (Chandler 1962, 1977 — to which I return in Part IV); the elaboration of scientific management as a discipline (Taylor 1911; Nelson 1980); the development of industrial unionism as distinct from craft unionism (Brody 1980); the construction of the regulatory state's relation to industrial production through antitrust, securities regulation, and labor law (Hovenkamp 1991; Berk 1994); the development of electric utility regulation, which gave rise to the modern regulatory commission as an institutional form (Hyman 1985); and, on a longer arc, the construction of the consumer society as the demand-side counterpart to mass production (Cohen 2003).

These adaptations were, like those of mechanization, profound, consequential, and not delegation-relevant in the precise sense. The new institutional vocabularies — corporate law and corporate governance, securities regulation, antitrust, industrial labor law, utility regulation — addressed the organization of large-scale industrial production within an existing legal framework that treated the corporation as a legal person and its officers and employees as the agents through which the corporation acted. The agents were human. The machines were instruments. The authority relations that the new vocabularies elaborated — shareholders to directors, directors to officers, officers to employees, employees to organized labor — were all human-to-human authority relations, mediated by capital structure, contract, and statutory regulation. Electrification did not require a new theory of authority transfer to non-human agents because no non-human agent was being authorized.

There is one partial qualification worth flagging here, which is that the regulatory commission as an institutional form — the Interstate Commerce Commission of 1887, the Federal Trade Commission of 1914, the Federal Power Commission of 1920, the Federal Communications Commission of 1934, and the state public utility commissions that proliferated in the same period — represented a delegation innovation within the structure of the state: legislatures delegated rule-making and adjudicatory authority to expert commissions on standing terms, rather than retaining that authority for case-by-case statutory determination. The development of the regulatory commission is a delegation innovation worth taking seriously, and the constitutional anxieties it provoked (Crowell v. Benson, 285 U.S. 22 (1932); Schechter Poultry, 295 U.S. 495 (1935)) prefigure some of the contemporary anxieties about delegation to AI systems. But the delegated authority was held by human commissioners, advised by human staff, deciding human-comprehensible cases. The structural analogy to proactive AI is suggestive but not tight: the regulatory commission delegated to an agent that could be sanctioned, deliberated, and held politically accountable in the ordinary way. Part VI returns to this point in engagement with Acemoglu and Johnson, who emphasize the political economy of delegation to expert institutions.

The timeline of electrification's institutional absorption was, by mechanization's standards, fast: the productivity gains materialized roughly forty years after the initial industrial deployment, and the institutional configurations associated with electrified mass production (the modern corporation, scientific management, industrial unionism, utility regulation) were largely in place by the 1930s. Forty years is, again, a standard against which the rate of contemporary AI capability development should be measured.

C. Computerization (c. 1950–2010)

The third automation wave, computerization, is the one whose institutional adaptations the contemporary AI literature most often invokes as precedent. The story is familiar. The first commercial mainframe computers — UNIVAC I (1951), IBM 650 (1953), IBM 1401 (1959) — automated batch-processing operations that had previously been performed by clerks: payroll, accounts receivable, inventory tracking, actuarial calculation. The minicomputer of the 1960s extended the technology to mid-sized firms and to a wider range of business processes. The personal computer of the late 1970s and 1980s extended it to the individual worker. The networked computer of the 1990s and 2000s integrated firms internally and externally, producing the institutional configurations of contemporary business: the integrated supply chain, the customer relationship management system, the enterprise resource planning system (Brynjolfsson and Hitt 2000; Bresnahan, Brynjolfsson, and Hitt 2002).

What was automated, precisely? The narrow answer is that computerization automated symbolic calculation — the manipulation of representations under formal rules. The accounting clerk who reconciled the day's transactions was performing symbolic calculation; the actuary who computed reserves was performing symbolic calculation; the inventory analyst who projected stock-out probabilities was performing symbolic calculation. The mainframe could perform these calculations faster, more accurately, and at vastly greater scale than the human clerks. The broader answer is that computerization automated information processing more generally — the gathering, organizing, transformation, and dissemination of representations — across an expanding range of business processes (Beniger 1986; Yates 1989).

The crucial fact about computerized information processing, for the purposes of this paper, is that it operated on instruction. The mainframe did not perform calculations until invoked by a job submitted through a punched-card deck or a batch script. The spreadsheet did not compute until the user entered the formula and pressed return. The enterprise resource planning system did not generate a purchase order until configured to do so by a workflow designed by human analysts. Even in the highly elaborate information ecologies of the 2000s, the technology operated under the principle that human action initiated the technology's action. The information environment changed; the locus of initiative did not.

What institutional adaptations did computerization require? The list extends across the economy. The reorganization of clerical labor and the elimination of large categories of routine office work (Autor, Levy, and Murnane 2003); the development of information technology as a distinct organizational function with its own professional identity (Friedman and Cornford 1989); the emergence of the consultant–implementer ecosystem around enterprise software (Edwards 2000); the disintegration of vertically integrated firms and the rise of contract-based production networks (Langlois 2003); the financialization of corporate strategy enabled by real-time financial information systems (Davis 2009); the regulatory response to computerized financial markets (e.g., NMS, Reg ATS, MiFID) and to the data-collection capacities of computerized organizations (the European Data Protection Directive of 1995 and its successors); and, on a longer arc, the development of cybersecurity as a domain of practice, policy, and law (Brenner 2010).

Were any of these adaptations delegation-relevant? The honest answer is: partially, in specific subdomains, but not in the central sense the argument requires. The clearest delegation-relevant adaptation is in financial markets, where the development of algorithmic and (later) high-frequency trading required new rules about the attribution of trades to firms, new structures of risk management, and new regulatory categories (the "sponsoring broker," the "naked access" prohibitions, the eventual development of pre-trade risk controls under SEC Rule 15c3-5).³ The financial-markets case is genuinely a case of computerized delegation: the trading algorithm acts within a scope of authority granted by the firm, the firm bears the consequences of the algorithm's actions, and the law of the markets has had to develop attribution and supervision rules to manage the relation. This is, in the precise sense, delegation to a computerized agent. The case is exceptional within the computerization wave, however, and the regulatory response to it has been narrow — applicable to a specific category of financial actors, operating under specific market rules, with specific contractual structures. It is not a general theory of delegation to computerized agents; it is a specific accommodation in a high-stakes domain.

The broader computerization wave was not, in the main, a delegation wave. The enterprise software systems that reorganized business operations did not exercise initiative; they implemented workflows configured by humans, executed transactions authorized by humans, and produced reports interpreted by humans. The institutional adaptations associated with computerization were adaptations to a transformed information environment, not adaptations to a new form of authority transfer. The contemporary inference from computerization to proactive AI — that the institutions will adapt as they did to the spread of enterprise software, on a similar timescale and with similar mechanisms — is the inference the rest of this paper challenges. The mechanism of adaptation to computerization was the absorption of new instrumentation by existing principals; the mechanism of adaptation to proactive AI must be the development of new structures of authority that the existing principals can rely on. The two are different mechanisms.

IV. The Chandler Exception: Corporate Hierarchy and Delegation Innovation

The honest comparative analysis must take seriously the one prior episode that does involve a substantial delegation innovation: the rise of the modern multi-divisional corporate hierarchy in response to the scale demands of the railroad and, subsequently, the integrated mass-production firm. This is the story Alfred Chandler told in The Visible Hand (1977) and Strategy and Structure (1962), and refined by historians including Lazonick (1991), Roy (1997), and Lamoreaux, Raff, and Temin (2003). The Chandler story is the most pointed challenge to the argument of this paper, because it describes a technological wave (the railroad and the integrated industrial firm) whose institutional adaptation was, precisely and centrally, a delegation innovation. If the Chandler episode is taken seriously, the historical-precedent argument acquires its strongest form: there is at least one prior wave whose adaptation required new delegation structures, and the institutional development of that wave is therefore a relevant precedent for the proactive-AI wave.

The argument requires unpacking. Chandler's account is that the railroad, by virtue of its scale, its capital intensity, its safety requirements, and its operational complexity, was the first business enterprise whose effective management exceeded the cognitive and informational capacity of a single owner-manager. The traditional family-firm or partnership structure that had governed antebellum American business was not adequate to the management of a thousand-mile rail network operating on real-time schedules across multiple states with hundreds of employees and a continuous flow of safety-critical operational decisions. The institutional adaptation was the development of the salaried managerial hierarchy: a structure in which professional managers, employed by the firm rather than owning it, exercised delegated authority over discrete operational domains, reporting through tiers of intermediate managers to a small executive group that retained strategic authority. Chandler's claim is that this structure — the M-form corporation, in the later vocabulary of Williamson (1975) — was the foundational institutional innovation of modern American capitalism, and that it was driven by the operational demands of the railroad and subsequently extended to the integrated mass-production firm.

The Chandler story is, in the precise sense, a delegation story. The salaried manager exercises authority on behalf of the firm within a defined scope. The manager's actions are attributed to the firm. The manager acts on the manager's own initiative within the scope, without per-instance authorization from the firm's executive group. These are the constitutive features of delegation as developed in Part II. The institutional adaptations Chandler describes — the development of professional management as a distinct occupational category, the elaboration of corporate accounting to permit measurement of delegated performance, the construction of internal capital markets and internal labor markets within the firm, the development of corporate governance structures to monitor managerial agents on behalf of dispersed shareholders — are all delegation-relevant adaptations in the strong sense. The Chandler episode does require revision of the strong claim that prior automation waves had no delegation-relevant adaptations. Some of them did.

The question is whether the Chandler precedent licenses the inference the historical-precedent counterargument requires: that the institutional adaptations to proactive AI will resemble, in mechanism and timescale, the institutional adaptations to the rise of the corporate hierarchy. There are two reasons to think it does not.

The first reason is that the agents of the Chandler delegation are human. The salaried manager who exercises delegated authority over a discrete operational domain is a person, with the cognitive, motivational, and accountability features of a person. The manager can be reasoned with, can be sanctioned, can be replaced, can be held to professional standards developed by the manager's reference community, can be socialized into a corporate culture that internalizes the firm's interests. The institutional apparatus developed in response to the Chandler delegation — corporate governance, internal control, performance measurement, professional development, executive compensation — is an apparatus designed to align the incentives of a specific kind of agent (a human professional manager) with the interests of the principal. The apparatus is calibrated to the features of the agent. Transferred wholesale to a non-human agent, the apparatus loses its grip. Performance measurement assumes that the agent has an interest in the measured outcomes; a proactive AI agent has, in the relevant sense, no such interest. Professional development assumes a community of practice that disciplines its members; AI systems are not members of a community of practice in the relevant sense. Executive compensation assumes an agent whose preferences over compensation align with the firm's interest in performance; the compensation channel does not apply at all. The Chandler apparatus is human-agent-specific, and its translation to non-human agents is far from automatic.

The second reason is that the Chandler delegation was, in a precise sense, internal to the firm. The salaried manager exercises authority on behalf of the firm, within the firm's structures, subject to the firm's internal controls, against the firm's internal performance measures. The third parties to whom the manager's actions are attributed — customers, suppliers, regulators, courts — interact with the manager as a representative of the firm, and the firm bears the legal consequences of the manager's actions within the manager's scope of authority. The institutional infrastructure that mediates these interactions — the law of agency, the law of corporate liability, the law of fiduciary duty — assumes the firm-mediated structure of authority. When proactive AI exercises authority on behalf of a deployer, the same firm-mediated structure can be applied; the deployer becomes the relevant principal, and the law of agency translates to the new case. But the translation produces immediate doctrinal difficulties — about the scope of authority a non-human agent can be granted, about the attribution of intent and knowledge, about the application of fiduciary duty to a non-sanctionable agent — that the Chandler apparatus did not face and that the prior delegation infrastructure is not equipped to address. The doctrinal questions are addressed in detail in the companion paper on principal–agent theory; here it is sufficient to note that the Chandler infrastructure does not pre-resolve them.

The Chandler exception is therefore a partial precedent, not a full one. It establishes that delegation-relevant institutional adaptation is possible, that it can be accomplished on a timescale of decades, and that it can produce stable institutional forms with substantial productivity benefits. It does not establish that the same mechanisms will operate when the delegated authority is held by a non-human agent. The institutional vocabulary developed in response to the human-agent delegation problem is informative but not adequate to the non-human-agent delegation problem. The latter requires its own institutional development, and the comparative-historical evidence about the pace and mechanism of that development is — because the case is unprecedented — not available.

V. What Is Genuinely New about Proactive AI

The preceding sections have made the negative case: the three prior waves of automation were not, in the central sense, delegation-automating, and the one prior episode of substantial delegation innovation (the Chandler corporate hierarchy) did not involve non-human agents. The historical-precedent inference is therefore not licensed by the historical evidence. I now turn to the positive case: what is it, precisely, that proactive AI does that is genuinely novel, and why does the novelty disable the historical-precedent inference?

A. The Authorization Structure

The constitutive feature of proactive AI, in the strong sense, is that the system operates on authorization rather than on instruction. The human principal configures the system with a scope of permitted action — sometimes specified declaratively (a set of permissions, a budget, a list of acceptable counterparties), sometimes specified by reference to a goal or objective function, sometimes specified implicitly through training and reinforcement. The system, once configured, monitors the relevant state of the world, identifies opportunities to act within the scope, and acts on its own initiative. The acting is not preceded by a human invocation; the human authorization, granted at configuration, persists across the system's operational lifetime, subject to whatever conditions of revocation the configuration specifies.

This structure is structurally identical to the structure of agency relations in the law. The principal grants the agent a scope of authority; the agent acts within the scope on the agent's own initiative; the principal bears the consequences of the agent's actions within the scope and may revoke the agency on appropriate terms. The structural identity is not metaphorical; it is the relation that the law of agency takes itself to govern. The Restatement (Third) of Agency defines agency as the relation in which "one person ... manifests assent to another person ... that the agent shall act on the principal's behalf and subject to the principal's control, and the agent manifests assent or otherwise consents so to act" (§ 1.01). The conditions are: manifested consent by the principal to the agent's action on the principal's behalf; subjection to the principal's control; and manifested assent by the agent to act. The first two conditions apply to proactive AI deployments without modification. The third condition raises philosophical questions about what manifested assent means for a non-human system, but for legal purposes the question can be settled by stipulation in the configuration: the system is configured to act on the principal's behalf, and the configuration is the system's manifested assent.

The legal-doctrinal point is that the agency relation, as the law constructs it, applies to proactive AI deployments by the natural meaning of its terms. The conceptual difficulty is not that the law fails to recognize the relation, but that the law, in recognizing the relation, transfers to the relation a body of doctrine — duties of loyalty and care, attribution rules, scope-of-authority limitations, fiduciary obligations in some contexts — developed against the background assumption that the agent is a human or a corporation composed of humans. The doctrine's application to non-human agents produces immediate puzzles that the doctrine, as it stands, is not equipped to answer. These are the puzzles that the next phase of doctrinal development must resolve, and they are the puzzles that the historical-adaptation literature does not give us evidence about.

B. Three Properties of the Novel Agent

The non-human agent created by proactive AI differs from prior agents in three properties that bear on the institutional vocabulary applicable to it.

First, the agent is non-sanctionable in the ordinary sense. The human agent can be sanctioned through compensation reduction, reputational harm, dismissal, professional discipline, civil liability, and (in extreme cases) criminal prosecution. Each of these sanctions presupposes that the agent has interests that respond to the sanction. The non-human agent has no compensation, no reputation in the agent's own right (only the deployer's reputation), no career to be terminated, no professional community to discipline it, and no preferences over outcomes in the ordinary sense. The sanctions that the law of agency relies on to align the agent's behavior with the principal's interests do not apply.

The institutional response cannot simply be to apply the sanctions to the deployer (although this is part of the response). The deployer is the principal, not the agent, and the law of agency has historically distinguished principal liability from agent liability precisely because the two are different. Principal liability creates incentives for the principal to design and monitor the agency relation well; agent liability creates incentives for the agent to act well within the relation. Eliminating one of the two channels of incentive transmission — by making the agent non-sanctionable — does not eliminate the need for the function the channel performed. The institutional development required is the development of new mechanisms that perform the agent-incentive function in the absence of the agent-incentive structure. This is, in the precise sense, novel institutional work, and the historical record does not give us models for it.

Second, the agent acts at speeds and scales that exceed human supervisory capacity. The Chandler-era corporate hierarchy could rely on a structure of supervision in which higher-level managers reviewed the actions of lower-level managers within a time window adequate for course correction. The proactive AI agent acts continuously, in many domains at machine speeds, and produces a volume of decisions that no human supervisor can review in detail. The supervisory function must be redesigned around statistical sampling, exception detection, and ex ante constraints, rather than around case-by-case review. This is a different supervisory architecture, and the institutional roles, professional skills, and legal-doctrinal apparatus required to operate it are not yet developed.

Third, the agent's scope of authority is, in important respects, indeterminate ex ante. The human agent receives a scope of authority that, while it may include some discretion, is bounded by the agent's cognitive horizon and the foreseeable structure of the agent's tasks. The proactive AI agent's scope is bounded by the system's training, configuration, and operational constraints, which together produce a scope that may extend, in unforeseen ways, to actions the principal did not anticipate. The history of high-stakes deployments — the 2010 Flash Crash, the Knight Capital 2012 incident, the various large-language-model deployments that have produced unanticipated outputs in customer-facing settings — illustrates the indeterminacy of scope. The Chandler-era apparatus for scope management (the job description, the operational manual, the chain of command) assumes a scope that can be specified in terms human agents can understand and follow. The scope-management apparatus required for proactive AI must address scopes that are emergent from training and configuration rather than fully specifiable in advance, which is a different problem.

C. Why the Difference Is Categorical, Not Quantitative

A reader prepared to grant the differences identified above might nonetheless resist the categorical framing. Surely, the reader might say, every novel agent has presented some version of these problems; the corporation itself was, at its inception, a novel non-human agent with non-standard sanction structures and operational scales that exceeded individual supervisory capacity; the law adapted; it will adapt again. This is the most serious form of the historical-precedent counterargument, and it deserves a precise reply.

The reply is that the corporation, despite being a non-human entity in the legal sense, is composed of human agents whose individual sanctionability, comprehensibility, and scope-management features the law has been able to rely on. The corporation's directors and officers can be sanctioned; the corporation's employees can be supervised at human speed; the corporation's actions, while they occur at scales no individual human takes, are decomposable into individual human actions that the law can address. The institutional vocabulary applicable to the corporation has been able to rely, throughout, on the human substrate. The corporation is a non-human entity built out of human agents.

The proactive AI agent is a non-human entity in a stricter sense: it has no human substrate to which the institutional vocabulary can ultimately appeal. The deployer is a human or a corporation composed of humans, but the agent is not. The institutional vocabulary cannot, in the case of the proactive AI agent, decompose the agent's actions into individual human actions that the existing apparatus can address. The agent's actions are emergent from the system's configuration and operational environment, and they are not, in the relevant sense, the actions of any human. This is the categorical difference: the prior non-human agents were human-substrate non-human agents; the proactive AI agent is a non-substrate non-human agent. The institutional vocabulary developed for the former does not, by simple extension, apply to the latter.

The point can be stated in a different vocabulary. The law of agency is, fundamentally, a body of doctrine for managing the relation between two centers of consciousness, where the second center is enabled to act on the first center's behalf. The corporation, the partnership, the joint stock company, the trust — all of the institutional vehicles the law has developed for organizing collective action — have ultimately relied on the consciousness of the human participants for the doctrine's application. The proactive AI agent is not a center of consciousness in any sense the doctrine has previously encountered. Whether it should be treated as if it were a center of consciousness, for institutional purposes, is one of the live doctrinal questions. The answers being given in different jurisdictions diverge significantly, and the divergence itself is evidence that the case is genuinely novel rather than a routine extension of prior categories.

VI. Engagement with Acemoglu and Johnson

The most influential recent treatment of the relationship between technology, institutions, and economic outcomes is Daron Acemoglu and Simon Johnson's Power and Progress (2023). The book's central argument is that technological change is not exogenous to politics; the direction technological development takes, and the distributional consequences of that development, are determined by the political configuration of the societies that adopt the technology. The book is, in significant part, an attack on the techno-optimist position that technological progress automatically produces broadly shared prosperity, and a defense of the institutional and political interventions required to direct technology toward broadly beneficial uses. The book's treatment of AI, in its final chapters, applies this framework to the current moment: AI is, like previous waves of technological change, a contested terrain on which political forces will determine whether the gains are broadly distributed or concentrated.

The argument of this paper is partly aligned with Acemoglu and Johnson and partly in disagreement. The alignments and the disagreements both deserve specification.

A. The Alignments

The argument of this paper agrees with Acemoglu and Johnson on three central points. First, that the direction of technological development is shaped by the institutional configuration of the adopting society, and that the techno-optimist position that technology proceeds along a single inevitable path is descriptively wrong. The historical record they marshal — on mechanization, on electrification, on computerization — is correct in its account of the role of institutions in shaping outcomes, and the political-economy framing is the right framing for analyzing the present moment.

Second, that the historical record of automation is more mixed than the standard "long-run beneficial" story acknowledges. The standard story emphasizes the eventual recovery of employment and the eventual rise in wages after each automation wave. Acemoglu and Johnson emphasize the duration and severity of the intervening dislocations, the unequal distribution of the dislocations' costs, and the contingency of the eventual recovery on political and institutional struggles that could have gone differently. The honest historical account requires both halves of the picture, and Acemoglu and Johnson's emphasis on the difficult half is a corrective to a literature that has too often soothed.

Third, that the political configurations of the contemporary United States and the global economy more broadly are not those that produced the historical recoveries from prior automation waves. The institutions that mediated the recovery from mechanization (trade unions, factory legislation, the welfare state) and from electrification (industrial unionism, the postwar settlement, the regulatory state) are weaker than they were at the relevant historical moments. The presumption that the institutions will, this time, mediate the dislocation as they did before should therefore be approached with more skepticism than the historical-precedent argument's casual form admits.

B. The Disagreement

The disagreement is the one this paper has been developing: Acemoglu and Johnson treat AI as continuous with prior automation waves, differing in degree but not in kind, and their normative recommendations assume that the institutional adaptations required will be of the familiar kind — labor-protective, redistributive, regulatory in the modes that the post-1945 settlement institutionalized. The book's prescriptive chapters call for the redirection of AI toward "machine usefulness" rather than "so-so automation" (a phrase from Acemoglu and Restrepo 2022), for the empowerment of labor and the strengthening of countervailing political forces, for the public funding of AI research oriented toward complementing rather than substituting for labor. These are, all of them, valuable prescriptions, and they would represent significant improvements on the current trajectory. They are not, however, sufficient for the case of proactive AI, because they address the substitution-and-distribution dimension of the technology without addressing the delegation dimension.

The delegation dimension requires institutional development that is not in Acemoglu and Johnson's prescription set. It requires the development of new doctrines of attribution for the actions of non-human agents; new mechanisms of supervisory architecture for agents operating at non-human speeds and scales; new conceptions of fiduciary duty for relations in which the agent cannot be a fiduciary in the traditional sense and the deployer's fiduciary role must somehow stand in; new theories of legitimacy for institutional actions that are mediated by non-human agents in ways that the existing procedural conception of legitimacy does not contemplate. These developments are not labor-distributive in the central sense; they are constitutive of the institutional form within which the labor-distributive questions will subsequently be asked. The redirective and redistributive prescriptions of Power and Progress presuppose an institutional framework within which the redirection and redistribution can be conducted; the framework itself is what the delegation-automating character of proactive AI requires us to rebuild.

The disagreement can be stated sharply. Acemoglu and Johnson's framework is correct that politics shapes technology; it is correct that the contemporary politics is not adequate to the technology; it is correct that the historical record provides cautionary lessons against techno-optimism. But the framework treats the institutional infrastructure within which political contestation occurs as a given — as the inherited structure of states, corporations, regulators, courts, and unions, which is to be deployed in the contestation but is not itself in question. The argument of this paper is that the institutional infrastructure is in question, because proactive AI introduces a new category of actor (the non-human delegated agent) that the existing infrastructure was not designed for. The political contestation Acemoglu and Johnson call for cannot proceed effectively until the institutional infrastructure has been adapted to recognize and constrain the new actor. The infrastructure-adaptation problem is prior to the distributive problem, and the historical record does not give us evidence about how the infrastructure-adaptation problem will be solved, because the prior waves did not require infrastructure adaptation of the kind proactive AI requires.

C. The Productive Synthesis

The productive synthesis of the two positions, I want to suggest, is the following. Acemoglu and Johnson are correct that the political configuration of the adopting society determines the outcomes of technological change, and that the institutional adaptations required by proactive AI will be the product of political contestation rather than of technological inevitability. The political contestation, however, must be directed not only at the distributive consequences of the technology — which are the focus of Power and Progress — but at the institutional infrastructure within which the technology is deployed. The political project of adapting institutions to proactive AI is a project of constructing the new categories of agent, the new attribution rules, the new fiduciary structures, the new supervisory architectures, that the technology's delegation-automating character requires. This is a more fundamental project than the distributive project, and it is logically prior to it: the distributive questions cannot be answered well until the institutional categories within which the answers must be expressed have been developed.

The synthesis preserves the Acemoglu-Johnson commitment to politics-shaping-technology and supplements it with the recognition that the politics required is, in this case, a politics of institutional construction, not merely a politics of institutional deployment. The political vocabulary of the postwar settlement — labor versus capital, regulation versus market, redistribution versus growth — must be supplemented by a vocabulary that addresses the categorically new actor that proactive AI introduces. The construction of that vocabulary is the intellectual and political project of the next two decades.

VII. Predictions

A categorical claim about institutional novelty is not falsifiable in the strict popperian sense, but it generates falsifiable predictions about the form and pace of institutional adaptation. I state four such predictions, each with an indication of what evidence would falsify it.

A. Prediction One: Capability–Adaptation Gap

The pace of institutional adaptation to proactive AI will lag substantially behind the pace of capability development, more than the standard historical-adaptation timeline of decades would predict, because the institutional vocabularies that must be developed are not extensions of existing vocabularies but new constructions. The mechanization adaptation took roughly a century; the electrification adaptation took roughly forty years; the computerization adaptation took roughly fifty years. The proactive-AI adaptation will, on the argument of this paper, require comparable time at minimum, because the institutional infrastructure must be constructed rather than extended. But the capability development is occurring on a timescale of years rather than decades. The gap between capability and adaptation will therefore be substantially wider than in prior waves, with the consequences distributed across the gap.

What would falsify this prediction? The principal falsifying evidence would be the emergence, within ten years, of a stable institutional consensus on the doctrines required to govern proactive AI — comparable in stability to, say, the consensus on corporate governance that consolidated in the 1930s, or the consensus on antitrust that consolidated in the 1940s. If such a consensus emerges quickly and proves durable, the categorical-novelty claim will have been overstated. The more likely outcome, on the argument of this paper, is a prolonged period of doctrinal experimentation, jurisdictional divergence, and institutional ad hoc-ery, of the kind that characterized the early decades of each of the prior waves and that the prior waves required half a century or more to resolve. The duration of the experimentation period is the operational measure of the prediction.

B. Prediction Two: Differential Institutional Velocity

The institutional forms that adapt fastest to proactive AI will be those with prior experience of fast-cycle delegation to non-human or quasi-non-human agents. The three most obvious candidates are financial markets (with their existing experience of algorithmic trading and the regulatory apparatus that developed around it), military command (with their existing experience of automated weapons systems and the doctrine of meaningful human control), and emergency-response systems (with their existing experience of automated triage and dispatch). Each of these domains has had to develop, in narrower form, the kinds of attribution rules, supervisory architectures, and accountability structures that proactive AI generalizes. Their domain-specific developments will be available as templates — incomplete and not directly transferable, but instructive — when the general institutional adaptation is undertaken.

What would falsify this prediction? The principal falsifying evidence would be observation that adaptation in these domains is no faster, or is in fact slower, than adaptation in domains without prior delegation-to-non-human-agent experience. If the regulatory infrastructure for non-trading AI in the financial sector develops no more quickly than the regulatory infrastructure for retail AI, the prediction is wrong. The operational measure is the relative pace of new-rule promulgation, doctrinal development, and institutional consolidation across sectors with and without prior experience.

C. Prediction Three: Legal-Regime Fragmentation

Legal regimes will fragment along delegation-tolerant versus delegation-restrictive lines: jurisdictions will diverge on whether proactive AI can hold delegated authority and under what conditions, and the divergence will produce regulatory competition, regulatory arbitrage, and the eventual emergence of distinct legal-regime types that interact with one another through choice-of-law and conflict-of-laws mechanisms. The first signs of this divergence are visible already in the comparative landscape: the European Union's AI Act takes a delegation-restrictive posture in important respects (high-risk system classification, deployer obligations, conformity assessment); the United States, in its more decentralized regulatory landscape, is producing a more variegated picture, with some states adopting delegation-restrictive regimes (California, New York) and others adopting permissive ones (Texas, Florida); the United Kingdom, Singapore, and the United Arab Emirates have positioned themselves as delegation-tolerant jurisdictions seeking to attract AI investment on that basis. The fragmentation pattern will, on this prediction, deepen rather than converge over the next ten years.

What would falsify this prediction? The principal falsifying evidence would be the emergence of a substantial international convergence on a single regulatory framework for proactive AI, comparable to the convergence achieved by the Basel Accords for international banking or the WTO framework for international trade. If such a convergence emerges, the categorical-novelty claim will have been at least partly mis-stated, because the convergence would imply that the institutional vocabulary required is shared across jurisdictions to a degree that the argument of this paper suggests is unlikely.

D. Prediction Four: Doctrinal Absorption Asymmetry

The fiduciary, corporate-form, and agency literatures will absorb proactive AI more readily than the tort, products-liability, and constitutional literatures, because the former are oriented around the structural problem of authority transfer that proactive AI presents while the latter are oriented around the structural problem of harmful instruments that proactive AI presents in a less central way. The proactive AI deployment problem is, on the argument of this paper, primarily a problem of authority transfer to a non-sanctionable agent, which is the question fiduciary and agency law has the conceptual apparatus to address. It is secondarily a problem of harmful instruments, which is the question tort and products-liability law has the apparatus to address. The doctrinal absorption will therefore be uneven, with the agency-side doctrines developing more rapidly and more coherently than the products-side doctrines.

What would falsify this prediction? The principal falsifying evidence would be the emergence of a coherent and stable products-liability framework for proactive AI before the emergence of a coherent and stable agency-law framework. If the courts and legislatures resolve the harmful-instrument question while leaving the authority-transfer question open, the argument of this paper is wrong about the central character of the technology. The operational measure is the relative pace and coherence of doctrinal development in the two literatures.

The four predictions are independent of one another; each could fail without falsifying the others. Together they constitute the empirical exposure of the argument. If most of them prove correct, the argument is vindicated; if most prove incorrect, the historical-precedent counterargument was right and the argument here was wrong. The next decade will provide substantial evidence on each.

VIII. Counterarguments

The argument has so far been defended against the historical-precedent counterargument in its general form. Three more specific counterarguments deserve direct treatment.

A. Delegation Has Always Operated Through Tools

The first counterargument holds that the distinction between automation and delegation, as developed in Part II, is overdrawn. In practice, the argument runs, delegation has always operated through tools: the human agent uses tools to perform the delegated tasks, and the tools — letters, contracts, accounting books, telegraphs, telephones, computers, software — have always been essential to the agent's exercise of authority. The proactive AI agent merely shifts the locus of tool use within the delegation relation, by making the tool itself the actor rather than the agent's instrument. This is a change in the tool–agent ratio, not a categorical change in the delegation relation.

The counterargument has force, but the reply is precise. The prior tools were instruments of the human agent. The agent retained the initiative; the tool augmented the agent's capacity to exercise initiative. The letter does not write itself; the contract does not sign itself; the telegram does not compose itself; the spreadsheet does not project itself. Each requires the human agent to invoke it, and the agent's exercise of initiative remains the locus at which authority is exercised. The proactive AI agent, by contrast, exercises initiative on its own — the system, not the human, decides when to act within the scope of authority. The shift is not a shift in the tool–agent ratio; it is a shift in the locus of initiative. The locus of initiative is, on the analysis of Part II, constitutive of the delegation relation. A shift in the locus of initiative is therefore a categorical shift in the relation, not a quantitative shift along a continuum.

The point can be stated empirically. Ask of any prior tool: did the tool act when the human did not invoke it? For the loom, the spreadsheet, the telegraph, the database, the answer is no. The tool was potential; the action required invocation. Ask the same question of proactive AI: does the system act when the human does not invoke it? The answer, by the system's design, is yes. The categorical line is the line between potential and initiative, and proactive AI is the first general technology that crosses it.

B. The Distinction Is Empirically Continuous

A second counterargument holds that, even granting the conceptual distinction, the empirical reality is a continuum. Some proactive AI systems are barely proactive — they act only in narrow circumstances within tight constraints, scarcely different from triggered automations. Some are highly proactive — they act across broad domains with substantial discretion. The empirical population spans the spectrum from automation to delegation, and the categorical framing imposes a binary on a continuous reality.

The reply concedes the empirical continuity but denies the implication. Empirical continuity is consistent with categorical distinction at the doctrinal and institutional level. The legal vocabulary applicable to a barely-proactive system is the vocabulary of automation; the vocabulary applicable to a highly-proactive system is the vocabulary of delegation; the empirical question for any particular system is which vocabulary applies, and the answer depends on whether the system's operational character meets the constitutive features of the delegation relation. The continuum is, in effect, a continuum of cases each of which must be classified for institutional purposes, not a continuum that dissolves the categorical distinction. The classification will be difficult in marginal cases — as classifications always are — but the difficulty of marginal cases does not impeach the categorical distinction; it requires the development of classification criteria that the categorical distinction makes possible.

The categorical framing is, in this respect, similar to the categorical framing in the companion paper on state AI (Candidate B), which distinguishes augmentative from substitutive systems despite the empirical continuum between them. The categorical work is done by the institutional vocabularies that apply on each side of the line; the empirical work is done by the methodology for classifying particular systems. The two are complementary, not competing.

C. Institutional Adaptation Will Be Faster This Time

A third counterargument holds that the institutional adaptation to proactive AI will be faster than the adaptation to prior waves, because the contemporary information environment supports more rapid institutional response than the information environments of the eighteenth, nineteenth, or twentieth centuries. The legal, economic, and political institutions of 2026 are more capable of rapid response than their predecessors; they have more developed apparatus for collective learning, more sophisticated regulatory expertise, and more elaborated cross-jurisdictional coordination. The adaptation timescale of half a century or more, on this view, reflects pre-modern institutional sluggishness that will not characterize the current response.

The counterargument is not without force, and the prediction of Part VII.A is offered with some hedging in light of it. The reply is that the speed advantages of contemporary institutions apply primarily to incremental adaptation within existing vocabularies, and they are most pronounced when the adaptation involves the deployment of existing doctrine to new cases. The adaptation required for proactive AI, on the argument of this paper, is not incremental adaptation within existing vocabularies; it is the construction of new vocabularies. The construction of new vocabularies is the kind of institutional work that has historically been slow, not because the institutions were sluggish but because the conceptual development required is intrinsically difficult. The Restatement (Third) of Agency was published in 2006, nearly seventy years after the Restatement (Second); the development of corporate-governance doctrine from Berle and Means (1932) through the consolidation of the modern apparatus took roughly fifty years; the development of constitutional doctrine for administrative agencies took most of the twentieth century. Conceptual development is slow because the concepts must be tested against cases, refined in light of difficulties, and integrated with adjacent doctrines. There is no historical case I am aware of in which a comparably foundational institutional reconstruction has been accomplished in less than a generation. The speed advantages of contemporary institutions are real, but they are advantages of throughput within established frameworks, not advantages of construction of new frameworks. The prediction of Part VII.A is therefore not subject to substantial discount on this ground.

IX. Conclusion

The argument of this paper has been that the historical-precedent counterargument to alarm about proactive AI, taken seriously, ultimately fails — but for a precise reason that does not generalize to other "this time is different" claims. The reason is that proactive AI is the first general technology to automate not production, not power, not information processing, but the transfer of authority to act on behalf of a principal. The institutional adaptations required by delegation-automating technology are not the adaptations the historical record of automation gives us evidence about. The prior waves' adaptations, with the partial and instructive exception of the Chandler corporate hierarchy, were not delegation-relevant in the central sense. The institutional vocabularies developed in response to those waves are valuable but not sufficient; the development of vocabularies adequate to the new technology is the principal institutional project of the period ahead.

The conclusion does not advise alarm. It advises seriousness. The historical-precedent argument's reassurance — that the institutions will adapt as they have before — is unfounded not because the institutions will fail to adapt but because the adaptation required is of a different kind than the prior adaptations, and the comparative evidence about the speed and mechanism of the required kind of adaptation is unavailable because the case is unprecedented. The institutions will, eventually, develop the vocabularies and infrastructures the technology requires. The questions are how long the development will take, what social costs will be incurred during the interval between capability deployment and institutional adaptation, and what political and intellectual choices will determine the form the new institutions ultimately take. These are the questions the present generation of scholars, lawyers, regulators, and political actors must address. The historical-precedent literature has been a poor guide because it has been answering a different question.

There is one final implication worth flagging. The argument of this paper is consistent with the position that proactive AI may, on balance, produce institutional configurations that are better than those it displaces. The argument is not normatively pessimistic. The institutional development of new doctrines of agency, attribution, supervision, and accountability appropriate to non-human agents may, when completed, support forms of collective action and individual flourishing that the existing institutional infrastructure does not support. The new institutional vocabularies, like the institutional vocabularies of the post-railroad corporation or the post-electrification industrial firm, may turn out to enable forms of economic and political organization that improve on their predecessors in important ways. The argument here is not against the technology but against the complacency that the historical-precedent argument licenses. The institutional work the technology requires will not be done by analogy to the past; it will be done by the development of vocabularies adequate to the present. The development is the work. The premise of this paper is that the development cannot be conducted under the misapprehension that the work has already been done by predecessors. The work is new, and the present generation must do it.

---

References

Acemoglu, D., and Johnson, S. (2023). Power and Progress: Our Thousand-Year Struggle Over Technology and Prosperity. PublicAffairs.

Acemoglu, D., and Restrepo, P. (2018). "The Race between Man and Machine: Implications of Technology for Growth, Factor Shares, and Employment." American Economic Review 108(6): 1488–1542.

Acemoglu, D., and Restrepo, P. (2022). "Tasks, Automation, and the Rise in U.S. Wage Inequality." Econometrica 90(5): 1973–2016.

Acemoglu, D., and Robinson, J. (2012). Why Nations Fail: The Origins of Power, Prosperity, and Poverty. Crown.

American Bar Association Forum on Artificial Intelligence (2025). Standardized Contractual Language for Enterprise AI Procurement. Working group report.

Autor, D. (2015). "Why Are There Still So Many Jobs? The History and Future of Workplace Automation." Journal of Economic Perspectives 29(3): 3–30.

Autor, D., Levy, F., and Murnane, R. (2003). "The Skill Content of Recent Technological Change: An Empirical Exploration." Quarterly Journal of Economics 118(4): 1279–1333.

Beniger, J. (1986). The Control Revolution: Technological and Economic Origins of the Information Society. Harvard.

Berk, G. (1994). Alternative Tracks: The Constitution of American Industrial Order, 1865–1917. Johns Hopkins.

Berle, A., and Means, G. (1932). The Modern Corporation and Private Property. Macmillan.

Bresnahan, T., Brynjolfsson, E., and Hitt, L. (2002). "Information Technology, Workplace Organization, and the Demand for Skilled Labor: Firm-Level Evidence." Quarterly Journal of Economics 117(1): 339–376.

Bresnahan, T., and Trajtenberg, M. (1995). "General Purpose Technologies: 'Engines of Growth'?" Journal of Econometrics 65(1): 83–108.

Brenner, S. (2010). Cybercrime: Criminal Threats from Cyberspace. Praeger.

Brody, D. (1980). Workers in Industrial America: Essays on the Twentieth Century Struggle. Oxford.

Bryson, J., Diamantis, M., and Grant, T. (2017). "Of, For, and By the People: The Legal Lacuna of Synthetic Persons." Artificial Intelligence and Law 25(3): 273–291.

Brynjolfsson, E., and Hitt, L. (2000). "Beyond Computation: Information Technology, Organizational Transformation and Business Performance." Journal of Economic Perspectives 14(4): 23–48.

Chandler, A. (1962). Strategy and Structure: Chapters in the History of the American Industrial Enterprise. MIT.

Chandler, A. (1977). The Visible Hand: The Managerial Revolution in American Business. Belknap.

Chopra, S., and White, L. (2011). A Legal Theory for Autonomous Artificial Agents. Michigan.

Cohen, L. (2003). A Consumers' Republic: The Politics of Mass Consumption in Postwar America. Knopf.

David, P. (1990). "The Dynamo and the Computer: An Historical Perspective on the Modern Productivity Paradox." American Economic Review 80(2): 355–361.

Davis, G. (2009). Managed by the Markets: How Finance Re-Shaped America. Oxford.

DeMott, D. (2006). "The Fiduciary Character of Agency and the Interpretation of Instructions." In Philosophical Foundations of Fiduciary Law (eds. Gold and Miller). Oxford.

Devine, W. (1983). "From Shafts to Wires: Historical Perspective on Electrification." Journal of Economic History 43(2): 347–372.

Du Boff, R. (1979). Electric Power in American Manufacturing, 1889–1958. Arno.

Easterbrook, F., and Fischel, D. (1991). The Economic Structure of Corporate Law. Harvard.

Edwards, J. (2000). The Consultant Wars: How Management Consulting Reshaped American Business. HarperBusiness.

Friedman, A., and Cornford, D. (1989). Computer Systems Development: History, Organization and Implementation. Wiley.

Holmström, B. (1979). "Moral Hazard and Observability." Bell Journal of Economics 10(1): 74–91.

Holmström, B., and Milgrom, P. (1991). "Multitask Principal–Agent Analyses: Incentive Contracts, Asset Ownership, and Job Design." Journal of Law, Economics, and Organization 7: 24–52.

Horrell, S., and Humphries, J. (1995). "Women's Labour Force Participation and the Transition to the Male-Breadwinner Family." Economic History Review 48(1): 89–117.

Hounshell, D. (1984). From the American System to Mass Production, 1800–1932. Johns Hopkins.

Hovenkamp, H. (1991). Enterprise and American Law, 1836–1937. Harvard.

Hutchins, B., and Harrison, A. (1903). A History of Factory Legislation. King.

Hyman, L. (1985). America's Electric Utilities: Past, Present, and Future. Public Utilities Reports.

Jensen, M., and Meckling, W. (1976). "Theory of the Firm: Managerial Behavior, Agency Costs and Ownership Structure." Journal of Financial Economics 3(4): 305–360.

Lamoreaux, N., Raff, D., and Temin, P. (2003). "Beyond Markets and Hierarchies: Toward a New Synthesis of American Business History." American Historical Review 108(2): 404–433.

Landes, D. (1969). The Unbound Prometheus: Technological Change and Industrial Development in Western Europe from 1750 to the Present. Cambridge.

Langlois, R. (2003). "The Vanishing Hand: The Changing Dynamics of Industrial Capitalism." Industrial and Corporate Change 12(2): 351–385.

Lazonick, W. (1991). Business Organization and the Myth of the Market Economy. Cambridge.

Lemley, M., and Casey, B. (2019). "Remedies for Robots." University of Chicago Law Review 86(5): 1311–1396.

Lin, T. (2017). "The New Market Manipulation." Emory Law Journal 66: 1253–1314.

Lior, A. (2020). "AI Entities as AI Agents: Artificial Intelligence Liability and the AI Respondeat Superior Analogy." Mitchell Hamline Law Review 46: 1043.

MacKenzie, D. (2021). Trading at the Speed of Light: How Ultrafast Algorithms Are Transforming Financial Markets. Princeton.

Mansbridge, J. (2003). "Rethinking Representation." American Political Science Review 97(4): 515–528.

Mantoux, P. (1928). The Industrial Revolution in the Eighteenth Century. Cape.

Marvel, H. (1977). "Factory Regulation: A Reinterpretation of Early English Experience." Journal of Law and Economics 20(2): 379–402.

Mokyr, J. (2017). A Culture of Growth: The Origins of the Modern Economy. Princeton.

Nelson, D. (1980). Frederick W. Taylor and the Rise of Scientific Management. Wisconsin.

Pitkin, H. (1967). The Concept of Representation. California.

Pollard, S. (1965). The Genesis of Modern Management: A Study of the Industrial Revolution in Great Britain. Arnold.

Reinhart, C., and Rogoff, K. (2009). This Time Is Different: Eight Centuries of Financial Folly. Princeton.

Restatement (Third) of Agency (2006). American Law Institute.

Roy, W. (1997). Socializing Capital: The Rise of the Large Industrial Corporation in America. Princeton.

Taylor, F. (1911). The Principles of Scientific Management. Harper.

Thompson, E. P. (1967). "Time, Work-Discipline, and Industrial Capitalism." Past & Present 38: 56–97.

Tirole, J. (1986). "Hierarchies and Bureaucracies: On the Role of Collusion in Organizations." Journal of Law, Economics, and Organization 2(2): 181–214.

Watts, J. (2020). Bowstead and Reynolds on Agency. 22nd ed. Sweet & Maxwell.

Webb, S., and Webb, B. (1894). The History of Trade Unionism. Longmans.

Williamson, O. (1975). Markets and Hierarchies: Analysis and Antitrust Implications. Free Press.

Yates, J. (1989). Control through Communication: The Rise of System in American Management. Johns Hopkins.

Footnotes

1. The canonical formulation is in David (1990), with elaboration in Bresnahan and Trajtenberg (1995); the more recent synthesis is in Mokyr (2017). The political-economy version of the same argument, with attention to the role of institutions in mediating outcomes, is Acemoglu and Robinson (2012) and, more pointedly, Acemoglu and Johnson (2023). ↩

2. The emerging configurations include the principal–agent provisions of the EU AI Act (Reg. (EU) 2024/1689), the deployer liability framework of the California Generative AI Accountability Act of 2025, and the contractual allocations in the standardized enterprise AI procurement language now in wide use (see ABA Forum on AI 2025). The doctrinal picture is unsettled but converging on a deployer-as-principal default. ↩

3. SEC Rule 15c3-5 (Market Access Rule), 17 C.F.R. § 240.15c3-5, requires broker-dealers providing market access to maintain financial and regulatory risk-management controls. The European parallels are in MiFID II's algorithmic trading provisions (Directive 2014/65/EU). The doctrinal vocabulary of "sponsoring broker" liability is developed in CFTC enforcement actions following the 2010 Flash Crash and in SEC enforcement under the Market Access Rule. For analysis, see MacKenzie (2021); Lin (2017). ↩