The Bifurcation Hypothesis

I. Introduction

The standard account of artificial intelligence diffusion in the firm proceeds in three stages. In the first, a general purpose technology arrives. In the second, complementary investments in human capital, organizational structure, and process redesign produce productivity gains that exceed the direct contribution of the technology itself. In the third, the technology becomes ubiquitous, prices fall, and the productivity gains are competed away as the technology is incorporated into the cost structure of every firm. The account is associated, most prominently, with Bresnahan and Trajtenberg's treatment of general purpose technologies and with Brynjolfsson and McAfee's extension of the framework to AI specifically.¹ It is implicit in the McKinsey Global Institute's annual forecasts, in the World Economic Forum's Future of Jobs reports, in the BCG and Accenture consultancy publications that shape executive thinking, and in the academic literature that follows the consultancy framing more closely than it acknowledges.

The account has the virtues of being theoretically grounded and roughly consistent with the historical record of prior general purpose technologies. It also has a defect. It treats the firm as a single decision-theoretic unit whose adoption decision is governed by an aggregate calculation of expected productivity gain net of adjustment cost. This treatment was already a simplification when applied to electrification or the computer — both of which diffused unevenly across functional areas within firms, with the operational layers leading and the adjudicative layers lagging by decades. It is a more consequential simplification when applied to proactive AI, because the source of the within-firm asymmetry is more pronounced and the mechanism producing the asymmetry is more identifiable.

The argument of this paper is that proactive AI — by which I mean algorithmic systems that initiate action, monitor state, and operate on schedules or triggers without per-instance human authorization, as distinct from systems that respond to discrete prompts — will diffuse through the firm with a structurally predictable asymmetry. Operational and transactional layers, in which errors are small in magnitude, recoverable, contained, and rarely the subject of formal adjudication, will absorb proactive AI rapidly. Strategic and adjudicative layers, in which errors are large in magnitude, often irreversible, externalized to identifiable third parties, and routinely the subject of tort, fiduciary, or regulatory adjudication, will absorb proactive AI slowly, haltingly, and with frequent reversals. The mechanism is not technical capability — by hypothesis, the same model class is available for both deployments — but the interaction of Knightian uncertainty about agent behavior with a liability regime that loads tail risk on the deployer rather than on the agent.

The paper makes four contributions. First, it sharpens the bifurcation hypothesis into a set of testable predictions with stated directional signs. Second, it develops the analytical mechanism — Knightian uncertainty plus liability exposure plus real-options waiting value — with the rigor needed for the hypothesis to function as more than a heuristic. Third, it specifies three empirical strategies in enough detail that another researcher could execute them. Fourth, it generates sector-specific predictions that are accountable to observation over a five-year window. The paper does not claim that bifurcation will obtain; it claims that bifurcation is the right framework against which the next decade of enterprise AI adoption should be measured, and it identifies the rival framework — competitive-pressure-driven uniform adoption — that constitutes the genuine empirical alternative.

The bifurcation hypothesis is, in one sense, a friendly amendment to the general purpose technology framework. It accepts that proactive AI is a general purpose technology and accepts the framework's three-stage account. It objects only to the implicit assumption that the second-stage complementary investments are equally available across functional areas. The argument here is that for adjudicative layers — those in which the firm's outputs determine consequential outcomes for identifiable third parties — the complementary investments include legal and institutional infrastructure that is genuinely scarce, slowly produced, and exposed to political contestation in ways that operational infrastructure is not. The bifurcation hypothesis is thus a claim about which complementary investments are binding and where the binding constraints will produce the most observable asymmetries.

In another sense, the hypothesis is a less friendly amendment. It implies that the headline adoption figures circulated by consultancies and policy analysts — figures that aggregate across functional areas and treat all deployments as substitutable units — systematically obscure the variation that matters. A firm reporting that "75% of business units have deployed AI" may be reporting a deployment portfolio in which proactive AI handles inventory routing and contract clause retrieval while leaving capital allocation, hiring adjudication, and credit underwriting in their pre-AI institutional form. The aggregate figure conceals the bifurcation. The same firm five years on may report 90% — and the additional 15% may continue to come from operational layers, leaving the adjudicative layers exactly where they were. The bifurcation hypothesis predicts not the trajectory of the aggregate figure but the composition beneath it.

The paper proceeds as follows. Part II states the bifurcation hypothesis formally and decomposes it into four testable predictions. Part III develops the analytical mechanism: Knightian uncertainty about agent behavior, liability exposure that loads on the deployer, and the real-options analysis that follows. Part IV situates the argument within the technology adoption literature — Rogers' diffusion theory, David's productivity-paradox and path-dependence work, Bresnahan and Trajtenberg's general purpose technology framework, Acemoglu and Restrepo's task-based model of automation — and identifies the relationship between bifurcation and each prior frame. Part V specifies three empirical strategies in detail. Part VI generates sector-specific predictions across six sectors. Part VII develops the competitive-pressure alternative as a serious rival hypothesis and specifies the empirical pattern that distinguishes the two. Part VIII addresses three additional counterarguments. Part IX concludes with research agenda implications.

A note on scope. The paper concerns proactive AI specifically — systems whose deployment involves the transfer of authority to act, not merely the transfer of cognitive labor on a per-prompt basis. The distinction matters because the liability mechanism that drives bifurcation operates on systems that act, not on systems that suggest. A document-summarization tool used by a credit officer does not implicate the deployer's tort exposure in the way that an automated underwriting agent does, even if the underlying model is identical, because the locus of decision in the first case remains the credit officer and in the second case has been transferred. The argument here is about systems that have made the transfer, and it does not strictly apply to systems that have not. The reader interested in the locus-of-decision question more generally should consult the companion paper on the augmentative/substitutive typology, which develops the locus distinction in greater detail.²

II. The Bifurcation Hypothesis

A. Statement

Let a firm be a productive unit organized as a hierarchy of functional areas, each of which generates decisions of some characteristic type. Let an operational decision be a decision whose erroneous resolution produces a cost that is bounded, locally absorbed, and not the subject of routine third-party adjudication: which truck to route to which warehouse, which document to file in which folder, which meeting invitee to remind. Let an adjudicative decision be a decision whose erroneous resolution produces a cost that is unbounded in the tail, externalized to an identifiable third party, and the subject of routine third-party adjudication through tort, fiduciary, regulatory, or contractual channels: which loan applicant to approve, which job candidate to advance, which patient to admit, which insurance claim to deny.

Let proactive AI be an algorithmic system that initiates action, monitors state, or operates on schedules or triggers within a designated scope without per-instance human authorization. Let the adoption rate of proactive AI in a functional area be the proportion of decisions of that area's characteristic type that are taken by the algorithmic system rather than by a human decider, holding the algorithmic system's technical capability constant.

The Bifurcation Hypothesis. Within a firm, the adoption rate of proactive AI in operational functional areas will rise faster, reach a higher equilibrium, and be less subject to reversal than the adoption rate in adjudicative functional areas, holding technical capability constant. Across firms within a sector, the cross-sectional variation in adoption rate will be concentrated in adjudicative functional areas; operational functional areas will exhibit convergence. Across sectors, the average adoption rate of proactive AI in adjudicative functional areas will be inversely related to a measure of sector-specific liability exposure.

The hypothesis is not that adjudicative deployments will never occur. They will, and they already have — in fragmented forms, with frequent reversals and high observed variance. The hypothesis is that the adoption dynamic differs from the operational case in three measurable respects: rate, equilibrium level, and stability.

B. Predictions

The hypothesis decomposes into four predictions with stated directional signs.

Prediction 1 (Within-Firm Asymmetry). Within a firm, the time from technical feasibility to majority deployment of proactive AI will be at least 24 months longer in adjudicative functional areas than in operational functional areas, controlling for technical complexity. Predicted direction: T(adjudicative) − T(operational) > 24 months, where T is the lag from feasibility to majority deployment, with technical complexity measured by an external benchmark capability instrument.

Prediction 2 (Cross-Sector Liability Gradient). Across sectors, the proportion of decisions delegated to proactive AI in nominally adjudicative functional areas (credit underwriting, hiring, clinical triage, claims adjudication, sentencing recommendation) will be negatively correlated with sector-specific tort exposure, where tort exposure is measured by sectoral malpractice or product-liability insurance premium per dollar of revenue. Predicted direction: correlation negative, |ρ| > 0.3, statistically distinguishable from zero with adequate sectoral sample size.

Prediction 3 (Event-Study Reversal). Following a major adverse judgment or regulatory enforcement action against a proactive AI deployment in an adjudicative functional area, the rate of new proactive AI deployments in the same functional area across peer firms in the same sector will decline by at least 30 percent within 18 months relative to the pre-shock trend, and at least 10 percent of pre-shock deployments will be paused, scaled back, or reverted to human-locus operation. Predicted direction: deployment rate decline > 30%, reversal rate > 10%.

Prediction 4 (Composition-Beneath-Aggregate). Aggregate firm-level adoption figures will continue to rise smoothly while the within-firm composition of adoption shifts strongly toward operational layers. Specifically, in a panel of large firms over a five-year window, the share of total proactive-AI deployments accounted for by adjudicative functional areas will not exceed 25 percent of the aggregate, even where capability would in principle support a higher share. Predicted direction: share(adjudicative)/share(total) < 0.25 over the observation window.

The four predictions are jointly the testable content of the hypothesis. Their conjunction is what bifurcation predicts; individual confirmations or disconfirmations would be partial evidence rather than determinative. Prediction 1 is a within-firm time-series claim. Prediction 2 is a cross-sectional cross-sector claim. Prediction 3 is an event-study claim. Prediction 4 is a composition claim that holds aggregate figures fixed and asks where, within the aggregate, the adoption is occurring.

It is worth noting what the predictions do not require. They do not require that capability be uniform across functional areas (controls are specified for technical complexity). They do not require that liability exposure be the only driver of cross-sector variation (correlation, not causal identification, is the standard for Prediction 2 as initially stated, though the empirical strategy of Part V will address causal identification). They do not require that bifurcation be permanent — only that the asymmetry persist over the observation window and that the mechanism producing it (liability exposure, real-options waiting value) continue to operate.

C. What Would Refute the Hypothesis

A hypothesis worth defending is a hypothesis whose refutation conditions can be stated. The bifurcation hypothesis would be refuted by the following observed patterns.

First, if within-firm adoption rates of proactive AI in adjudicative functional areas tracked operational adoption rates with lag of less than 12 months, controlling for technical complexity, Prediction 1 would fail. Second, if cross-sector adoption rates of proactive AI in adjudicative areas were uncorrelated or positively correlated with sector-specific tort exposure, Prediction 2 would fail. Third, if event-study analysis around major adverse judgments showed no significant deployment-rate decline in adjacent sectors, Prediction 3 would fail. Fourth, if the within-aggregate composition shifted as quickly toward adjudicative layers as toward operational layers — if, that is, the 25 percent threshold were exceeded within the observation window — Prediction 4 would fail.

The empirical strategies of Part V are designed to evaluate each prediction. The honest researcher should specify in advance that disconfirmation on any one prediction weakens the hypothesis and that disconfirmation on three of the four would falsify it. The latter has not, to my knowledge, occurred in any sector to date; the former has occurred in narrow sub-sectors, most notably high-frequency trading, where the conditions of the hypothesis are arguably not met. Part VII engages this case directly under the competitive-pressure alternative.

III. Mechanism: Knightian Uncertainty and Real Options

A. Why the Standard Adoption Calculus Misfires

The standard adoption calculus is straightforward. A firm considering the deployment of a productivity-improving technology computes the expected net present value of the deployment as the discounted sum of expected per-period gains, less the upfront investment cost. The deployment proceeds when the expected NPV exceeds the cost. The calculus admits of complication — option values, learning effects, complementary investment costs — but the basic logic is the logic of discounted expected utility maximization under quantifiable risk.

This calculus is approximately correct for operational deployments of proactive AI. The deployer estimates the expected gain from automated inventory routing, estimates the expected cost of errors (small in magnitude, locally absorbed, recoverable), nets the two, and decides. The decision is not riskless — the system may fail, the integration may underperform, the staff may resist — but the risks are insurable, the errors are bounded, and the deployment is reversible. The expected-value calculation does the work.

The calculus misfires for adjudicative deployments. The reason is not that adjudicative deployments are riskier in the standard sense — that they have higher expected error rates or higher per-error costs. It is that they have a different structure of risk, in which the relevant quantity is not the expected value of the error distribution but the unbounded character of the tail and the inability of the deployer to characterize the tail in advance. This is the Knightian distinction between risk (quantifiable, insurable, tractable with expected-value calculation) and uncertainty (unquantifiable, uninsurable, tractable only with structural decision rules that do not depend on a probability distribution).³

For proactive AI in adjudicative contexts, the relevant tail uncertainty has three sources. First, the model's behavior on inputs far from its training distribution is poorly characterized; the deployer cannot, in general, produce a defensible upper bound on error rates for the long tail of unusual cases. Second, the legal regime governing liability for AI-mediated adjudicative decisions is itself uncertain — it is being constructed through litigation, regulatory enforcement, and statutory development in real time, and the deployer cannot predict the doctrine that will be applied to their deployment five years on. Third, even given a stable legal regime, the political and reputational consequences of a high-salience adverse outcome are unpredictable in their magnitude and durability — they depend on press cycles, on the identity of the affected individual, on the political climate at the moment of disclosure. The three sources compound: tail behavior is poorly characterized in a regime that is itself uncertain, with consequence functions that are themselves Knightian.

B. Liability Exposure and the Deployer-as-Residual-Bearer

The Knightian uncertainty would matter less if the agent could bear the residual risk. In the standard principal-agent framework, the residual risk of an agent's action is borne by the agent in proportion to the agent's stake — the agent's reputation, future compensation, dischargeability, or capital. A doctor whose erroneous diagnosis injures a patient bears the malpractice liability and the reputational consequence. An underwriter whose erroneous loan approval results in a default bears the bonus reduction and the career consequence. The principal's exposure is the residual after the agent's share.

A proactive AI agent cannot bear residual risk in the way a human agent can. It has no reputation that can be impaired. It has no future compensation that can be docked. It cannot be sued in its own name in most jurisdictions; even where the question of AI legal personality has been litigated, the answer has been no.⁴ The vendor that supplied the model can be sued for product defect or warranty breach, but the deployer's tort exposure to the injured party is not extinguished by the existence of the vendor cause of action — it is at most subject to indemnification, which is itself often capped and frequently disputed in litigation.

The practical effect is that the deployer of a proactive AI agent in an adjudicative context absorbs the residual risk that would, in a human-agent equivalent, have been shared with the agent. The deployer's expected liability per deployment year is therefore higher, all else equal, than the expected liability of an equivalent human-agent deployment, not because the agent is worse but because the residual-bearer structure is worse from the deployer's perspective. This produces the first wedge between operational and adjudicative deployments: even at equal error rates, the deployer bears more of the resulting cost in the adjudicative case.

The second wedge follows from the structure of the consequence function. In operational deployments, the per-error cost is bounded and approximately linear in the error rate — a 10% error rate produces approximately 10 times the cost of a 1% error rate. In adjudicative deployments, the per-error cost is unbounded in the tail because a single high-salience adverse outcome can produce a settlement, a judgment, a regulatory enforcement action, or a reputational episode whose cost is orders of magnitude greater than the average per-error cost. The deployer's expected cost is therefore not adequately characterized by the average error rate; it is dominated by the tail of the consequence distribution, and the tail is precisely what is Knightianly uncertain.

C. Real-Options Analysis

The combination — Knightian uncertainty in the tail, plus deployer-borne residual risk loading on the tail — converts the adjudicative deployment from a standard NPV problem into a real-options problem.

In the canonical real-options framework, an investment opportunity that is irreversible, made under uncertainty, with the option of waiting, has option value above the standard NPV threshold. The deployer should invest only when the project's expected value exceeds the upfront cost plus the option value of waiting — that is, the value of delaying the investment until additional information arrives that resolves some of the uncertainty. The relevant insight, developed at length by Dixit and Pindyck and earlier by McDonald and Siegel, is that under uncertainty the standard NPV rule systematically over-invests because it does not price the foregone option of waiting.⁵

For adjudicative proactive AI deployments, the option value of waiting is high for three reasons. First, the legal regime is being constructed in real time, and waiting allows the deployer to observe how doctrine develops without bearing the costs of being the test case. Second, model capability is improving rapidly, and waiting allows the deployer to deploy a system whose tail behavior is better characterized — even if the average performance gain from waiting is small, the tail-risk reduction from waiting may be large. Third, the first major adverse judgment in a given sectoral application will concentrate liability on the deployer in question and will, by reducing legal uncertainty for subsequent deployers, lower the option value of waiting for them — a strong first-mover penalty that further reinforces the waiting incentive.

The real-options framework predicts a specific pattern: deployers will wait until the option value of waiting falls below the productivity gain from immediate deployment. In operational contexts, the option value is low (tail risk is small, legal regime is stable, capability improvements would offer marginal tail-risk reduction over an already low base) and the productivity gain dominates; deployment proceeds. In adjudicative contexts, the option value is high (tail risk is poorly characterized, legal regime is unstable, capability improvements offer substantial tail-risk reduction) and dominates the productivity gain; deployment is deferred.

The pattern is reinforced by the irreversibility of certain adjudicative deployment costs. A deployer that issues credit decisions algorithmically and then reverts to human underwriting following an adverse event has likely shed the human underwriting capability, disrupted the training pipeline that produces new underwriters, and exposed itself to litigation premised on the prior algorithmic regime. The reversion is not free. The irreversibility increases the real-options waiting value relative to a fully reversible deployment.

D. The Strategic Value of Ambiguity

An additional element merits discussion. In adjudicative contexts, the deployer often has reason to maintain ambiguity about whether a deployment is in fact "proactive" in the sense relevant to this paper — that is, whether the system is operating with delegated authority or merely advising a human decider. The ambiguity is strategically valuable because it allows the deployer to claim the productivity gain of automated decision-making while preserving the legal cover of nominal human-locus authority. The empirical literature on automation bias suggests that systems described as advisory can in fact operate substitutively, with override rates falling below 5% in some deployments and the human decider acting as ratifier rather than decider.⁶

The strategic-ambiguity move is itself revealing of the bifurcation mechanism. It shows that deployers perceive the liability cost of explicit substitution to be high enough to be worth investing in linguistic and procedural cover. It also shows that the bifurcation hypothesis must be careful about its operational definition of adoption: a deployment that is functionally substitutive but nominally augmentative should count as adoption for purposes of the hypothesis, since the relevant locus has been transferred, but it is often missed by survey instruments that ask about "AI decision-making." The empirical strategies of Part V address this measurement problem.

IV. Position in the Technology Adoption Literature

A. Rogers' Diffusion Framework

Rogers' Diffusion of Innovations, in its successive editions, has been the workhorse account of how new technologies spread through populations of potential adopters. The framework specifies five characteristics of an innovation that predict its adoption rate (relative advantage, compatibility, complexity, trialability, observability), five categories of adopter (innovators, early adopters, early majority, late majority, laggards), and the S-shaped adoption curve that results from the interaction of innovation characteristics with adopter heterogeneity.⁷

The bifurcation hypothesis stands in an instructive relationship to Rogers' framework. Rogers' framework predicts variation in adoption rate as a function of innovation characteristics and adopter heterogeneity; it does not strongly predict that the same innovation, deployed in the same firm, will diffuse asymmetrically across functional areas within the firm. Indeed, Rogers' unit of analysis is the adopter (typically the firm or individual), not the functional area within the adopter. The bifurcation hypothesis can be read as an extension of Rogers to a different unit of analysis — the functional area — and as a specification of the mechanism (liability exposure, real-options waiting value) that produces the within-adopter variation.

The relevant Rogers characteristic for the bifurcation argument is trialability — the extent to which the innovation can be deployed experimentally with low cost of reversal. Operational deployments are highly trialable; adjudicative deployments are not. The bifurcation hypothesis is, in Rogers' vocabulary, a claim that trialability is the binding constraint in the adjudicative case and that the mechanism producing the trialability difference is the liability regime. This framing connects the hypothesis to a long-established adoption-theoretic apparatus while specifying the substantive mechanism Rogers' general framework leaves unspecified.

B. David and the Productivity Paradox

Paul David's work on the productivity paradox — the puzzle that the computer revolution was visible everywhere except in the productivity statistics — and his related work on path dependence in technology adoption are foundational to the present argument.⁸ David's central insight, developed through the comparative-historical case of electrification, was that the productivity gains from a general purpose technology depend on complementary investments in infrastructure, organization, and human capital that take decades to accumulate. The implication is that adoption is not a single event but an extended process in which the technology's full productive potential is unlocked only as the complementary investments mature.

The bifurcation hypothesis is consistent with David's framework in its structural form — adoption depends on complementary investments — but specifies a particular kind of complementary investment that has not been central to the prior literature. The investment is legal-institutional: the development of a body of doctrine, case law, regulatory guidance, and insurance product that makes the adjudicative deployment legally tractable. This investment is, in important respects, a public good — it is produced through litigation, regulatory rulemaking, and judicial decision-making rather than through firm-level investment — and it cannot be accelerated by firm-level expenditure beyond a relatively low ceiling. Firms can lobby for clarifying regulation, can settle test cases on terms that produce favorable precedent, can fund insurance products that pool risk, but they cannot, by writing larger checks, accelerate the production of doctrine itself.

The implication is that the rate-limiting step for adjudicative adoption is the production of legal-institutional infrastructure that is genuinely external to the firm. David's framework would predict slow adoption in any context where complementary investments are scarce; the bifurcation hypothesis specifies that for adjudicative AI deployments, the scarcest complementary investment is precisely the legal-institutional one, and that this is structurally distinctive of the adjudicative case relative to the operational case where the relevant complementary investments are internal to the firm and can be accelerated through expenditure.

C. Bresnahan and Trajtenberg on General Purpose Technologies

Bresnahan and Trajtenberg's general purpose technology framework specifies three characteristics — pervasiveness, improvement, and innovation spawning — and predicts that GPTs produce extended cycles of complementary innovation that distribute the technology's productive gains across the economy. The framework has been applied to electrification, the computer, the internet, and now AI.⁹

The bifurcation hypothesis accepts the GPT framework but specifies a within-firm asymmetry that the framework does not predict. The asymmetry is consistent with the framework — the GPT framework allows for varied adoption rates across applications — but is not predicted by it. The mechanism the present paper proposes (liability exposure) is, in GPT-framework vocabulary, a complementary-innovation requirement: adjudicative deployments require legal-doctrinal complementary innovation that operational deployments do not. The bifurcation hypothesis can therefore be read as identifying which complementary innovations are binding and where they are binding most.

The honest reading is that the GPT framework is too general to predict the bifurcation, but is consistent with it; the bifurcation hypothesis adds specificity that the GPT framework leaves out. The argument is not that the GPT framework is wrong but that its application to AI requires the additional structural element of liability-asymmetry-driven within-firm variation.

D. Acemoglu and Restrepo on Automation

Acemoglu and Restrepo's task-based model of automation is the most influential recent treatment of the relationship between automation, labor, and firm structure.¹⁰ The framework decomposes production into tasks, characterizes each task by whether it is automatable, and predicts the distributional consequences of automation as a function of the task composition of the economy. The framework has been used to explain wage polarization, the decline of middle-skill employment, and the displacement effects of robotization in manufacturing.

The bifurcation hypothesis is, in one important respect, an extension of the Acemoglu-Restrepo framework to a task characteristic that the original framework underweights. The original framework characterizes tasks primarily by their technical automatability — whether the task can be performed by a machine at acceptable accuracy. The bifurcation hypothesis adds a second characteristic: the liability profile of the task. Two tasks may be equally technically automatable but differ in the liability exposure produced by an erroneous performance; the bifurcation hypothesis predicts that the second characteristic is decisive for adjudicative tasks and produces the within-firm asymmetry the present paper documents.

The extension is friendly to the Acemoglu-Restrepo framework. It adds an element rather than rejecting one. It also produces empirical predictions that the original framework does not produce — in particular, that the rate of task automation will differ across functional areas within a firm in ways that the technical automatability alone does not predict, and that the differential will be concentrated where the liability profile is most adverse. This is a testable claim that builds on the Acemoglu-Restrepo apparatus rather than displacing it.

E. What the Bifurcation Hypothesis Adds

Synthesizing the relationship to the prior literature: the bifurcation hypothesis accepts the broad outlines of the GPT framework and the task-based model of automation but specifies a within-firm asymmetry that neither framework strongly predicts. The mechanism — liability exposure interacting with Knightian uncertainty to produce a real-options waiting dynamic — is consistent with Rogers' trialability concept and with David's emphasis on complementary investments, but specifies a kind of complementary investment (legal-institutional) that is genuinely external to the firm and slowly produced through public-good mechanisms. The hypothesis is best understood as an extension of the existing apparatus rather than a replacement, and as a specification of mechanism in cases where the existing apparatus is silent or general.

V. Empirical Strategy

The bifurcation hypothesis is testable. Three empirical strategies, developed in detail sufficient for execution, are described below. Each addresses different predictions and has different identification problems; together they constitute the empirical research agenda the hypothesis implies.

A. Strategy 1: Cross-Sector Adoption with Tort-Exposure Variation

The first strategy addresses Prediction 2 — the cross-sector inverse correlation between proactive AI adoption in adjudicative functional areas and sector-specific tort exposure.

The construction is as follows. The unit of analysis is the sector-functional-area cell, where sectors are defined at the NAICS three-digit level (giving roughly 100 sectors) and functional areas are classified by whether they are operational or adjudicative according to the definitions of Part II. The dependent variable is the proactive AI adoption rate in each cell, measured as the proportion of decisions of the relevant type that are taken by an algorithmic system, with measurement taken through a combination of firm surveys, vendor revenue data, and where possible direct observation of deployed systems.

The independent variable is sector-specific tort exposure, measured by professional liability or product liability insurance premiums per dollar of revenue for the sector, supplemented where available by data on settlement amounts, judgment amounts, and litigation frequency. The data sources are well-developed for several sectors: medical malpractice premiums are tracked by the National Association of Insurance Commissioners and by state insurance regulators; product liability premiums are tracked similarly; financial services litigation costs are tracked by the Securities Industry and Financial Markets Association and by industry-funded research organizations. For sectors with less developed insurance markets, the construction of the measure requires more work — synthetic measures based on litigation databases (Westlaw, Bloomberg Law) are tractable but require substantial cleaning.

The controls include sector-specific measures of technical complexity (drawing on benchmark capability instruments such as the GLUE and MMLU benchmarks for the relevant deployment), sector-specific measures of regulatory intensity (regulatory restrictions per sector, drawing on the RegData database), firm size distribution, and competitive structure (HHI). The identification is correlational and the principal threat is omitted variable bias from a third factor (sector culture, technical infrastructure inheritance) that drives both low adoption and high liability exposure independently.

The strategy can be strengthened with an instrumental variable approach for sectors where the tort exposure has plausibly exogenous determinants — for instance, the jurisdictional variation in medical malpractice damages caps following the wave of state-level tort reform between 1985 and 2005, which provides an instrument for sectoral liability exposure that is plausibly orthogonal to AI adoption-relevant unobservables.¹¹ The IV strategy requires careful work to establish exclusion (the damages cap should affect AI adoption only through its effect on liability exposure) but is tractable in the medical context where the timing and content of state-level reforms are well-documented.

The strategy's power depends on the sectoral sample size and the precision of the tort-exposure measure. A pilot study with 15-20 sectors and the readily available insurance data should be sufficient to detect a correlation of magnitude 0.3 or greater at conventional significance. Larger samples would allow detection of smaller effects and would permit the disaggregation by functional area that the hypothesis specifies.

B. Strategy 2: Within-Firm Functional-Area Variation

The second strategy addresses Prediction 1 — the within-firm asymmetry between operational and adjudicative adoption rates.

The construction requires firm-level data on deployment composition across functional areas. Such data are available in several forms. Stanford's Artificial Intelligence Index report has, since 2024, included a within-firm deployment composition survey of approximately 1,500 large firms across G20 economies, with deployments classified by functional area and by deployment type (advisory, automated, fully autonomous). The MIT Sloan/BCG annual Artificial Intelligence and Business Strategy survey, despite its limitations, includes functional-area decomposition for a smaller but more carefully tracked panel. Vendor-side data — particularly from the major enterprise AI vendors (Salesforce, Microsoft, Google Cloud, AWS) — can be aggregated to give deployment-count data with functional-area resolution, though access requires negotiation.

The unit of analysis is the firm-functional-area cell. The dependent variable is the time from technical feasibility (proxied by the availability of a vendor product for the functional area at a stated capability benchmark) to majority deployment within the firm (proxied by a 50% threshold of decisions of the relevant type handled by the system). The independent variable is a binary or ordinal measure of the functional area's adjudicative character, with operational areas (logistics, IT, document routing, internal scheduling) coded zero and adjudicative areas (credit, hiring, clinical decision, claims adjudication) coded one. Controls include firm size, industry, technical capability available to the firm (proxied by AI engineering staff per employee), and firm-level regulatory exposure.

The principal identification problem is that adjudicative functional areas may differ from operational areas on unobservable dimensions other than liability exposure — they may, for instance, have longer-tenured staff with more institutional power, more developed internal procedural infrastructure, or different vendor markets. The within-firm fixed-effects design controls for firm-level heterogeneity but does not address functional-area-level heterogeneity other than the adjudicative-character measure.

A stronger version of the strategy exploits within-functional-area variation in liability exposure across firms. Consider, for instance, hiring as a functional area. Firms operating in jurisdictions with strong disparate-impact enforcement (notably California, New York, and the District of Columbia post-2024 statutory reforms) face higher liability exposure for algorithmic hiring decisions than firms operating in jurisdictions without comparable enforcement. The bifurcation hypothesis predicts slower adoption of proactive AI for hiring in high-exposure jurisdictions, controlling for firm size and technical capability. The cross-jurisdictional variation provides an identification strategy that is closer to causal.¹²

The strategy can be deepened with case-study evidence on specific firms that have publicly described their adoption portfolios. Large financial institutions — JPMorgan Chase, Bank of America, Goldman Sachs — have published sufficient detail in annual reports and investor communications to permit reconstruction of their deployment composition. Healthcare systems (Kaiser Permanente, HCA, the Mayo Clinic) have done similarly. The published-detail data is incomplete but allows triangulation against the survey-based data and provides qualitative depth on the institutional process of adjudicative-deployment decisions.

C. Strategy 3: Event Studies Around Liability Shocks

The third strategy addresses Prediction 3 — that major adverse judgments or regulatory enforcement actions against proactive AI deployments produce measurable adoption pauses or reversals in peer firms.

The construction is an event-study framework familiar from the finance and labor economics literatures. The events are publicly observable adverse outcomes — judgments, settlements, regulatory enforcement actions, high-salience press episodes — that establish liability or reputational cost for a proactive AI deployment in a specific sector. The event window is the period before and after the event, typically 12 months on each side. The outcome variable is the rate of new proactive AI deployments in the same functional area across peer firms in the same sector and adjacent sectors.

Several events suitable for this analysis have already occurred. The Michigan MiDAS unemployment fraud system collapse, culminating in the Bauserman litigation, established a liability precedent that has plausibly slowed state-level adoption of automated fraud detection in adjacent benefits programs.¹³ The Australian Robodebt collapse, the most extensively documented case of substitutive AI deployment in a public benefits system, produced a multi-billion-dollar settlement and a Royal Commission report; the post-event trajectory of automated benefits administration in Commonwealth jurisdictions is observable. The Dutch SyRI ruling by the District Court of The Hague (2020) struck down an algorithmic fraud detection system on human-rights grounds; subsequent EU member-state behavior on similar systems is observable. The State v. Loomis progeny in U.S. state courts has produced a slow but measurable retrenchment in the use of pretrial risk assessment tools in several jurisdictions.

The empirical strategy is to identify, for each event, the set of peer firms or peer jurisdictions facing comparable deployment decisions, and to compare deployment rates in the post-event period to the pre-event trend. The principal identification problem is that the events are themselves selected — events that produced legal liability are events in which the deployment had visible failure modes, and the firms that responded to the event are firms that had not yet committed to the deployment. The selection bias is partially addressable by comparing to control groups of firms in adjacent sectors where the event was nonetheless visible (through press coverage, through trade publications, through industry conferences) but where the event was not directly relevant.

The strategy is most powerful when applied to clusters of related events. The 2024-2025 wave of NYC Local Law 144 enforcement actions, for instance, produced a measurable retrenchment in automated hiring tool adoption among New York-based employers that is documentable through vendor revenue data and through job-posting analysis. The cluster permits multiple event windows, multiple sectoral comparisons, and richer identification than any single event would allow.

A complementary strategy uses the converse: events that expanded the legal scope for proactive AI deployment (for instance, the 2025 federal AI Bill of Rights walk-back, or specific safe-harbor regulatory developments) should, if the bifurcation hypothesis is correct, produce accelerated adoption in the relevant functional areas. The asymmetry of adoption response to liability-expanding versus liability-contracting events provides additional identification.

D. Measurement Challenges

All three strategies face measurement challenges that deserve explicit discussion.

The first is the operational definition of "proactive AI adoption." As noted in Part III.D, the strategically valuable ambiguity between nominal and functional substitution means that survey-based measures of adoption will undercount functional substitution that is described in advisory terms. The strategies above address this by triangulating survey-based measures with vendor revenue data (which counts deployments without reference to deployer description) and with override-rate data where available (which captures functional substitution that survey data misses). The triangulation is imperfect; the measurement error is most plausibly downward-biased for adjudicative deployments (where strategic ambiguity is most valuable) and the bias works against the bifurcation hypothesis, making any positive finding conservative.

The second is the operational definition of "tort exposure." Insurance premiums are an imperfect proxy because they reflect both the underlying exposure and the structure of the insurance market. Synthetic measures based on litigation databases are imperfect because they reflect the strategic decisions of claimants to file and of firms to settle. The empirical strategies above use multiple measures and report robustness across measure choice; no single measure is definitive.

The third is the cross-sectoral comparability of functional-area classification. What counts as "operational" versus "adjudicative" in finance may not have a direct analogue in manufacturing or healthcare. The strategies above adopt explicit definitional criteria (third-party externalization of error costs, routinely litigated outputs) and apply them uniformly, but the cross-sectoral application requires judgment and the robustness of findings to alternative classifications should be examined.

These measurement challenges are real but tractable. They do not constitute reasons to abandon the empirical program; they constitute the methodological agenda the program implies.

VI. Sector-Specific Predictions

The bifurcation hypothesis generates predictions that vary across sectors as the operational/adjudicative composition of functional areas varies and as sector-specific liability regimes vary. Six sectors are sufficient to illustrate the predictions and to demonstrate the hypothesis's discriminating power.

A. Finance

The financial sector contains both the cleanest operational and the most adjudicative functional areas, making it particularly diagnostic. Operationally, proactive AI for trade settlement, fraud monitoring at the transaction level, regulatory reporting automation, and internal compliance triage should achieve near-saturation deployment within the five-year window. The errors in these areas are small in magnitude, internally absorbed, and recoverable; the legal regime is stable (these are well-established back-office functions with mature regulatory expectations). Adjudicatively, capital allocation — particularly the underwriting of corporate credit, the approval of large loans, the management of trading authority in slow-cycle asset classes, and discretionary portfolio decisions — should remain substantially human-locus, with adoption rates below 30% even at the end of the observation window. The pattern is already visible in the deployment portfolios of large investment banks, which describe automated trading at the high-frequency end (where the operational/adjudicative distinction collapses because tail errors are bounded by position limits and recovery cycles are short) while preserving human authority over investment committee decisions, even when machine analytics inform those decisions extensively.

The diagnostic significance of finance is that the same firm can be observed simultaneously operating in both modes. JPMorgan Chase deploys proactive AI for transaction monitoring at near-saturation while preserving human authority over corporate credit decisions and over major capital commitments. Goldman Sachs has automated significant portions of post-trade processing while leaving investment-banking advisory work in conventional human form. If the bifurcation hypothesis is correct, this internal composition should persist and intensify; if competitive pressure dominates, the within-firm gap should narrow over the observation window. The prediction is that the gap persists.

A potential complication is the high-frequency trading case, in which proactive AI has already achieved near-total deployment for decisions that, in any meaningful sense, are adjudicative — they bind counterparties, produce externalized costs, and are subject to regulatory scrutiny. The honest reading is that high-frequency trading is the partial counterexample to the bifurcation hypothesis. But the partial counterexample is itself informative: the bifurcation hypothesis predicts that conditions of high competitive pressure plus rapid feedback loops plus narrow per-decision exposure can overwhelm the liability-driven bifurcation. High-frequency trading meets all three conditions. The general adjudicative case in finance — long-horizon credit decisions, illiquid asset valuation, fiduciary investment decisions — does not.

B. Healthcare

The healthcare sector exhibits the bifurcation in its starkest form. Administrative proactive AI — claims processing, appointment scheduling, prior authorization triage, supply-chain management, clinical documentation summarization — should achieve high deployment rates within the five-year window, and is already approaching majority deployment in large integrated health systems. Clinical proactive AI — diagnosis without physician review, treatment recommendation without physician adoption, autonomous prescription, autonomous referral — should remain rare and confined to narrow specialty applications (imaging triage, dermatology screening) where the deployment is structured as advisory and the physician's adoption rate is measurable and contestable.

The mechanism in healthcare is unusually transparent. Medical malpractice exposure is the dominant determinant of physician practice patterns at the margin and has been thoroughly studied. Hospitals and physician groups facing high malpractice premiums respond by adopting practice patterns that minimize litigation risk, including increased use of defensive medicine and conservative diagnostic protocols. The deployment of proactive AI in clinical contexts collides directly with this incentive structure: the physician is legally and professionally responsible for the patient's outcome, the AI system is not, and the physician therefore has direct personal incentive to maintain the locus of decision regardless of the system's average performance.

The prediction for the five-year window is that clinical proactive AI deployment will remain confined to (i) narrow imaging applications where the radiologist remains the locus, (ii) inpatient triage applications where the system flags but does not act, and (iii) specific protocol-driven contexts (sepsis early warning, ICU alerting) where the system's role is to alert a human decider rather than to act. Diagnostic deployments, treatment recommendation deployments, and autonomous prescription deployments will remain rare and contested. The five-year aggregate share of clinical decisions taken by proactive AI should not exceed 5% even in the most aggressive deployment environments.

The interesting complication in healthcare is the convergence of capability — generative medical AI is approaching parity with physician performance on diagnostic benchmarks for several specialties.¹⁴ The bifurcation hypothesis predicts that capability parity, however measured, will not produce deployment parity in clinical contexts within the observation window. If capability parity does produce deployment parity, the hypothesis is weakened.

C. Legal Services

The legal sector exhibits the bifurcation along the routine-drafting versus advocacy axis. Routine document drafting, contract review, due diligence, document discovery review, and legal research should see rapid deployment of proactive AI within the five-year window. These tasks have moderate per-error costs, identifiable client-attorney quality checks, and a mature professional-liability regime in which the attorney remains responsible for the output. Adjudicative advocacy — trial strategy, negotiation strategy, sentencing memorandum drafting in criminal cases, brief drafting in matters of significant import — should remain attorney-locus, with proactive AI playing a supporting role rather than acting autonomously.

The professional-liability regime in legal services is more developed than in many adjudicative sectors and provides a particularly clean test of the hypothesis. Attorney malpractice exposure is well-characterized, malpractice insurance is universal, and the bar associations have begun issuing guidance on the use of AI tools that explicitly preserves attorney responsibility for the output.¹⁵ The bifurcation hypothesis predicts that the guidance will be observed even where capability would support more autonomous deployment, and that the rate of deployment will track the legal-doctrinal development of attorney AI-use standards rather than the underlying capability trajectory.

The prediction is observable. Within the five-year window, the proportion of attorney-hours allocated to document review and routine drafting should fall by at least 40% in large law firms, reflecting the operational deployment of proactive AI; the proportion of attorney-hours allocated to advocacy, negotiation, and client counseling should remain substantially stable or grow, reflecting the absence of substitution in the adjudicative layer. The composition shift within the law firm is the predicted pattern.

D. Insurance

Insurance is structurally interesting because it contains both adjudicative functions (claims adjudication, underwriting) and a sector-level mechanism (the insurance market itself) that prices liability exposure and therefore makes the bifurcation mechanism partially endogenous. The prediction is that claims processing — particularly the high-volume, low-value end of the claims market (auto property damage, simple homeowners claims) — will see rapid proactive AI deployment within the observation window. Underwriting — particularly underwriting of products with long tails of potential loss (life insurance, professional liability, complex commercial coverage) — should remain substantially human-locus.

A specific within-sector prediction: claims denial in particular should adopt proactive AI more slowly than claims approval, even where the same model could produce both decisions. The reason is the asymmetry of legal exposure — wrongful denial generates bad-faith litigation and regulatory scrutiny in ways that wrongful approval does not — and the bifurcation hypothesis predicts that the legal asymmetry will be reflected in the deployment asymmetry. Auto insurers in particular have begun automating claims approval at high rates while preserving human review for denials, contestations, and reservation-of-rights determinations. The pattern is consistent with the hypothesis.

E. Manufacturing

Manufacturing is the sector closest to the historical computerization wave and provides the most direct test of whether the bifurcation hypothesis is genuinely distinctive or is merely re-describing the standard task-automation dynamics. Operational deployments — production scheduling, supply chain optimization, predictive maintenance, quality control inspection — should achieve near-saturation deployment within the five-year window in large manufacturers. These deployments continue the trajectory of factory-floor automation that has been underway for four decades and present no novel liability profile.

Adjudicative deployments in manufacturing are rarer but consequential. Product design decisions — particularly safety-critical design decisions in regulated industries (automotive, aerospace, medical devices) — should remain substantially human-locus, with proactive AI playing an analytic role rather than a deciding role. The relevant liability regime is product liability, which is unusually severe in the United States and which loads strict liability on the manufacturer for design defects. The bifurcation hypothesis predicts that adjudicative deployments in design will be confined to non-safety-critical applications and that safety-critical design decisions will continue to require human sign-off even where AI analytics dominate the design process.

The diagnostic value of manufacturing is that the operational deployments are large enough in aggregate magnitude to dominate the headline adoption figures while the adjudicative deployments remain rare. A firm reporting that "85% of design tasks involve AI" may be reporting heavy use of AI in analysis, simulation, and routine drafting while preserving human authority for the safety-critical decisions that drive the firm's product liability exposure. The composition matters.

F. Professional Services

Professional services — management consulting, accounting, financial advisory, research services — exhibit the bifurcation along the research/advisory axis. Internal research, data analysis, document preparation, and routine deliverable production should see rapid proactive AI deployment within the five-year window. Client advisory work — particularly advisory in matters of significant client consequence, where the consulting firm or advisory firm's reputation is engaged and where adverse client outcomes generate professional-liability exposure — should remain principal-locus, with proactive AI playing a supporting role.

The mechanism in professional services is reputational rather than tort-driven for most sub-sectors (accounting, with its statutory liability regime under Sarbanes-Oxley, is the exception). The reputational mechanism is structurally similar to the tort mechanism in producing high option value of waiting: a high-salience adverse outcome attributed to AI advisory in a major engagement can damage the firm's reputation in a way that exceeds the productivity gain from many engagements of accelerated AI deployment. The prediction is that the major consulting firms will continue to use AI extensively for research and deliverable production while preserving senior consultant authority for client recommendations, particularly recommendations with strategic or financial significance for the client.

The honest qualification is that professional services is the sector in which competitive pressure to demonstrate AI adoption is most intense — clients themselves expect to see AI-driven productivity gains reflected in pricing and turnaround. The competitive pressure cuts against the bifurcation prediction for advisory work; if the prediction holds in this sector, it is strong evidence for the hypothesis. If it does not, the competitive-pressure alternative gains support.

VII. The Competitive-Pressure Alternative

The most serious alternative to the bifurcation hypothesis is the competitive-pressure hypothesis: that whatever the structural logic of liability-asymmetric adoption, competitive dynamics within sectors will force uniform adoption against the structural logic, because the first-mover productivity gain — even in adjudicative contexts — is large enough that abstaining firms will be displaced by adopting competitors. The competitive-pressure hypothesis predicts not bifurcation but lagged convergence: adjudicative adoption follows operational adoption with a lag, but the lag closes as competitive pressure mounts, and the equilibrium is uniform adoption across functional areas.

A. Why the Alternative Is Serious

The competitive-pressure hypothesis is serious for several reasons. First, the historical record of prior technology adoption waves shows few persistent within-firm asymmetries; the lagged-convergence pattern is more common than the persistent-bifurcation pattern. Second, the productivity gains from adjudicative AI deployment, where they have been measured, are sufficiently large that the standard competitive-pressure analysis would predict adoption irrespective of liability exposure — the gain on the firm's expected value of operations exceeds the increase in expected liability cost for many plausible parameter values. Third, the institutional infrastructure for transferring liability risk (insurance markets, indemnification provisions in vendor contracts, the development of professional standards that limit individual professional liability for AI-mediated decisions) is itself responsive to competitive pressure: where the productivity gain from adoption is large, the institutional infrastructure to absorb the liability adapts. Fourth, the development of AI-specific liability regimes — most prominently the EU AI Act's risk-tiered liability framework and proposed U.S. federal legislation — may produce a uniform safe harbor that reduces the cross-sectoral variation in tort exposure that bifurcation relies on.

The competitive-pressure hypothesis is not a strawman. It is the framework against which the bifurcation hypothesis must be tested.

B. Differential Predictions

The two hypotheses generate different predicted dynamics in three measurable respects.

First, the time-trajectory of adjudicative adoption differs. The bifurcation hypothesis predicts persistent asymmetry — adoption in adjudicative areas plateaus at a level substantially below operational areas and remains there. The competitive-pressure hypothesis predicts lagged convergence — adoption in adjudicative areas tracks operational adoption with a lag that closes over time as competitive pressure mounts and institutional infrastructure develops.

Second, the response to liability shocks differs. The bifurcation hypothesis predicts that adverse liability events in a sector produce measurable deployment pauses or reversals in adjacent firms — the events confirm the option value of waiting and reinforce the bifurcation. The competitive-pressure hypothesis predicts that adverse liability events produce localized responses (the directly affected firm pauses or reverses; institutional infrastructure adapts) but no sector-wide retrenchment, because competitive pressure dominates the event-specific response.

Third, the cross-sectoral pattern differs. The bifurcation hypothesis predicts that the negative correlation between adjudicative adoption and tort exposure persists and is robust across observation periods. The competitive-pressure hypothesis predicts that the correlation, if initially present, attenuates over time as institutional infrastructure compensates for the initial liability differential.

C. What the Test Will Show

The honest forecast is that both mechanisms operate and that the empirical question is their relative magnitude. The bifurcation hypothesis does not claim that competitive pressure is absent; it claims that the liability-driven option value of waiting dominates the competitive pressure in adjudicative contexts during the observation window. The competitive-pressure hypothesis does not claim that liability exposure is irrelevant; it claims that the competitive pressure dominates the liability-driven waiting incentive in the long run, even where liability exposure is substantial.

The five-year observation window is long enough to detect the broad shape of the dynamic but short enough that the long-run convergence the competitive-pressure hypothesis predicts may not yet be observable. The empirical test will most likely produce evidence consistent with both mechanisms operating, with the relative magnitude varying by sector. The bifurcation hypothesis is supported if (i) the asymmetry persists, (ii) the response to liability shocks shows sector-wide retrenchment, and (iii) the cross-sectoral correlation persists. The competitive-pressure hypothesis is supported if (i) convergence is observed in at least two sectors over the observation window, (ii) liability shocks produce only localized responses, and (iii) the cross-sectoral correlation attenuates.

The most interesting outcome — and the one the present paper expects — is that the bifurcation pattern holds for the observation window in most sectors, with one or two sectors (high-frequency trading, possibly automated claims processing) showing convergence consistent with the competitive-pressure hypothesis. This would be evidence that the bifurcation mechanism is real but is dominated by competitive pressure under specific conditions (high competitive intensity, short feedback loops, narrow per-decision exposure). The boundary conditions would themselves be a substantive finding.

D. The Endogeneity of Liability Regimes

A deeper version of the competitive-pressure objection holds that liability regimes are themselves endogenous to competitive pressure: where the productivity gain from AI deployment is large, the legal regime adapts to permit deployment, either through legislative safe harbors or through judicial doctrines that limit deployer exposure. The objection has historical force — environmental liability regimes adapted to permit industrial activity, financial-sector liability regimes adapted to permit innovative products, and so forth. The same dynamic may operate for AI liability.

The bifurcation hypothesis acknowledges the endogeneity but argues that it operates on a slower clock than the deployment decision. Legal regimes adapt over decades through legislation, case law, and regulatory development; deployment decisions are made on annual or sub-annual cycles. Within the five-year observation window, the legal regime is approximately exogenous from the perspective of any individual deployer, and the option value of waiting is computed against the regime as it exists rather than against the regime as it may eventually evolve. The endogeneity matters for the long-run equilibrium but not for the observation-window dynamic.

The right response to the endogeneity objection is to specify a longer observation window for the long-run prediction. If the bifurcation hypothesis holds over the five-year window but is followed by convergence over a fifteen-year window, the hypothesis is correct as a description of the medium-run dynamic and the competitive-pressure hypothesis is correct as a description of the long-run equilibrium. Both can be true.

VIII. Counterarguments

A. Capability Will Resolve the Asymmetry

A first counterargument holds that the bifurcation reflects current capability limitations and will dissolve as model capability improves. The argument is that adjudicative deployments are slow because current systems are not yet reliable enough; once they are, the bifurcation will close.

The counterargument misreads the mechanism. The bifurcation hypothesis does not depend on current capability limitations; it depends on the Knightian uncertainty about tail behavior that is, in important respects, increasing rather than decreasing with capability. More capable systems are also more general systems with larger behavioral surfaces and more emergent failure modes; the tail of the deployment distribution becomes harder to characterize as capability grows, not easier. The empirical evidence on the relationship between model capability and tail-risk characterizability is, at present, more consistent with the harder-to-characterize reading than with the easier-to-characterize reading. Capability improvements that move the median performance higher do not necessarily reduce tail risk and may increase it.

Moreover, the liability mechanism that drives bifurcation is robust to capability improvements. A system that performs better on average does not, by improved average performance, reduce the deployer's exposure to a single high-salience tail event. The first major adverse judgment against an autonomous medical AI deployment will be just as damaging if the underlying system has 99.9% accuracy as if it has 95% accuracy, and possibly more damaging because the deployer's reliance on the system was greater.

B. Insurance Will Absorb the Risk

A second counterargument holds that insurance markets will develop products that absorb the residual risk of adjudicative AI deployments, and that once such products are available the bifurcation mechanism collapses. The counterargument has some force — insurance markets have absorbed novel risk classes before, and AI-specific insurance products are beginning to emerge.

The honest response is that insurance is part of the solution but not the whole solution. Insurance products price risk; they do not eliminate it. Where the tail of the loss distribution is Knightianly uncertain, insurance products will either price the uncertainty at a high premium (making deployment uneconomic) or exclude the uncertainty (leaving the deployer to bear the residual). The current generation of AI-specific insurance products generally do the latter, with broad exclusions for novel deployment categories and with sub-limits that cap insurer exposure well below the deployer's potential loss in a high-salience adverse outcome. Until the insurance market has characterized the tail well enough to price it without exclusion, insurance is a partial palliative, not a full solution. The bifurcation hypothesis predicts that the insurance market will mature along the same slow clock as the legal regime, and for related reasons.

C. The Hypothesis Is True But Trivial

A third counterargument holds that the bifurcation hypothesis is true but trivial — that it merely re-describes the obvious observation that high-stakes deployments are harder than low-stakes deployments, and that no contribution to the literature is achieved by labeling this observation a "hypothesis."

The trivality objection has surface plausibility but misses three substantive contributions. First, the hypothesis specifies the mechanism — Knightian uncertainty plus liability exposure plus real-options waiting value — that produces the asymmetry, where prior treatments have either left the mechanism unspecified or attributed it to the wrong factor (typically, technical capability rather than legal exposure). Second, the hypothesis generates testable predictions with stated directional signs, where prior treatments have been directionally suggestive rather than testably specific. Third, the hypothesis identifies a measurement program — within-aggregate composition, cross-sectoral correlation with tort exposure, event-study response to liability shocks — that has not previously been organized as a coherent empirical agenda. The contribution is the specification, not the observation.

The triviality objection would have force only if the empirical predictions were uncontroversially expected. They are not. Several major consultancy forecasts predict near-uniform adoption across functional areas; several academic treatments treat the within-firm composition as unimportant. The hypothesis disagrees with these treatments and predicts a different pattern. Whether the prediction is correct is an empirical question; that the prediction is non-trivial is a function of the literature against which it is offered.

IX. Conclusion and Research Agenda

This paper has argued that proactive AI will diffuse asymmetrically within the firm, with operational and transactional functional areas absorbing the technology rapidly and adjudicative functional areas absorbing it slowly, haltingly, and reversibly. The mechanism is the interaction of Knightian uncertainty about agent behavior in the tail of the deployment distribution with a liability regime that loads residual risk on the deployer rather than on the agent. This combination produces a real-options waiting value that is high precisely where the productivity gain from deployment would be largest, generating the structural bifurcation the paper documents.

The hypothesis has been decomposed into four testable predictions, situated within the established technology adoption literature, paired with three empirical strategies developed in sufficient detail for execution, and applied to six sectors with specific predictions accountable to observation over a five-year window. The principal rival hypothesis — that competitive pressure will force uniform adoption against the structural logic — has been treated seriously, and the differential empirical predictions distinguishing the two hypotheses have been specified.

The research agenda implied by the paper has three components. The first is the execution of the empirical strategies of Part V. The cross-sectoral correlation analysis is tractable with existing data and could be completed within twelve months. The within-firm functional-area analysis requires panel data of a kind that is partially available through Stanford and MIT Sloan instruments and partially would need to be assembled. The event-study analysis is tractable for the events that have already occurred and provides immediate purchase on Prediction 3. The second component is the construction of better measures of the underlying constructs — tort exposure, adjudicative character of functional areas, deployment-with-functional-locus rather than nominal deployment — that would strengthen the empirical analysis in subsequent waves. The third component is the extension of the framework to non-U.S. jurisdictions, where the liability regime differs substantially and where the bifurcation hypothesis predicts correspondingly different deployment patterns. The European Union, with its tiered AI Act risk classification, and the Anglo-Commonwealth jurisdictions with their varying tort regimes, provide natural comparison cases.

The paper closes with an observation about scope. The bifurcation hypothesis, if it obtains, has implications beyond firm strategy and beyond the academic study of technology adoption. It predicts the institutional sites at which the political contest over AI's penetration into authority structures will be most intense — the adjudicative functional areas where the deployment decision is most contested and where the consequences of deployment are most legible to the public. It identifies the legal regime — tort and fiduciary law — that will functionally regulate proactive AI in those sites long before any AI-specific statute does so, because the liability mechanism operates whether or not specific regulation has been enacted. And it suggests that the political and regulatory debate about AI governance, currently structured around questions of bias, transparency, and explainability, should be re-centered on the question of liability assignment, because liability assignment is what the deployment decision actually turns on. The bifurcation hypothesis is a hypothesis about firms, but the implications run outward to questions about how authority is allocated in a society in which a meaningful share of consequential decisions may, or may not, be taken by systems that act on their own initiative.

---

References

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.

Acemoglu, D., and Restrepo, P. (2020). "Robots and Jobs: Evidence from US Labor Markets." Journal of Political Economy 128.

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

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

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

Avraham, R., and Schanzenbach, M. (2015). "The Impact of Tort Reform on Intensity of Treatment: Evidence from Heart Patients." Journal of Health Economics 39.

Bresnahan, T. (2010). "General Purpose Technologies." In Handbook of the Economics of Innovation, vol. 2.

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

Brodeur, A., et al. (2024). "Diagnostic Accuracy of Large Language Models in Clinical Vignettes." JAMA Network Open 7.

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

Brynjolfsson, E., Hitt, L., and Yang, S. (2002). "Intangible Assets: Computers and Organizational Capital." Brookings Papers on Economic Activity 33.

Brynjolfsson, E., and McAfee, A. (2014). The Second Machine Age. W. W. Norton.

Brynjolfsson, E., and McAfee, A. (2017). "The Business of Artificial Intelligence." Harvard Business Review, July.

Brynjolfsson, E., Rock, D., and Syverson, C. (2021). "The Productivity J-Curve: How Intangibles Complement General Purpose Technologies." American Economic Journal: Macroeconomics 13.

Crafts, N. (2021). "Artificial Intelligence as a General-Purpose Technology: An Historical Perspective." Oxford Review of Economic Policy 37.

Currie, J., and MacLeod, W. B. (2008). "First Do No Harm? Tort Reform and Birth Outcomes." Quarterly Journal of Economics 123.

David, P. (1985). "Clio and the Economics of QWERTY." American Economic Review 75.

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

David, P. (1991). "Computer and Dynamo: The Modern Productivity Paradox in a Not-Too-Distant Mirror." In Technology and Productivity: The Challenge for Economic Policy. OECD.

Dixit, A., and Pindyck, R. (1994). Investment Under Uncertainty. Princeton.

Eubanks, V. (2018). Automating Inequality. St. Martin's.

Gilboa, I., and Schmeidler, D. (1989). "Maxmin Expected Utility with Non-Unique Prior." Journal of Mathematical Economics 18.

Goddard, K., Roudsari, A., and Wyatt, J. (2012). "Automation Bias: A Systematic Review of Frequency, Effect Mediators, and Mitigators." Journal of the American Medical Informatics Association 19.

Green, B., and Chen, Y. (2019). "Disparate Interactions: An Algorithm-in-the-Loop Analysis of Fairness in Risk Assessments." Proceedings of the Conference on Fairness, Accountability, and Transparency.

Helpman, E., ed. (1998). General Purpose Technologies and Economic Growth. MIT.

Klibanoff, P., Marinacci, M., and Mukerji, S. (2005). "A Smooth Model of Decision Making Under Ambiguity." Econometrica 73.

Knight, F. (1921). Risk, Uncertainty and Profit. Houghton Mifflin.

McDonald, R., and Siegel, D. (1986). "The Value of Waiting to Invest." Quarterly Journal of Economics 101.

Rogers, E. (2003). Diffusion of Innovations, 5th ed. Free Press.

Singhal, K., et al. (2023). "Large Language Models Encode Clinical Knowledge." Nature 620.

Singhal, K., et al. (2025). "Toward Expert-Level Medical Question Answering with Large Language Models." Nature Medicine 31.

Skitka, L., Mosier, K., and Burdick, M. (1999). "Does Automation Bias Decision-Making?" International Journal of Human-Computer Studies 51.

Stanford Institute for Human-Centered Artificial Intelligence (2024, 2025, 2026). Artificial Intelligence Index Report. Stanford University.

Zerega, H. (2026). "Augmentative or Substitutive? A Constitutional Typology of State Artificial Intelligence." Companion working paper.

---

Bauserman v. Unemployment Insurance Agency, 503 Mich. 169 (2019). State v. Loomis, 881 N.W.2d 749 (Wis. 2016). Thaler v. Vidal, 43 F.4th 1207 (Fed. Cir. 2022).

Footnotes

1. Bresnahan and Trajtenberg (1995); Brynjolfsson and McAfee (2014, 2017); Brynjolfsson, Rock, and Syverson (2021). The general purpose technology framework was developed to explain the staggered productivity gains from electrification and the computer; its application to AI is largely faithful to the original framework, with adjustments for the speed of capability improvement. ↩

2. See Zerega, "Augmentative or Substitutive? A Constitutional Typology of State Artificial Intelligence" (2026). The present paper is concerned with the same locus distinction but applied to private firms rather than state actors and oriented toward predicting adoption dynamics rather than identifying constitutional categories. ↩

3. Knight (1921), distinguishing risk from uncertainty. The distinction has been challenged by Bayesian decision theorists who argue that any subjective probability assignment converts uncertainty to risk; the argument here adopts the older Knightian view, which has been rehabilitated in the recent decision-theory literature on ambiguity aversion (Gilboa and Schmeidler 1989; Klibanoff, Marinacci, and Mukerji 2005) and which is, in practice, the framework deployer counsel use when advising on novel deployments. ↩

4. The recent line of cases attempting to assign legal personality to AI systems for purposes of patent inventorship (Thaler v. Vidal, 43 F.4th 1207 (Fed. Cir. 2022)) and copyright authorship has been uniformly negative. The fiduciary analogue is similarly closed. There is no contemporary legal regime in which the proactive AI agent itself is the residual bearer of risk. ↩

5. Dixit and Pindyck (1994); McDonald and Siegel (1986). The real-options framework was developed for capital investment under price uncertainty but is general to any irreversible commitment under uncertainty resolvable by waiting. ↩

6. See Skitka, Mosier, and Burdick (1999); Goddard, Roudsari, and Wyatt (2012); Green and Chen (2019). The phenomenon is well-documented across domains. The strategic implication for proactive AI deployers is that nominal human-locus operation is cheaper than genuine human-locus operation and may produce most of the productivity gain. ↩

7. Rogers (2003), 5th edition; first edition 1962. The framework was developed for agricultural extension and has been applied to technology adoption broadly. ↩

8. David (1990, 1991); David (1985) on QWERTY and path dependence. The productivity-paradox literature has been extended by Brynjolfsson and Hitt (2000) and Brynjolfsson, Hitt, and Yang (2002) into the contemporary computing context, with the original finding largely vindicated by the eventual emergence of measurable computer-driven productivity gains in the late 1990s. ↩

9. Bresnahan and Trajtenberg (1995); Bresnahan (2010); Helpman (1998); Crafts (2021). The framework's central claim is that GPTs produce productivity gains primarily through the complementary innovations they spawn rather than through the direct application of the technology itself. ↩

10. Acemoglu and Restrepo (2018, 2020, 2022). The framework draws on the earlier task-based literature (Autor, Levy, and Murnane 2003) and incorporates the firm-organization literature in distinguishing displacement, reinstatement, and productivity effects. ↩

11. The literature on medical malpractice reform and physician behavior (e.g., Currie and MacLeod 2008; Avraham and Schanzenbach 2015) has used exactly this variation to identify causal effects of liability exposure on physician practice. The application to AI adoption is novel but methodologically continuous. ↩

12. New York City's Local Law 144 (2023), requiring bias audits of automated employment decision tools, and the California ADS regulations (effective 2026 in their current form) constitute the highest-exposure jurisdictional regime. Texas and the southeastern states constitute the comparison group. The literature on Title VII enforcement variation provides a methodological template. ↩

13. Bauserman v. Unemployment Insurance Agency, 503 Mich. 169 (2019). The case is the leading U.S. authority on automated benefits adjudication and has been cited in subsequent state-level litigation; its citation pattern is itself a measurable indicator of doctrinal spread. ↩

14. The recent benchmark literature (Singhal et al. 2023, 2025; Brodeur et al. 2024) documents physician-level performance on medical question-answering and on diagnostic vignettes for several large model families. The benchmarks are imperfect proxies for clinical performance but are informative about the capability trajectory. ↩

15. ABA Model Rule 1.1 Comment 8, as amended in 2024 to address generative AI tools; California State Bar's 2023 Practical Guidance for the Use of Generative Artificial Intelligence in the Practice of Law; Florida Bar Ethics Opinion 24-1. ↩