Organizations increasingly replace variable, relationship-dependent, and judgment-intensive capabilities with engineered substitutes: automated service agents replace human representatives, algorithmic scores replace deliberation, dashboards replace local knowledge, scripts replace discretion, and standardized platforms replace context-sensitive routines. Existing research explains automation, technology acceptance, algorithmic management, organizational ambidexterity, service relationships, and sociotechnical systems, but it lacks a unified theory explaining why an engineered substitute can appear superior while progressively weakening the adaptive capability that created value in the first place. This thesis proposes Robotic Dog Theory (RDT), a middle-range theory of engineered substitution.
The robotic dog is a generative analogy: it can imitate selected outputs of a living dog—availability, responsiveness, predictability, and low maintenance-while lacking the full reciprocal, emergent, and context-sensitive relationship through which many benefits of companionship arise. RDT argues that firms adopt engineered substitutes when the perceived cost of biological or social variability exceeds the perceived value of adaptive reciprocity. Substitution initially improves controllability, scalability, measurability, and unit efficiency. However, when the substitute becomes dominant in a task environment characterized by ambiguity, novelty, emotional consequence, or exception diversity, the organization loses signal variety, tacit knowledge, relational legitimacy, and local recovery capacity. The resulting accumulation is termed adaptive-value debt: a deferred liability created when current efficiency is purchased by reducing the system’s future ability to interpret, learn, and respond.
The theory specifies eight constructs, a staged process model, 14 propositions, boundary conditions, a formal model, operational measures, multilevel research designs, rival explanations, and managerial applications. It predicts a nonlinear relationship between engineered substitution and organizational performance, moderated by environmental variety, task equivocality, consequence severity, reversibility, and retained human agency. RDT contributes a falsifiable explanation for why automation succeeds in some domains, disappoints in others, and becomes dangerous when managers confuse behavioral resemblance with functional equivalence.
Keywords: Robotic Dog Theory, engineered substitution, automation, artificial intelligence, adaptive-value debt, organizational learning, relational value, algorithmic management, sociotechnical systems
Introduction
A robotic dog can be designed to sit, respond to a name, follow its owner, simulate affection, require no food, shed no hair, and remain available on demand. Evaluated against a narrow list of desirable behaviors, it may appear to be a superior dog. It is predictable, measurable, programmable, and inexpensive to reproduce. Yet the judgment depends entirely on what “dog” means. If the desired object is a compliant bundle of outputs, the robot may be adequate or better. If the desired relationship includes mutual adaptation, spontaneous behavior, embodied vulnerability, trust developed over time, and the possibility that each participant changes the other, resemblance is not equivalence.
Modern organizations face this problem at scale. A chatbot resembles a service representative because it answers questions. A productivity score resembles managerial judgment because it ranks performance. A dashboard resembles operational understanding because it displays indicators. A standardized script resembles expertise because it encodes a preferred response. A digital platform resembles a market relationship because it coordinates exchange. These substitutes are not inherently deficient. Many are faster, safer, more accessible, and more consistent than what they replace. The theoretical problem is that organizations frequently evaluate them by the outputs the substitute was designed to imitate, while excluding the less visible capabilities of the original system from the comparison.
This produces a systematic bias. Engineered substitutes are legible to management because their outputs can be specified and counted. Living and social systems produce variation, tacit judgment, emotional reciprocity, anomaly detection, improvisation, and learning that are difficult to isolate. The substitute is credited for lower current costs; the original is blamed for inconsistency. Lost adaptability often becomes visible only later, when a novel customer need, crisis, exception, ethical conflict, or market discontinuity exceeds the substitute’s model.
Robotic Dog Theory (RDT) explains this pattern. It is a proposed middle-range business theory addressing the selection, adoption, and consequences of engineered substitutes for adaptive human, social, or ecological capabilities. The theory does not claim that automation is false, inhuman, or inevitably harmful. It predicts when substitution will create value, when hybridization will outperform replacement, and when short-term control will generate a hidden future liability.
The thesis answers five questions:
- Why do organizations systematically overvalue the controllability of engineered substitutes?
- Which forms of value are most likely to be omitted when a substitute is evaluated?
- Under what conditions does substitution improve sustained performance?
- When does substitution create adaptive-value debt?
- How can the theory be measured, tested, challenged, and used in managerial decisions?
The central claim is that the net value of engineered substitution depends not on how closely the substitute reproduces visible behavior, but on whether the organization preserves the variety, reciprocity, judgment, and recovery capacity required by the environment. In stable, repetitive, observable, low-consequence, and reversible work, a robotic substitute can produce genuine gains. In ambiguous, novel, emotionally consequential, tightly coupled, or irreversible work, aggressive substitution can create a control paradox: the organization becomes more consistent in expected conditions and less capable in the conditions that matter most.
Theoretical Status and Claim of Originality
RDT is presented as a new theoretical synthesis and named explanatory framework. Absolute historical novelty cannot be proved; related insights appear in research on the ironies of automation, requisite variety, sociotechnical systems, organizational learning, relationship marketing, service automation, and algorithmic management. A credible originality claim must therefore be narrower. RDT’s proposed novelty lies in combining five elements within a single causal theory:
- the distinction between behavioral resemblance and capability equivalence;
- control attraction as the mechanism encouraging over-substitution;
- adaptive reciprocity as a form of productive, two-way adjustment that engineered substitutes may attenuate;
- substitution blindness as a measurement and governance failure; and
- adaptive-value debt as the delayed liability produced by lost signal variety, tacit knowledge, legitimacy, and recovery capacity.
The robotic-dog analogy is not the evidence for the theory. It is a compact representation of the theory’s causal logic. The theoretical contribution must stand independently through defined constructs, predicted relationships, boundary conditions, empirical measures, and the possibility of refutation.
Literature Foundations and the Unresolved Gap
Bounded rationality and managerial legibility
Organizations cannot perceive or optimize every relevant feature of a complex environment. Decision makers operate under bounded rationality and seek satisfactory rather than perfectly optimal choices (Simon, 1955). Standardization and measurement reduce cognitive load. They render work comparable, governable, and scalable. This creates a preference for capabilities that can be decomposed into observable outputs.
Yet what is measurable may become privileged over what is important. Campbell (1979) warned that the more a quantitative indicator is used for social decision-making, the more it becomes subject to corruption and distorts the process it was intended to monitor. Strathern (1997) expressed a related problem: when a measure becomes a target, it ceases to function well as a measure. Engineered substitutes intensify this tendency because they are usually designed and evaluated against preselected metrics. A service bot optimized for average handle time can satisfy its target while increasing repeat contacts or eroding trust. A hiring algorithm can produce consistent rankings while reproducing omissions embedded in its training data.
RDT extends this literature by arguing that legibility influences not only how work is measured but which form of capability an organization chooses to retain. The substitute gains an institutional advantage because its value is visible in the management system; the displaced capability loses because its adaptive contribution is often counterfactual, distributed, or delayed.
Requisite variety and environmental complexity
Ashby’s (1956) law of requisite variety holds that effective regulation requires a regulator to possess sufficient variety to address the variety of the system being regulated. Organizations similarly need interpretive and response repertoires adequate to their environments. Employees contribute variety through experience, judgment, cultural knowledge, empathy, and improvisation. Customers and partners contribute feedback that reshapes offerings and routines.
An engineered substitute can contain enormous computational variety while remaining narrow in contextual variety. A language model may generate many responses but still lack authority, embodied knowledge, causal access, or the ability to recognize that a situation falls outside its valid domain. Variety is therefore not synonymous with output volume. It concerns the number of materially different states the system can correctly distinguish and the number of safe responses it can execute.
RDT predicts that substitution becomes fragile when the response variety retained by the organization falls below environmental variety. This deficit may remain hidden during normal operation because expected cases dominate the data. It becomes visible during exceptions and discontinuities.
The ironies and levels of automation
Bainbridge (1983) observed that automation can remove humans from routine control while leaving them responsible for rare, difficult situations, thereby degrading the very skills needed when intervention becomes necessary. Parasuraman et al. (2000) showed that automation varies across information acquisition, analysis, decision selection, and action implementation; its consequences depend on both type and level. These insights remain foundational in contemporary AI adoption.
RDT differs by focusing on the organization’s evaluation of the displaced capability. The theory asks why managers may interpret lower human participation as evidence of maturity even when that participation supplied anomaly detection, relationship repair, and institutional memory. It also explains the delayed balance-sheet-like accumulation created by lost capability.
Sociotechnical systems and technology-in-practice
Technology does not operate independently of people, roles, incentives, routines, and institutions. Orlikowski (1992) argued that technology both shapes and is shaped by human action. Sociotechnical research likewise treats performance as an outcome of joint optimization, not the isolated optimization of a technical subsystem (Trist & Bamforth, 1951). Digital transformation scholarship emphasizes that technology-driven change alters value creation, organizational structures, and strategic paths rather than simply installing tools (Vial, 2019).
RDT builds on these ideas but specifies a recurring substitution error. Organizations compare a designed technical artifact with an oversimplified model of the human or relational system. They then optimize the technical component while treating adaptation costs imposed on people as externalities.
Exploration, exploitation, and organizational learning
March (1991) distinguished exploitation of known capabilities from exploration of new possibilities. Excessive exploitation produces near-term refinement but can make an organization vulnerable to environmental change. Levitt and March (1988) further showed that organizations learn through routines and interpretations shaped by experience, but learning can become myopic.
Engineered substitutes excel at repeating and scaling what has been formalized. Even learning systems generally learn from selected historical data, objective functions, and feedback channels. If substitution removes conversations, anomalies, and discretionary encounters, it can narrow the organization’s future learning data. RDT therefore treats substitution as an exploration-allocation decision even when managers frame it as cost reduction.
Relational value and service co-creation
Relationship marketing and service-dominant logic reject the idea that value is produced entirely inside a firm and delivered to passive customers. Value emerges through interaction and use (Grönroos, 1994; Vargo & Lusch, 2004). Trust reduces perceived risk and enables exchange, but it is difficult to reproduce through surface behaviors alone (Mayer et al., 1995).
Research on robotic companions provides a literal foundation for the analogy. People can form meaningful responses to robotic pets, and such devices can produce benefits in some care contexts. However, studies comparing children’s relationships with Sony’s AIBO and living dogs found both similarities and important differences in how participants understood biology, mental states, and moral standing (Melson et al., 2009). Ethical analyses of robotic care similarly distinguish experienced comfort from genuine reciprocity and warn against deceptive substitution (Sharkey & Sharkey, 2012). RDT does not deny the benefit of a robotic companion. It asks which benefit is being produced, which is being imitated, and which is being lost.
Algorithmic management and human–AI complementarity
Algorithmic management distributes work, evaluates performance, and structures behavior through digital systems. It can improve coordination while producing opacity, surveillance, reduced autonomy, and contested legitimacy (Kellogg et al., 2020). Human–AI research indicates that complementarity is possible but not automatic. Human judgment and machine prediction contribute differently, and combined systems can underperform when task allocation, trust, or interface design is poor (Jarrahi, 2018; Raisch & Krakowski, 2021).
The unresolved gap is a general theory linking the attraction of substitution to its delayed organizational consequences. RDT addresses that gap.
Definition of Robotic Dog Theory
Robotic Dog Theory states that an organization will prefer an engineered substitute when it can reproduce highly visible desired outputs with greater controllability than the adaptive system it replaces; however, as substitution increases, unmeasured losses in reciprocal adaptation, signal variety, tacit knowledge, legitimacy, and recovery capacity accumulate as adaptive-value debt. When environmental demands exceed the substitute’s modeled variety and the organization’s retained adaptive capacity, performance reverses sharply.
The theory applies to an engineered substitute, defined as a technical, procedural, contractual, or metric-based arrangement intentionally designed to perform selected outputs formerly produced by a more adaptive human, social, organizational, or ecological capability. Examples include:
- automated customer service replacing open-ended human service;
- algorithmic scheduling replacing supervisor discretion;
- performance metrics replacing developmental judgment;
- standardized scripts replacing professional interpretation;
- outsourced transactional capacity replacing internal expertise;
- synthetic data replacing direct observation;
- predictive maintenance replacing operator sensing;
- digital communities replacing reciprocal stakeholder relationships.
The term “replace” includes functional displacement even when people remain employed. If the system reduces their authority to judgment-free execution, an adaptive capability has been substituted.
Core Constructs
- Engineered Substitution Intensity
Engineered substitution intensity (ESI) is the proportion of a capability’s sensing, interpretation, decision, action, and learning authority transferred from adaptive actors to a designed substitute. ESI is multidimensional. An organization may automate information collection but retain human decisions, or automate decisions while requiring human execution. A simple headcount measure would therefore be misleading.
ESI can be measured across five stages:
- sensing: who notices and captures the condition;
- interpretation: who assigns meaning;
- choice: who selects a response;
- execution: who acts;
- learning: who changes future rules.
High ESI exists when the substitute dominates all five stages and human actors lack practical authority to override or revise it.
- Control Attraction
Control attraction (CA) is the managerial preference for a substitute because it appears more predictable, scalable, inspectable, compliant, and economically legible than the original capability. Control attraction is intensified by cost pressure, audit demands, managerial distance from operations, faith in quantification, and incentives tied to near-term efficiency.
Control attraction is not irrational. Predictability and standardization are valuable. The bias arises when management applies a larger evidentiary burden to the adaptive capability than to the substitute or treats variance as waste without distinguishing harmful inconsistency from informative variation.
- Behavioral Resemblance
Behavioral resemblance (BR) is the degree to which the substitute reproduces observable outputs associated with the original capability. A chatbot that provides grammatically appropriate answers has high surface resemblance to an agent. A recommendation engine that suggests products resembles a knowledgeable seller. BR is generally the easiest construct to demonstrate in a pilot.
RDT distinguishes BR from capability equivalence (CE), the degree to which the substitute produces the full outcome—including exception recognition, contextual judgment, ethical responsibility, relationship effects, and downstream learning—under the range of conditions the organization actually faces. Substitution blindness occurs when BR is treated as proof of CE.
- Adaptive Reciprocity
Adaptive reciprocity (AR) is the two-way process through which actors alter their understanding and behavior in response to one another over time. Reciprocity includes feedback, mutual accommodation, negotiated meaning, trust calibration, and co-learning. It is not synonymous with warmth. A candid disagreement between a customer and an employee can contain more adaptive reciprocity than a pleasant but closed automated exchange.
AR produces information. The employee learns why the customer’s need does not fit the product; the customer learns what constraints are real; both may redefine the problem. When the substitute is designed to terminate the interaction efficiently, that learning may disappear.
- Signal Variety
Signal variety (SV) is the breadth and meaningful diversity of information available to detect changes, anomalies, weak signals, and emergent needs. Data volume can rise while SV falls. Ten million standardized clicks may contain less strategic variety than twenty candid conversations with departing customers.
Substitution often structures input so that it can be processed. Categories, menus, scores, and templates improve throughput but discard unanticipated meanings. SV measures what remains available to the organization after this compression.
- Retained Adaptive Capacity
Retained adaptive capacity (RAC) is the organization’s remaining human and relational ability to interpret novelty, exercise judgment, improvise safely, challenge the substitute, and recover when modeled routines fail. RAC includes expertise, role authority, staffing, practice, psychological safety, cross-functional relationships, and accessible fallback processes.
Nominal oversight is not retained capacity. A “human in the loop” who must approve hundreds of machine decisions, lacks the information to disagree, or is punished for slowing throughput does not provide meaningful RAC.
- Substitution Blindness
Substitution blindness (SB) is a governance condition in which measurement systems capture the substitute’s intended outputs and current efficiencies while omitting displaced adaptive value and downstream costs. It has three forms:
- construct blindness: measuring the wrong concept, such as equating response completion with problem resolution;
- temporal blindness: capturing immediate savings but not later learning or recovery losses;
- boundary blindness: recording benefits in the automation unit while externalizing work to customers, employees, partners, or other departments.
SB is the mechanism through which adaptive-value debt remains institutionally invisible.
- Adaptive-Value Debt
Adaptive-value debt (AVD) is the accumulated future liability created when engineered substitution reduces the organization’s ability to sense, interpret, respond to, learn from, or legitimately govern conditions outside the substitute’s design envelope. Like technical debt, AVD permits current gains while increasing future change and recovery costs. Unlike ordinary technical debt, it is stored partly in lost relationships, degraded expertise, narrowed data, reduced trust, and unpracticed human judgment.
AVD has four components:
- epistemic debt: loss of knowledge and weak signals;
- relational debt: loss of trust, reciprocity, and legitimacy;
- capability debt: atrophy of expertise, discretion, and recovery skill;
- option debt: reduction in feasible future responses because processes, contracts, data, and roles have been optimized around the substitute.
Causal Process Model
RDT proposes a six-stage cycle.
Stage 1: Capability simplification
The organization decomposes an adaptive capability into visible outputs. Customer service becomes “answering contacts”; management becomes “assigning and rating work”; recruitment becomes “ranking candidates.” This abstraction is necessary for design but becomes dangerous when excluded features are assumed irrelevant.
Stage 2: Resemblance demonstration
The engineered substitute demonstrates that it can reproduce selected outputs in a bounded setting. The pilot usually contains cleaner data, lower exception variety, expert supervision, and an explicit success definition. Behavioral resemblance is interpreted as proof of broader equivalence.
Stage 3: Control capture
The organization scales the substitute because its benefits are legible: lower unit cost, faster throughput, continuous availability, consistent execution, and auditable records. Budgets, roles, and performance targets are redesigned around the expected savings. Human capacity is removed or deskilled.
Stage 4: Adaptive attenuation
Interactions become more structured, exceptions are redirected or suppressed, and people have fewer opportunities to exercise judgment. Signal variety and adaptive reciprocity decline. Because standard cases still perform well, this stage can resemble maturity.
Stage 5: Debt accumulation
The environment changes while the organization’s ability to perceive and interpret change weakens. Work moves outside the measured boundary. Employees and customers compensate. Expertise decays. Metrics remain favorable if they exclude this hidden labor and unobserved abandonment.
Stage 6: Variety breach
A novel, ambiguous, consequential, or cascading condition exceeds the substitute’s modeled response variety. The organization discovers that its nominal human fallback lacks current knowledge, authority, staffing, or practice. Performance deteriorates nonlinearly. Recovery requires rebuilding relationships and capabilities that cannot be purchased instantly.
The cycle is not inevitable. Governance can interrupt it through bounded automation, meaningful human authority, exception telemetry, relational channels, capability rehearsal, and deliberate reinvestment of efficiency gains into exploration.
Formal Model
Let:
- (s) = engineered substitution intensity, where (0 \le s \le 1);
- (C(s)) = gains from controllability, consistency, scale, and unit efficiency;
- (E) = environmental variety;
- (Q) = task equivocality, or the existence of multiple plausible interpretations;
- (K) = consequence severity;
- (V(s)) = response variety retained in the total sociotechnical system;
- (R(s)) = adaptive reciprocity;
- (H(s)) = retained adaptive capacity;
- (B) = substitution blindness;
- (D_t) = accumulated adaptive-value debt at time (t).
Short-run performance can be represented as:
[
P_{short}(s)=P_0+C(s)-I(s),
]
where (I(s)) represents implementation and coordination cost. For many repetitive tasks, (C'(s)>0), so early substitution improves performance.
Long-run performance is:
[
P_{long}(s,t)=P_0+C(s)-I(s)-\lambda D_t-\Omega_t,
]
where (\lambda) is the rate at which debt affects realized performance and (\Omega_t) is the loss from a variety breach.
Debt accumulates as:
[
D_{t+1}=D_t+\alpha B,[E+Q+K-V(s)-H(s)]_+ +\beta [R(0)-R(s)]-M_t,
]
where ([x]_+=\max(0,x)), (M_t) represents mitigation and capability reinvestment, and (\alpha) and (\beta) scale the effects. The expression predicts little debt when total response variety and retained human capacity match environmental demand. Debt rises when external variety, ambiguity, and consequence exceed the system’s remaining capacity, especially under substitution blindness.
A variety breach occurs when:
[
E_t+Q_t > V(s)+H(s),
]
with the magnitude of loss increasing with consequence severity and coupling to downstream processes. The optimal substitution level (s^*) is therefore contextual, not universal. It is generally higher for stable, decomposable, reversible tasks and lower for novel, relational, equivocal, or irreversible tasks.
The model predicts an inverted-U relationship between substitution intensity and sustained performance in high-variety environments. The curve may remain positive or plateau in low-variety environments.
Theoretical Propositions
Proposition 1: Control-attraction proposition
The greater the perceived unpredictability, labor cost, audit difficulty, or managerial distance associated with an adaptive capability, the stronger the organization’s preference for an engineered substitute, independent of demonstrated capability equivalence.
This proposition predicts that substitution decisions are influenced by the governability of the new capability, not only its outcome performance.
Proposition 2: Resemblance-equivalence error
The greater an engineered substitute’s behavioral resemblance on visible metrics, the more likely decision makers are to infer capability equivalence across untested conditions.
The effect should be stronger when evaluators lack direct experience with the original work.
Proposition 3: Legibility premium
Engineered substitutes will receive a valuation premium when their benefits are measured within the adopting unit while adaptive losses occur across time or organizational boundaries.
This predicts greater over-adoption under siloed budgets and short executive evaluation horizons.
Proposition 4: Initial efficiency proposition
For stable, repetitive, observable, and reversible tasks, increasing engineered substitution will improve consistency, throughput, availability, and unit cost until coordination or exception costs offset the gains.
RDT is not an anti-automation theory. Failure to observe this effect would weaken its assumptions about the legitimate attraction of substitution.
Proposition 5: Reciprocal attenuation proposition
As substitution intensity increases in relationship-dependent work, adaptive reciprocity will decline unless the substitute is explicitly designed to create and route open-ended feedback into accountable human learning.
An automated channel that merely collects satisfaction scores does not necessarily preserve reciprocity.
Proposition 6: Signal-compression proposition
The more an engineered substitute constrains interactions to predefined categories, the more efficiently it will process known conditions and the less effectively the organization will detect unmodeled needs and weak signals.
This proposition predicts simultaneous operational improvement and strategic sensing loss.
Proposition 7: Skill-atrophy proposition
When human actors are removed from routine participation but remain responsible for rare exceptions, their intervention performance will decline with time since meaningful practice, unless expertise is deliberately rehearsed and refreshed.
This extends Bainbridge’s (1983) automation irony into organizational capability management.
Proposition 8: Pseudo-oversight proposition
Human oversight will fail to moderate substitution risk when overseers lack time, information, authority, competence, or psychological safety to contradict the substitute.
The proposition distinguishes symbolic “human-in-the-loop” controls from effective retained adaptive capacity.
Proposition 9: Adaptive-value debt proposition
The combination of high substitution intensity and high substitution blindness will produce adaptive-value debt even when current operating metrics improve.
This is RDT’s central intertemporal prediction.
Proposition 10: Variety-breach proposition
When environmental variety and task equivocality exceed the combined response variety of the substitute and retained adaptive capacity, performance will deteriorate nonlinearly rather than gradually.
The threshold effect differentiates RDT from simple diminishing returns.
Proposition 11: Consequence amplification proposition
The organizational harm caused by a variety breach will increase with task consequence severity, process coupling, and irreversibility.
The same classification error produces different theoretical significance in entertainment recommendations and clinical triage.
Proposition 12: Hybrid superiority proposition
In environments combining high transaction volume with high exception variety, a deliberately designed hybrid will outperform both fully human and fully substituted models over time.
The condition “deliberately designed” is essential. Simply adding people after failure does not create complementarity.
Proposition 13: Reinvestment proposition
Organizations that reinvest a portion of substitution gains in exploration, exception analysis, human capability, and relationship channels will accumulate less adaptive-value debt than organizations that remove the gains entirely from the capability system.
This proposition makes mitigation empirically testable.
Proposition 14: Trust-asymmetry proposition
When stakeholders believe they are receiving reciprocal human attention but later discover that the interaction was substantially simulated, relational legitimacy will decline more than when the engineered nature and limits of the substitute were transparent from the outset.
Transparency does not guarantee trust, but deceptive resemblance creates a distinct legitimacy risk.
Boundary Conditions
RDT does not predict an inverted-U effect everywhere. Its claims are bounded by task, environment, system, and governance conditions.
Conditions favoring extensive substitution
High substitution is more likely to create sustained value when work is:
- repetitive and decomposable;
- governed by stable rules;
- supplied with reliable and representative data;
- objectively verifiable;
- low in emotional or moral consequence;
- reversible at low cost;
- weakly coupled to consequential downstream systems;
- rich in exception detection and escalation mechanisms; and
- supported by retained capability that is practiced rather than ceremonial.
Examples may include routine data reconciliation, standardized status notifications, known fraud-pattern screening with appeal, or machine monitoring that augments rather than silences operator expertise.
Conditions favoring low substitution or hybridization
RDT predicts greater risk when work includes:
- ambiguous problem definition;
- changing goals or categories;
- rare but severe exceptions;
- emotional vulnerability or identity consequence;
- ethical discretion;
- negotiation and mutual adaptation;
- incomplete causal data;
- tightly coupled processes;
- irreversible decisions;
- adversarial adaptation; or
- responsibility that cannot legitimately be delegated.
These conditions occur in crisis leadership, complex sales, employee discipline, medical diagnosis, social services, strategic negotiation, novel product discovery, and high-stakes risk decisions.
Technology capability as a moving boundary
The boundary is not fixed by a permanent distinction between humans and machines. Improved technology can increase response variety, interpretive accuracy, and recovery capability. RDT therefore does not ground its predictions in an assumption that machines cannot learn or relate. The relevant question is empirical: does the total deployed system possess the required variety, legitimate authority, reciprocal feedback, and recovery capacity under real conditions? As technology changes, the optimal design can change.
Relationship to Existing Theories
RDT overlaps with but remains distinct from several established theories.
Theory
Central concern
What RDT adds
Bounded rationality
Decision under cognitive limits
Why legibility biases capability substitution
Requisite variety
Regulatory capacity must match system variety
How substitution reduces or relocates organizational variety over time
Ironies of automation
Automation can weaken human intervention
A multilevel debt mechanism and governance theory
Sociotechnical systems
Joint optimization of social and technical systems
A specific error: behavioral resemblance mistaken for capability equivalence
Exploration–exploitation
Balance refinement and discovery
Substitution as removal of future learning channels
Technology acceptance
Factors shaping use and adoption
Why adoption and surface performance may coexist with capability loss
Transaction-cost economics
Governance choice minimizes transaction costs
Costs omitted when reciprocity, learning, and options cross boundaries
Principal–agent theory
Misaligned interests and monitoring
Why increased monitoring may itself attenuate adaptive value
Dynamic capabilities
Sensing, seizing, and transforming
How current automation can erode the microfoundations of future sensing
Universal Resilience Theory
Resilience across interconnected systems
A specific pathway through which control optimization consumes resilience
Dynamic Value Networks Theory
Value emerges across interconnected actors
Why replacing relational nodes can compress value-bearing feedback
Universal Resilience Theory and Dynamic Value Networks Theory provide complementary foundations without subsuming RDT. Universal Resilience Theory explains the importance of adaptive capacity across layers (Pirro, 2024b). Dynamic Value Networks Theory locates value in exchanges across actors rather than isolated assets (Pirro, 2024a). RDT specifies a decision mechanism that can weaken both: an organization substitutes a controllable representation for an adaptive node, records the immediate efficiency, and fails to record the network and resilience capabilities displaced.
Operationalization and Measurement
A theory becomes useful when its constructs can be observed independently of the outcomes they are meant to explain. RDT proposes the following measurement architecture.
Measuring engineered substitution intensity
Researchers can score each task from 0 to 4 across sensing, interpretation, choice, execution, and learning:
- 0 = adaptive actor holds full authority;
- 1 = substitute advises;
- 2 = substitute defaults, actor can readily override;
- 3 = substitute acts, override is difficult or exceptional;
- 4 = substitute acts and practical override is unavailable.
The weighted sum produces ESI. Weights should reflect where judgment is consequential rather than treating every stage equally.
Measuring adaptive reciprocity
AR can be measured through stakeholder surveys, interaction analysis, and process evidence. Indicators include whether participants can introduce unstructured information, whether the organization responds to that information, whether repeated interactions change future behavior, whether reasons are explained, and whether disagreements can alter the outcome. Network measures may examine reciprocal communication rather than one-way message volume.
Measuring signal variety
SV requires more than counting data fields. Researchers can measure category diversity, novelty rate, entropy of issue types, proportion of free-form versus predefined input, weak-signal escalation, and the number of new product or process changes originating from frontline or customer interactions. Qualitative coding can identify themes excluded by the formal taxonomy.
Measuring retained adaptive capacity
RAC indicators include current expert staffing, frequency of meaningful manual practice, override use and success, time available for review, cross-training, recovery drill performance, access to contextual data, decision authority, and psychological safety. A particularly revealing measure is time-to-competent-recovery after automation failure.
Measuring substitution blindness
SB can be assessed by auditing the business case and scorecard. Researchers should determine whether measures capture repeat contacts, customer abandonment, work transferred to other units, uncompensated customer effort, exception severity, employee workarounds, skill decay, trust, and long-term learning. The larger the unmeasured boundary, the higher the blindness score.
Measuring adaptive-value debt
AVD is a latent construct and should be measured through multiple indicators:
- cost and time to handle novel exceptions;
- dependence on a shrinking number of experts;
- recovery time after system failure;
- growth of manual workarounds;
- decline in override accuracy;
- rising repeat-contact or escalation rates;
- customer effort and silent abandonment;
- employee-reported inability to exercise judgment;
- reduction in novel insights reaching decision makers;
- cost to restore a human or relational channel;
- option loss created by rigid data, contracts, or organizational design.
A validated AVD scale would require exploratory and confirmatory factor analysis, tests of discriminant validity, and longitudinal predictive validation.
Empirical Research Program
Study 1: Construct development
The first study should develop and validate scales for control attraction, substitution blindness, adaptive reciprocity, retained adaptive capacity, and adaptive-value debt. Researchers could interview executives, frontline workers, technology leaders, customers, labor representatives, and risk professionals across industries. Item generation should be followed by expert sorting, pilot surveys, exploratory factor analysis, confirmatory factor analysis, and measurement-invariance tests across sectors.
Study 2: Longitudinal field study
A multisite organization introducing an automated service or decision system offers a strong quasi-experimental setting. Researchers should collect data before deployment and at several points afterward. Outcome measures should include unit cost, cycle time, first-contact resolution, repeat contacts, exception handling, customer effort, trust, employee discretion, signal diversity, innovation inputs, and recovery performance. Difference-in-differences analysis could compare early- and late-adopting units.
RDT predicts early improvements in visible efficiency, followed by divergent long-term outcomes based on environmental variety, substitution blindness, and retained adaptive capacity.
Study 3: Scenario experiment
Managers could be randomly assigned to review business cases with identical expected performance but different levels of predictability, auditability, and relational loss visibility. The dependent variable would be willingness to substitute and the premium assigned to control. This would test control attraction and the legibility premium.
Study 4: Variety-breach simulation
Teams could manage a simulated operation under human-led, automated, and hybrid designs. The environment would begin with stable demand, followed by an unannounced novel condition. Researchers would measure normal-period efficiency, anomaly detection, response quality, recovery time, and post-event learning. RDT predicts that automation may lead under stable conditions while a well-designed hybrid produces the best full-cycle performance.
Study 5: Archival event study
Researchers could examine public service outages, recalls, algorithmic failures, or automation reversals. Proxies for ESI, RAC, task variety, and consequence severity could be coded from disclosures. The dependent variables would include recovery duration, customer harm, regulatory response, and restoration cost. Although causal inference would be limited, a large sample could test the variety-breach and consequence-amplification propositions.
Study 6: Robotic companion comparison
Because the theory’s name invites literal testing, a bounded study could compare stakeholder responses to living, robotic, and transparently hybrid companion services in care or workplace-wellness settings. The objective would not be to determine which companion is universally “real,” but to test behavioral resemblance, disclosed artificiality, reciprocity expectations, and trust asymmetry. Ethical safeguards would be essential.
Falsifiability and Disconfirming Evidence
RDT would be weakened or rejected under several findings.
First, if substitution intensity consistently improved long-term performance in high-variety, equivocal, and consequential tasks without meaningful retained human capacity, the predicted variety constraint would be false or seriously incomplete. Second, if constrained automated interactions produced equal or greater novel signal discovery than open reciprocal channels across settings, the signal-compression proposition would fail. Third, if human skill did not decay after prolonged removal from meaningful practice, the skill-atrophy mechanism would require revision. Fourth, if measurement boundary and executive time horizon did not influence substitution decisions, control attraction and substitution blindness would lack support. Fifth, if adaptive-value debt indicators failed to predict recovery cost, learning loss, trust decline, or performance reversal beyond established constructs such as technical debt and organizational slack, AVD would not demonstrate incremental validity.
The theory also risks tautology if “adaptive-value debt” is inferred only after a failure. Researchers must therefore measure debt indicators prospectively. The strongest evidence would show that pre-event measures of expertise atrophy, signal compression, workaround accumulation, and relational decline predict later variety-breach losses.
Rival Explanations
Poor implementation rather than substitution
A system may fail because of low-quality data, weak integration, inadequate testing, or poor training. RDT does not replace implementation theory. It predicts additional risk even when implementation is technically competent, particularly if the design removes adaptive channels. Studies should control for implementation quality.
Simple cost cutting
Management may knowingly accept lower service quality in exchange for savings. This is not substitution blindness if the displaced value and risks are explicitly understood. RDT becomes relevant when decision makers assume equivalence, omit cross-boundary costs, or underestimate future adaptive loss.
Technology immaturity
Failures may disappear as technology improves. RDT accommodates this possibility: better systems can increase response variety. However, technical capability does not automatically supply legitimate responsibility, reciprocal governance, or environmental access. The theory’s unit of analysis is the deployed sociotechnical system, not a fixed belief about machine limitations.
Resistance to change
Employees may resist automation because it threatens identity, status, or employment. Such resistance can reduce implementation performance. Yet dismissing all objections as resistance can itself create substitution blindness. The empirical task is to distinguish self-protective resistance from valid knowledge about task variety, exceptions, and consequences.
Transaction-cost optimization
An organization may rationally internalize or automate an activity to reduce transaction costs. RDT argues that standard transaction-cost accounting may omit learning, reciprocity, resilience, and option value. If those costs are fully included and substitution remains superior, RDT predicts no necessary reversal.
Managerial Diagnostic: The RDT Test
Before replacing an adaptive capability, leaders should answer ten questions:
- What is the real outcome? Define the stakeholder result, not the visible activity.
- What does the original capability do that is not currently measured? Include judgment, sensing, trust, teaching, recovery, and learning.
- Where is resemblance being mistaken for equivalence? Identify which conditions the pilot did not represent.
- How much environmental variety exists? Count meaningful exception classes and assess novelty, not just transaction volume.
- What information will the substitute compress or exclude? Determine whether stakeholders can introduce an unanticipated need.
- Who may override the substitute? Verify authority, time, information, competence, and protection from retaliation.
- How will human capability remain current? Define practice, rotation, simulation, and professional development.
- Where will work move? Measure effort transferred to customers, employees, suppliers, and adjacent departments.
- What signals indicate adaptive-value debt? Establish leading measures before launch.
- What is the restoration plan? Estimate the time and cost required to rebuild the displaced capability.
The test should be applied at the level of a critical outcome, not an entire technology. The same AI platform may be appropriate for summarizing routine correspondence and inappropriate for autonomously resolving a vulnerable customer’s disputed account.
Portfolio Strategy: Automate, Augment, Buffer, or Preserve
RDT yields four design choices.
Automate
Use extensive substitution when task variety is low, outcomes are verifiable, errors are reversible, and the substitute’s response variety is sufficient. Maintain monitoring and appeals proportional to consequence.
Augment
Let the substitute acquire information, generate options, or execute routine steps while adaptive actors retain meaningful interpretation and decision authority. Augmentation is preferred when volume is high but exceptions require judgment.
Buffer
Place engineered substitutes at the edge of a process to absorb predictable volume while preserving direct access to adaptive capability when signals indicate novelty, vulnerability, or high consequence. The escalation route must be easy and should not punish the stakeholder with repeated authentication or restatement.
Preserve
Retain predominantly human or relational capability when value arises principally through trust, negotiation, ethical responsibility, mutual discovery, or embodied expertise and when errors are difficult to reverse. Technology can still support documentation and information access without becoming the governing substitute.
Applications
Customer service
A chatbot can resolve known questions continuously and cheaply. RDT predicts value when intent is clear and resolution is verifiable. Risk rises when the system interprets abandonment as resolution, blocks human access, or filters novel complaints into old categories. A resilient design uses the bot as a buffer, preserves conversation history during escalation, detects repeated failure, and treats unclassified contacts as strategic signals.
Human resources
Algorithmic screening and employee analytics can process large volumes consistently. They may also convert contextual judgments into proxies, constrain appeals, and weaken developmental relationships. High-consequence decisions such as termination, accommodation, promotion, and misconduct investigation require accountable human interpretation. The RDT question is not whether an algorithm participates, but whether stakeholders retain meaningful voice and whether the organization continues learning from cases that challenge its categories.
Sales and relationship management
Automated outreach can increase contact volume while reducing reciprocal discovery. If success is measured by opens, meetings, or immediate conversion, the firm may miss how repetitive synthetic interaction affects brand trust or product learning. RDT predicts that transactional segments can tolerate higher substitution than complex, negotiated, or trust-intensive sales.
Operations and maintenance
Sensors and predictive models often outperform episodic manual inspection. However, operators possess embodied and contextual knowledge that may not be captured by formal data. Organizations should preserve channels for anomalous observations and avoid punishing departures from the model when evidence supports them. Maintenance experts must practice degraded operations so that automation failure does not reveal a hollow fallback.
Professional services
Generative AI can accelerate research, drafting, analysis, and documentation. Behavioral resemblance is high because outputs are fluent. Capability equivalence depends on source verification, problem framing, confidentiality, accountability, and judgment. RDT predicts that firms treating fluency as expertise will accumulate epistemic debt: junior professionals may produce more output while receiving fewer opportunities to develop the tacit reasoning needed to detect when the output is wrong.
Strategy and leadership
Dashboards and predictive analytics improve visibility but can turn strategy into management of predefined indicators. Leaders need exposure to disconfirming information, frontline experience, and stakeholder dialogue. Otherwise the organization may become exceptionally efficient at answering yesterday’s question.
Publishing and scholarship
Automated writing systems can generate coherent manuscripts, summarize literature, and standardize style. They can also fabricate citations, flatten authorial difference, reproduce dominant assumptions, and reward publication volume over knowledge contribution. RDT suggests a hybrid model: automation supports discovery, structure, and editing; accountable scholars verify sources, define claims, expose uncertainty, and contribute original judgment. The objective is not text that resembles scholarship but scholarship that survives scrutiny.
Ethical Implications
Engineered substitution raises ethical questions when organizations capture efficiency while transferring burden or risk to less powerful stakeholders. A customer may spend additional time navigating automation; an employee may absorb monitoring and loss of discretion; a patient may receive simulated empathy without accountable care. These effects are not incidental if they are predictable consequences of the business model.
Transparency is necessary but insufficient. Telling a person that a system is automated does not make an inaccessible appeal process fair. Ethical governance requires proportional human review, contestability, data stewardship, accessibility, and an accountable party capable of changing the decision. In high-consequence domains, responsibility cannot be dissolved across developers, vendors, operators, and models.
The robotic-dog analogy also cautions against contempt for people who find value in substitutes. A robotic companion may reduce loneliness; an automated service may increase access; a script may protect employees from abusive discretion. RDT does not rank “natural” above “artificial.” It requires precision about the value actually created and honesty about the capability actually present.
Strategic Implications
RDT reframes automation strategy from labor substitution to capability architecture. The strategic unit is not the job but the portfolio of sensing, interpretation, decision, action, learning, relationship, and recovery capabilities required to produce an outcome. Leaders should ask where machines provide superior consistency and scale, where people provide contextual variety and legitimacy, and how the two will correct one another.
This approach changes the automation business case. Savings should not be recognized without reserves for monitoring, exception handling, capability maintenance, appeal, and restoration. Efficiency gains should partly fund exploration and learning. A firm that removes every dollar of human capacity after automation may report a stronger first-year return while increasing its exposure to adaptive-value debt.
RDT also predicts a strategic differentiation opportunity. As competitors automate the same visible interactions, genuine reciprocal access may become scarce and valuable. A firm can automate routine work while deliberately making accountable human judgment easier to reach in consequential moments. The advantage is not nostalgia; it is superior allocation of adaptive capacity.
Discussion
Robotic Dog Theory begins with a deceptively simple observation: an engineered artifact can imitate the outputs that an evaluator chooses to count. The theory becomes consequential when the omitted features are the organization’s means of adaptation. Control attraction encourages leaders to value predictability, consistency, and scale. Behavioral resemblance makes the substitute appear equivalent. Substitution blindness hides displaced effort and delayed losses. Adaptive-value debt then accumulates until environmental variety breaches the combined capability of the substitute and the diminished human system.
The theory resolves an apparent contradiction in debates about automation. Advocates can correctly demonstrate productivity, access, consistency, and safety gains. Critics can correctly identify deskilling, opacity, trust loss, and brittleness. These outcomes can coexist because they occur at different levels, boundaries, and times. RDT predicts the conditions governing their balance.
The theory also rejects a crude human-versus-machine opposition. Humans are inconsistent, biased, fatigued, political, and sometimes unsafe. Engineered substitutes can expand human possibility. The critical error is replacement based on an impoverished model of the original capability. A living system’s variation may contain both waste and information. Good design removes harmful variance while preserving informative variance and the ability to respond to what was not anticipated.
Limitations
RDT is conceptual and has not yet received empirical validation. Its constructs may overlap with organizational resilience, dynamic capabilities, technological frames, sociotechnical fit, and organizational slack. Empirical work must demonstrate discriminant and incremental validity. Adaptive-value debt is especially vulnerable to retrospective storytelling; prospective measurement is essential.
The robotic-dog analogy may also provoke an unintended interpretation that human workers are analogous to pets. They are not. The analogy concerns the evaluator’s substitution error: reproducing selected visible behaviors does not establish equivalence of the underlying capability or relationship. Researchers and managers should use the term without demeaning labor or romanticizing biological systems.
Finally, technological progress can change the theory’s empirical boundaries. Systems may become more capable of contextual learning and reciprocal interaction. RDT remains applicable only if constructs are measured at the level of the total deployed system and revised when evidence contradicts its assumptions.
Conclusion
Organizations are entering an era in which almost any visible behavior can be simulated, scored, standardized, or automated. The central managerial danger is no longer simply adopting weak technology. It is defining the original capability so narrowly that imitation is mistaken for equivalence.
Robotic Dog Theory proposes that engineered substitution produces an early control dividend and can create lasting value in stable, observable, reversible work. Yet in environments characterized by novelty, ambiguity, emotional or ethical consequence, tight coupling, and irreversible outcomes, aggressive substitution can compress signals, weaken reciprocity, atrophy skill, and reduce recovery capacity. When scorecards omit those losses, adaptive-value debt accumulates. The debt becomes visible when the environment asks a question the substitute was not designed to answer and the organization discovers that the displaced capability is no longer available.
The theory’s practical conclusion is neither “automate” nor “do not automate.” It is: never confuse a system that behaves like the capability with a system that preserves the value of the capability. Firms should automate where rules are stable, augment where judgment and scale must coexist, buffer routine demand while maintaining accessible expertise, and preserve reciprocal human responsibility where the outcome depends on trust, discovery, or moral accountability. The organizations that master this distinction will gain the efficiency of engineered systems without surrendering the adaptive intelligence required to survive what they cannot predict.
References
Ashby, W. R. (1956). An introduction to cybernetics. Chapman & Hall.
Bainbridge, L. (1983). Ironies of automation. Automatica, 19(6), 775–779. https://doi.org/10.1016/0005-1098(83)90046-8
Campbell, D. T. (1979). Assessing the impact of planned social change. Evaluation and Program Planning, 2(1), 67–90. https://doi.org/10.1016/0149-7189(79)90048-X
Grönroos, C. (1994). From marketing mix to relationship marketing: Towards a paradigm shift in marketing. Management Decision, 32(2), 4–20. https://doi.org/10.1108/00251749410054774
Jarrahi, M. H. (2018). Artificial intelligence and the future of work: Human–AI symbiosis in organizational decision making. Business Horizons, 61(4), 577–586. https://doi.org/10.1016/j.bushor.2018.03.007
Kellogg, K. C., Valentine, M. A., & Christin, A. (2020). Algorithms at work: The new contested terrain of control. Academy of Management Annals, 14(1), 366–410. https://doi.org/10.5465/annals.2018.0174
Levitt, B., & March, J. G. (1988). Organizational learning. Annual Review of Sociology, 14, 319–338. https://doi.org/10.1146/annurev.so.14.080188.001535
March, J. G. (1991). Exploration and exploitation in organizational learning. Organization Science, 2(1), 71–87. https://doi.org/10.1287/orsc.2.1.71
Mayer, R. C., Davis, J. H., & Schoorman, F. D. (1995). An integrative model of organizational trust. Academy of Management Review, 20(3), 709–734. https://doi.org/10.2307/258792
Melson, G. F., Kahn, P. H., Jr., Beck, A. M., Friedman, B., Roberts, T., Garrett, E., & Gill, B. T. (2009). Children’s behavior toward and understanding of robotic and living dogs. Journal of Applied Developmental Psychology, 30(2), 92–102. https://doi.org/10.1016/j.appdev.2008.10.011
Orlikowski, W. J. (1992). The duality of technology: Rethinking the concept of technology in organizations. Organization Science, 3(3), 398–427. https://doi.org/10.1287/orsc.3.3.398
Parasuraman, R., Sheridan, T. B., & Wickens, C. D. (2000). A model for types and levels of human interaction with automation. IEEE Transactions on Systems, Man, and Cybernetics—Part A: Systems and Humans, 30(3), 286–297. https://doi.org/10.1109/3468.844354
Pirro, N. J. (2024a). Dynamic Value Networks Theory: A new framework for interconnected systems. Pyrrhic Press Publishing. https://doi.org/10.5281/zenodo.11120455
Pirro, N. J. (2024b). Universal Resilience Theory: A comprehensive framework for understanding and enhancing resilience in diverse systems. Pyrrhic Press Publishing. https://doi.org/10.5281/zenodo.10997597
Raisch, S., & Krakowski, S. (2021). Artificial intelligence and management: The automation–augmentation paradox. Academy of Management Review, 46(1), 192–210. https://doi.org/10.5465/amr.2018.0072
Sharkey, A., & Sharkey, N. (2012). Granny and the robots: Ethical issues in robot care for the elderly. Ethics and Information Technology, 14, 27–40. https://doi.org/10.1007/s10676-010-9234-6
Simon, H. A. (1955). A behavioral model of rational choice. The Quarterly Journal of Economics, 69(1), 99–118. https://doi.org/10.2307/1884852
Strathern, M. (1997). “Improving ratings”: Audit in the British university system. European Review, 5(3), 305–321. https://doi.org/10.1002/(SICI)1234-981X(199707)5:3%3C305::AID-EURO184%3E3.0.CO;2-4
Trist, E. L., & Bamforth, K. W. (1951). Some social and psychological consequences of the longwall method of coal-getting. Human Relations, 4(1), 3–38. https://doi.org/10.1177/001872675100400101
Vargo, S. L., & Lusch, R. F. (2004). Evolving to a new dominant logic for marketing. Journal of Marketing, 68(1), 1–17. https://doi.org/10.1509/jmkg.68.1.1.24036
Vial, G. (2019). Understanding digital transformation: A review and a research agenda. The Journal of Strategic Information Systems, 28(2), 118–144. https://doi.org/10.1016/j.jsis.2019.01.003