The AI Configuration Paradox

Contents

Abstract
1. Introduction: The Hidden Labor of Delegated Improvement
2. Canonical Definition and Discriminant Boundary
3. Research Gap and Adjacent Traditions
4. Human Meta-Work: Taxonomy, Attribution, and Measurement
5. Agentic AI as an Integrative Technical Substrate
6. Classical Human-Sciences Mechanisms
7. Integrated Causal Architecture
8. Formal Conceptual Framework
9. Baselines, Materiality, Distribution, and Exit
10. Empirical Hypotheses and Research Program
11. Governance Derivation and the Configuration Continuation Gate
12. Strongest Objections and Replies
13. Falsification, Boundary Conditions, and Maturity
14. Conclusion: When the Improvement Mechanism Becomes the Work
Glossary
References

Abstract

Agentic artificial intelligence is commonly evaluated through the human execution it appears to replace. This evaluation can remain incomplete because automation may relocate human effort into specification, configuration, orchestration, supervision, verification, governance, correction, and configuration-debt repair. This article develops the AI Configuration Paradox as a candidate socio-technical theory of the conditions under which that higher-order human work ceases to remain subordinate to the objective that justified the AI system.

The paradox is defined conjunctively. It exists only where: P1, an identifiable evaluative objective is present; P2, the AI configuration regime generates substantial incremental human meta-work; P3, that burden materially displaces the objective or expected benefit; and P4, the regime persists without a bounded learning rationale or an effective path to simplification, restriction, suspension, or rollback. The framework therefore distinguishes the paradox from high implementation cost, legitimate experimentation, necessary safety oversight, technical debt, organizational inertia, and complex but beneficial automation.

Human meta-work is represented as the multidimensional profile:

M(T)=(C,O,S,V,G,Dr)

comprising configuration, orchestration, supervision, verification, governance, and realized configuration-debt cost. External benefit and human burden remain separate profiles unless a transparent domain-specific valuation supports scalar comparison. The analysis further requires credible baselines, predeclared materiality rules, actor-level distribution, and explicit exit and reversibility conditions.

The article integrates established mechanisms—including formal rationalization, goal displacement, bounded rationality, indicator corruption, escalation of commitment, exploration–exploitation imbalance, automation irony, and out-of-the-loop performance—with technical properties of agentic systems such as recursive configurability, cheap generation, persistent state, internal observability, distributed ownership, and exit dependence. These mechanisms are treated as candidate causal pathways rather than definitional conditions.

The framework derives falsifiable hypotheses and a candidate Configuration Continuation Gate with four authorized states: SHIP, RESTRICT, HOLD, and ROLLBACK. Governance is defined as operative only where evidence is connected to a threshold, authorized decision, and enforceable consequence.

The theory is conceptually advanced but empirically unvalidated. Its independent status depends on whether P1–P4 can be measured reliably and explain persistent configuration expansion and failure to simplify beyond total cost of ownership, workload, technical debt, sunk cost, and general organizational dysfunction.

1. Introduction: The Hidden Labor of Delegated Improvement

Agentic artificial intelligence is commonly evaluated through what it can perform without direct human execution.

A system may decompose an objective, generate plans, select tools, delegate subtasks, retain state, review intermediate outputs, revise its approach, and execute actions across several technical environments.

Measured at the level of task execution, the apparent direction of improvement is straightforward:

More Machine Execution → Less Human Work

That inference is incomplete.

Human labor may not disappear when execution is automated. It may move into a higher-order control layer. A person who previously performed a task directly may become responsible for defining the objective, translating it into system instructions, configuring models and roles, coordinating agents and tools, supervising ongoing behavior, verifying generated outputs, correcting failures, governing permissions, maintaining rollback capacity, and repairing configuration debt.

The visible task may become faster while the complete configuration regime becomes more demanding.

Consider a research unit that replaces a largely human workflow with a multi-agent system. One agent plans the inquiry, several agents collect material, another evaluates sources, another drafts the report, and a final agent critiques the result. Direct drafting time falls sharply. Yet the unit must now determine which sources the agents may use, how state is transferred, how conflicting findings are resolved, which claims require human verification, whether the critic is genuinely independent, how persistent memory is governed, who may approve external actions, and what happens when the resulting report is plausible but wrong.

The unit may also need to maintain dashboards, evaluation criteria, permission structures, audit records, and procedures for recovering from failure. The system has automated execution while generating a new form of human work.

That transformation is not necessarily a failure. The resulting burden may be justified by greater quality, faster completion, lower risk, expanded coverage, improved accessibility, or a genuinely valuable learning objective.

The analytical problem begins when the burden required to operate and govern the improvement mechanism materially displaces the objective it was introduced to serve. The problem becomes a paradoxical configuration when the regime continues even though the relevant evidence should support simplification, restriction, suspension, or rollback.

This article defines that condition as the AI Configuration Paradox.

The AI Configuration Paradox is a condition in which a socio-technical configuration regime organized around an AI system and introduced to improve an identifiable objective—one not satisfied merely by the regime’s own existence, continuation, or expansion—generates substantial incremental human meta-work relative to a credible baseline; that burden materially displaces the objective or expected benefit under a declared materiality rule; and the regime persists without a bounded learning rationale or an effective path to simplification, restriction, suspension, or rollback.

The term paradox does not denote a formal logical contradiction. It denotes a counterproductive reversal:

Mechanism Introduced to Reduce Human Burden → Higher-Order Human Burden → Displacement of the Intended Benefit

The theory is therefore not a general argument against AI, autonomy, configuration, governance, or technical complexity. A complex agentic system may create substantial net value. Extensive verification may be necessary in a high-stakes domain. A long configuration period may be justified by bounded experimentation. Governance may protect rights, safety, and accountability even where it increases cost.

The paradox requires a narrower conjunctive structure:

AICPd,a ∣ b,T ⇔ P1d,T ∧ P2a ∣ b,T ∧ P3d,a ∣ b,T ∧ P4a,T

where:

P1 is an identifiable evaluative objective;
P2 is substantial incremental human meta-work attributable to the AI configuration regime;
P3 is material displacement;
P4 is persistent continuation failure.

No single condition is sufficient. High workload alone is not the paradox. Material displacement followed by rapid correction is not the full paradox. Persistent use of an unpopular system is not the paradox unless configuration-regime-attributable meta-work materially displaces the evaluative objective.

The conjunctive definition is intended to prevent the concept from becoming another name for hidden labor, total cost of ownership, technical debt, poor implementation, or ordinary organizational inertia.

The article’s proposed contribution is socio-technical. The relevant mechanisms are not created entirely by AI. Research on bureaucracy, bounded rationality, social indicators, commitment, organizational learning, automation irony, and out-of-the-loop performance already explains why human organizations can confuse means with ends, continue search without a stable adequacy criterion, reorganize around decision-bearing metrics, escalate commitment after weak results, remain trapped in experimentation, and distribute knowledge, burden, risk, and authority across actors.

Agentic AI supplies a distinctive technical substrate through which these mechanisms may become unusually extensible, recursive, observable, and interconnected.

The central interaction can be stated as:

Classical Human and Institutional Mechanisms × Agentic Technical Properties → Candidate Persistent Configuration Regime

The multiplication sign is conceptual rather than numerical. It indicates interaction. Agentic AI does not independently cause the paradox. Its technical properties may increase the probability, scale, speed, or coupling through which existing human and institutional mechanisms operate.

The relevant unit of analysis is therefore not the model alone. It is the complete configuration regime:

Evaluative Objective + AI Architecture + Human Meta-Work + External Outcome + Baseline + Burden Distribution + Correction Structure

This shift in unit of analysis changes the governing question. The conventional question is: How much of the task can the AI system perform? The question developed here is: Does the complete configuration regime improve the evaluative objective once all incremental human specification, orchestration, supervision, verification, governance, correction, and debt-repair work is included?

A second question follows: If the answer becomes negative, uncertain, or materially weaker than a credible simpler alternative, can the relevant evidence change the system’s state?

The article’s originality claim does not depend on asserting first use of the phrase configuration paradox. Nor does it claim discovery of hidden cost, automation irony, goal displacement, proxy corruption, escalation of commitment, or exploration–exploitation tension.

Its proposed contribution consists of five integrated moves:

defining the AI Configuration Paradox through the P1–P4 conjunction;
modeling human meta-work through six functional components;
separating multidimensional benefit and burden profiles from optional scalar valuation;
integrating technical, cognitive, organizational, and institutional mechanisms into a plural causal architecture;
deriving a candidate governance structure capable of producing SHIP, RESTRICT, HOLD, or ROLLBACK decisions.

The article proceeds accordingly. Section 2 establishes the canonical definition and discriminant boundary. Section 3 positions the framework relative to adjacent research traditions. Section 4 develops the human meta-work taxonomy. Section 5 explains why agentic AI functions as an integrative technical substrate. Section 6 reconstructs the relevant human-sciences and Human Factors mechanisms. Section 7 develops the causal architecture. Sections 8 and 9 formalize benefit, burden, baselines, materiality, distribution, debt, and exit. Section 10 presents the empirical hypotheses and research program. Section 11 derives the Configuration Continuation Gate. Section 12 confronts the strongest objections. Section 13 defines falsification conditions, scope boundaries, and the theory’s current maturity.

The article does not claim that the paradox is prevalent, that its mechanisms have been empirically confirmed, or that the proposed governance architecture has been validated. Its present claim is more limited:

The AI Configuration Paradox can be formulated as an advanced candidate socio-technical theory whose constructs, mechanisms, measurements, rejection conditions, and governance implications are sufficiently explicit to support disciplined empirical testing.

The governing principle of the theory is therefore not that automation should always reduce human involvement. It is that automation of execution does not by itself establish improvement. The complete human burden required to make the system useful, acceptable, correctable, and governable must remain instrumentally subordinate to the evaluative objective that justified the system in the first place.

2. Canonical Definition and Discriminant Boundary

The classification is indexed to a domain or objective d, configuration alternative a, credible baseline b, and evaluation period T. A case is therefore not classified in the abstract; it is classified relative to a specified comparison and time horizon.

2.1 Canonical Definition

The AI Configuration Paradox is a condition in which a socio-technical configuration regime organized around an AI system and introduced to improve an identifiable objective—one not satisfied merely by the regime’s own existence, continuation, or expansion—generates substantial incremental human meta-work relative to a credible baseline; that burden materially displaces the objective or expected benefit under a declared materiality rule; and the regime persists without a bounded learning rationale or an effective path to simplification, restriction, suspension, or rollback.

The definition contains four constitutive conditions:

AICPd,a ∣ b,T ⇔ P1d,T ∧ P2a ∣ b,T ∧ P3d,a ∣ b,T ∧ P4a,T

P1 — Identifiable evaluative objective
P2 — Substantial incremental human meta-work attributable to the AI configuration regime
P3 — Material displacement
P4 — Persistent continuation failure

These conditions are conjunctive. A case that lacks any one of them should not receive the full classification.

2.2 P1 — Identifiable Evaluative Objective

The configuration regime must be introduced in relation to an objective that is not satisfied merely by the regime’s own existence, continuation, or expansion. “External” in this article is evaluative rather than spatial: the objective may be internal to an organization or may concern research, learning, resilience, or configuration craftsmanship, provided success can be assessed independently of simply continuing to configure.

A valid objective should identify, as far as the domain permits:

the condition to be improved;
the principal beneficiary or affected population;
the relevant baseline;
the evaluation period;
and the evidence through which improvement can be assessed.

Examples include reducing decision latency, improving diagnostic accuracy, producing a decision-ready research report, lowering error rates, expanding access, or reducing direct execution burden.

“Use AI,” “increase AI adoption,” “build an agentic architecture,” or “improve the configuration” are not, by themselves, evaluative objectives. They describe a technology or mechanism rather than the human, organizational, or domain condition the mechanism is intended to improve.

An objective need not be purely economic or quantitative. It may concern quality, risk, accessibility, autonomy, learning, resilience, or subjective utility. It must nevertheless be identifiable enough to support comparison.

Objective revision

Objectives may change legitimately. Revision is acceptable where it is explicit, temporally recorded, justified, and evaluated as a new decision.

Objective revision must not erase the historical record. If a system introduced to save time later becomes valuable as a research or learning environment, the later objective may be legitimate. It does not retroactively establish that the original time-saving objective was achieved.

2.3 P2 — Substantial Incremental Human Meta-Work Attributable to the Configuration Regime

The socio-technical configuration regime organized around the AI system must generate human work beyond what would have been required under the selected baseline.

Human meta-work includes:

configuration and specification;
orchestration;
supervision;
verification;
governance;
and realized configuration-debt repair.

P2 does not require the burden to be unjustified. A high-value or high-stakes system may rationally require substantial meta-work.

The required finding is attribution:

Δ Ma ∣ b(T)=Ma(T)-Mb(T)

The analysis must distinguish incremental burden attributable to the configuration regime from pre-existing coordination, review, compliance, or management work. Attribution may be causal, quasi-experimental, documentary, or carefully reconstructed; where it cannot be established, the weaker phrase AI-associated burden should be used.

2.4 P3 — Material Displacement

The incremental burden must materially displace the evaluative objective, expected benefit, or another resource necessary to achieve it.

Material displacement may occur where:

net value becomes negative against a credible baseline;
scarce expert capacity is consumed beyond a declared threshold;
verification or governance delay destroys the usefulness of the output;
another activity of equal or greater value is displaced;
burden is transferred to an actor beyond an accepted limit;
or the system fails against a credible simpler alternative.

P3 requires a predeclared or independently defensible materiality rule. Visible complexity, frustration, or high workload alone is insufficient.

2.5 P4 — Persistent Continuation Failure

The configuration regime must continue after relevant displacement evidence emerges, without either:

a bounded learning rationale; or
an effective path to simplification, restriction, suspension, or rollback.

A bounded learning rationale requires:

a declared learning objective;
testable hypotheses;
a resource or time budget;
specified evidence;
stopping criteria;
and periodic reauthorization.

P4 does not require that the system remain completely unchanged. Continuous modification can itself be the form of continuation.

The operative question is whether evidence can change the authorized state of the system through:

Signal → Threshold → Authority → Consequence

A review process that can document problems but cannot alter scope, permissions, architecture, or continuation is not an effective stopping path.

2.6 Lifecycle Phases

The theory distinguishes four phases:

Setup — initial specification, integration, training, and architecture construction.
Stabilization — bounded correction and learning after first use.
Routine maintenance — ordinary operation, updating, and proportionate control.
Persistent reconfiguration — continuing modification without convergence, bounded learning, or terminal acceptance.

High burden during setup or stabilization does not by itself establish P4. The relevant evidence is whether the regime transitions toward stable external value production or remains in a continuing state of configuration.

2.7 False Positives and Exclusion Rules

Observed condition	Why it is not sufficient
High implementation cost	P3 and P4 may be absent
Extensive safety review	May be proportionate to risk reduction
Long research program	May have a bounded learning rationale
Complex architecture	Complexity may create greater external value
Technical debt	May be temporary, visible, and effectively governed
Organizational inertia	P2 or P3 may be absent
User frustration	Does not establish material displacement
High AI usage	Internal activity is not external benefit
Configuration as creative practice	Configuration may be the declared objective

The exclusion rule is:

¬ P1d,T ∨ ¬ P2a ∣ b,T ∨ ¬ P3d,a ∣ b,T ∨ ¬ P4a,T ⇒ AICPd,a ∣ b,T=0

Partial findings should be reported precisely. A case may exhibit configuration-regime-attributable meta-work, material displacement, or weak correction without satisfying the complete classification.

3. Research Gap and Adjacent Traditions

The AI Configuration Paradox does not emerge in an empty intellectual field. Research on technological productivity, lifecycle cost, cooperative work, hidden integration labor, machine-learning technical debt, automation, human oversight, organizational dysfunction, social indicators, commitment, and organizational learning already explains many processes that can appear inside an AI configuration regime.

The originality claim must therefore be made against the nearest explanatory neighbors, not only against broad classical traditions. The article does not claim that previous scholarship failed to notice hidden work, labor transfer, repair, technical debt, or weak productivity. It asks whether these elements can form a distinguishable conjunctive configuration:

AICPd,a ∣ b,T ⇔ P1d,T ∧ P2a ∣ b,T ∧ P3d,a ∣ b,T ∧ P4a,T

The residual research question is whether a framework that joins attributable meta-work, material displacement, baseline sensitivity, actor-level distribution, exit conditions, and operative stopping authority explains something that the adjacent traditions do not explain separately.

3.1 Information-Technology Productivity Paradox

Research on the information-technology productivity paradox established that investment in computing capability does not automatically appear as proportional productivity improvement. Brynjolfsson organized explanations around measurement error, delayed organizational adjustment, redistribution, and mismanagement rather than treating the absence of aggregate gain as a simple technological verdict (Brynjolfsson, 1993).

This tradition establishes an essential negative inference:

Technological Investment ⇏ Realized External Productivity

It does not ordinarily provide a functional taxonomy of the human work generated by an AI configuration, actor-level incidence, a materiality rule, or an institutional account of why weak-value architecture continues after relevant evidence emerges. The present framework shifts the question from Where is the productivity gain? to Which work was generated, did it materially displace the objective, and could the evidence alter continuation?

3.2 Total Cost of Ownership and Lifecycle Cost

Total cost of ownership expands analysis beyond acquisition price to the costs of obtaining, operating, supporting, maintaining, and replacing an asset or supplier relationship (Ellram, 1995). It directly supports the article’s rejection of model price, task throughput, or visible execution savings as complete measures of AI value.

The theory nevertheless differs from TCO in three respects. First, the meta-work profile is functional before it is monetary; it preserves scarce expertise, cognitive interruption, downstream correction, and authority separation that may be hidden by accounting categories. Second, the theory includes mechanisms explaining why complete-cost evidence may fail to produce correction. Third, P4 requires an observable continuation failure rather than merely a high lifecycle cost.

If rigorous TCO predicts persistent continuation and failure to simplify as well as P1–P4, the independent theory should be reduced.

3.3 Articulation Work and Cooperative-Work Infrastructure

The CSCW tradition treats cooperative work as requiring articulation work: the work necessary to align distributed tasks, actors, resources, schedules, and contingencies so that substantive work can proceed. Schmidt and Bannon made support for articulation work central to the conceptual identity of CSCW (Schmidt & Bannon, 1992).

This tradition is a direct antecedent of the article’s configuration and orchestration components. It shows that coordination is not an accidental residue but an intrinsic requirement of interdependent work arrangements.

The difference is not that human meta-work discovers articulation. The proposed extension is narrower:

it separates six functional burden components;
attributes their increment to a configuration alternative relative to a baseline;
requires a declared materiality test;
examines distribution among beneficiaries, operators, verifiers, and authorities;
and adds persistent continuation failure as a separate constitutive condition.

Articulation work may be productive, necessary, and well governed. The paradox requires more than its presence.

3.4 Patchwork and Hidden AI-Integration Labor

Fox and colleagues document patchwork: situated, often invisible human labor required to integrate AI into essential work, bridge organizational gaps, repair failures, and make technically incomplete systems function in practice (Fox et al., 2023). Their analysis is among the closest empirical neighbors to human meta-work because it treats integration labor as constitutive of real deployment rather than as noise outside the system boundary.

Larsen-Ledet similarly distinguishes productive articulation—interaction that helps refine substantive objectives—from mundane articulation devoted to repeatedly supplying context and clarification so that generative AI can produce usable results (Larsen-Ledet, 2026).

These works directly challenge narratives in which AI output is treated as self-sufficient. They also sharpen the present theory’s obligation not to classify all configuration as waste.

The residual contribution of P1–P4 is not identification of invisible integration work. It is the distinction among:

integration work that is necessary and proportionate;
integration work that becomes materially displacing;
displacement that is corrected;
and displacement that persists because evidence lacks an operative route to consequence.

3.5 Pseudo-Automation and Labor Offsetting

Moradi, Levy, and Cheyre define pseudo-automation through technologies that appear to automate work while offsetting labor onto customers and reconfiguring the work that remains for frontline employees. Their study of self-checkout finds parallel demands, problem-oriented monitoring and policing, and relational repair rather than simple labor elimination (Moradi et al., 2025).

This is directly relevant to configuration-burden substitution. Both frameworks reject measurement that credits the beneficiary with automation while excluding displaced labor elsewhere.

The present theory adds a different analytical object. Pseudo-automation can occur without an agentic configuration regime, without configuration debt, and without P4. Conversely, the AI Configuration Paradox can arise through verification, governance, or exit dependence even where labor is not transferred to customers. The overlap lies in labor offsetting; the residual contribution lies in the full configuration-regime classification and stopping architecture.

3.6 Hidden Technical Debt in Machine-Learning Systems

Sculley and colleagues show that machine-learning systems accumulate system-level technical debt through entanglement, hidden feedback loops, undeclared consumers, data dependencies, configuration issues, and changes in the external world (Sculley et al., 2015).

This literature is the nearest technical antecedent of configuration debt. The article adopts its central warning that the model is only a small part of a production system and that apparently local changes can create broad maintenance obligations.

The present distinction is a stock-flow and labor-incidence distinction:

Ds represents accumulated configuration-debt stock;
Dr represents realized human cost during the evaluation period;
only realized cost enters the current meta-work vector;
and debt is analyzed together with objective, benefit, distribution, exit, and stopping authority.

Technical debt may be substantial yet visible, justified, and effectively retired. It becomes part of the paradox only through P2–P4.

3.7 Automation Irony and Out-of-the-Loop Performance

Human Factors research provides the strongest antecedent for the proposition that automation can transfer rather than eliminate human responsibility. Bainbridge showed that automation may handle normal operation while leaving people responsible for abnormal conditions for which reduced participation makes them poorly prepared (Bainbridge, 1983). Endsley and Kiris linked reduced active control to impaired situation awareness and takeover performance (Endsley & Kiris, 1995). Parasuraman and Riley distinguished use, misuse, disuse, and abuse rather than treating automation reliance as a single variable (Parasuraman & Riley, 1997).

Recent work extends this concern to generative and agentic systems. Simkute and colleagues analyze production-to-evaluation shifts, workflow restructuring, and productivity loss in human–AI interaction (Simkute et al., 2024). Dhanorkar, Passi, and Vorvoreanu identify a priori control, co-planning, real-time monitoring, and post hoc review in developers’ practical oversight of software agents (Dhanorkar et al., 2026).

The present theory expands the unit from operator–automation interaction to the complete configuration regime, adding configuration, orchestration, governance, debt, baseline comparison, actor distribution, and institutional termination.

3.8 Verification Bottlenecks

Agentic systems can increase the rate and volume of candidate production more rapidly than responsible acceptance can scale. Emerging research frames human specification, judgment, verification, or novelty as potential constraints on AI-assisted work. Liang’s novelty-bottleneck model is explicitly stylized rather than a universal law, but it clarifies why improved generation does not mechanically remove human effort (Liang, 2026).

Verification-bottleneck accounts support H3 and H4. They do not by themselves establish that verification is materially displacing, that a simpler baseline is superior, or that the regime lacks an effective stopping path. They may establish P2 and contribute to P3; P4 remains analytically separate.

3.9 Goal Displacement, Indicators, Commitment, and Search

Merton’s account of bureaucratic goal displacement explains how rules and procedures created as means can acquire independent practical value (Merton, 1940). Campbell explains how consequential indicators can become targets of adaptation and lose validity as representations of the underlying condition (Campbell, 1979). Staw shows why negative feedback can increase commitment under conditions of responsibility for the original choice (Staw, 1976). March distinguishes exploration from exploitation and the different temporal structures of their returns (March, 1991). Simon’s bounded-rationality framework explains why decision procedures require aspiration levels capable of terminating search (Simon, 1955).

These traditions supply mechanisms, not the full classification. The article re-specifies them under agentic conditions and tests whether they mediate burden formation, displacement, or persistence.

3.10 Contemporary Agentic-AI Definitions and Governance

Terminology in the agentic-AI field remains unsettled. The OECD distinguishes AI agents from agentic AI while emphasizing autonomy, goal pursuit, tool use, multi-agent coordination, task decomposition, longer operation, and socio-technical embedding (OECD, 2026). NIST’s 2026 AI Agent Standards Initiative similarly centers agents capable of autonomous action, secure operation on behalf of users, interoperability, identity, and authorization (NIST, 2026a).

This article adopts a broader functional continuum rather than making multi-agent composition a definitional requirement. It does so because the causal properties at issue—planning depth, action authority, tool use, persistence, recursive configuration, and reversibility—can vary within single-agent and multi-agent systems. The difference from the OECD terminology is explicit rather than accidental.

Monitoring research also reinforces the need to separate deployment from validated governance. NIST’s review of post-deployment monitoring reports that methods, terminology, and best practices remain nascent and identifies field studies and incident monitoring as continuing research needs (Rao et al., 2026). The Configuration Continuation Gate is therefore presented as a candidate architecture, not a validated standard.

3.11 Residual Contribution Matrix

Adjacent tradition	What it already explains	Residual question addressed by P1–P4
Productivity paradox	Capability and investment may not yield measured productivity	Which attributable human burden displaced the objective, and why did the regime persist?
Total cost of ownership	Acquisition price omits lifecycle cost	Does complete-cost evidence connect to materiality and operative correction?
Articulation work	Cooperative systems require coordination and alignment	When does necessary articulation become materially displacing?
Patchwork	AI integration depends on hidden situated repair and bridging	Was the burden proportionate, distributed visibly, and correctable?
Pseudo-automation	Reported automation can offset labor onto others	Does labor offset form part of a persistent configuration regime with weak stopping authority?
ML technical debt	ML systems accumulate hidden dependencies and maintenance obligations	How much debt cost is realized by whom, against which baseline, and with what exit path?
Automation irony	Automation can transfer responsibility and weaken intervention capacity	Does transferred oversight materially displace the objective and persist?
Verification bottlenecks	Cheap production may remain constrained by human acceptance	Why is the architecture not narrowed when acceptance becomes the dominant cost?
Goal displacement and proxy capture	Means and metrics can become operative ends	How do those mechanisms connect to P3 and P4 under agentic conditions?
Escalation and exploration	Organizations may continue investment and search	Is continuation bounded learning, or a terminal correction failure?

3.12 Research-Gap Formulation

The unresolved research problem can be stated as follows:

Existing traditions identify hidden coordination, repair, labor offset, lifecycle cost, technical debt, oversight transfer, indicator corruption, commitment, and search imbalance. They do not yet provide a unified, conjunctively defined, and operationalized framework for determining when an AI configuration regime generates substantial incremental human meta-work, materially displaces a declared objective relative to a credible baseline, and persists because burden evidence lacks an effective route to simplification, restriction, suspension, or rollback.

The proposed framework fills that gap through five moves:

Unit-of-analysis integration — the complete socio-technical configuration regime rather than the model alone.
Discriminant definition — P1–P4 separates the paradox from high cost, legitimate learning, necessary oversight, and corrected failure.
Measurement integration — benefit, meta-work, baselines, materiality, actor incidence, debt, and exit.
Causal integration — antecedents, mediators, moderators, and feedback loops remain distinct from constitutive conditions.
Governance integration — evidence must connect to signal, threshold, authority, gate, and consequence.

3.13 Reduction Test

No gap statement proves independent theoretical status. The decisive test is comparative:

After rigorous TCO, articulation-work, Human Factors, technical-debt, sunk-cost, and authority-fragmentation measures are included, does P1–P4 improve classification, explanation of persistence, prospective prediction, or correction?

If not, the framework should be reduced to a synthesis or diagnostic extension of those traditions. Its legitimacy depends on empirical incremental value, not vocabulary alone.

4. Human Meta-Work: Taxonomy, Attribution, and Measurement

The central empirical object is not all work associated with an AI-supported activity. It is the incremental human work required to make the AI configuration usable, acceptable, correctable, and governable relative to a selected baseline.

The taxonomy is adjacent to articulation work and patchwork but is not intended to rename them. Articulation-work research explains coordination necessary for cooperative arrangements (Schmidt & Bannon, 1992); patchwork documents situated labor that bridges incomplete AI systems and institutional practice (Fox et al., 2023). The present taxonomy adds functional separation, baseline attribution, materiality, actor incidence, and a direct link to continuation decisions.

The primary burden profile is:

M(t)=(C,O,S,V,G,Dr)

where:

C — Configuration and specification
O — Orchestration
S — Supervision
V — Verification
G — Governance
D_r — Realized configuration-debt cost

The vector is primary. A scalar total should be constructed only after transparent normalization and valuation.

4.1 Configuration and Specification — C

Configuration and specification include human work that defines, structures, or revises what the system is expected or permitted to do.

Included activities may involve:

defining objectives and constraints;
writing and revising prompts or policies;
selecting models, agents, tools, and data sources;
designing role structures and routing rules;
establishing permissions;
defining memory behavior;
creating evaluation criteria;
and translating domain requirements into machine-operable form.

Configuration is distinct from direct domain execution. Writing an analysis is direct work. Designing the agent workflow that will generate the analysis is configuration.

Configuration may be productive and reusable. Its inclusion in the burden profile does not classify it as waste.

4.2 Orchestration — O

Orchestration is the human work required to coordinate components, actors, states, and handoffs within the configuration regime.

It may include:

sequencing agents or tools;
resolving conflicting outputs;
synchronizing state;
assigning tasks;
transferring context;
managing dependencies;
and coordinating human and automated roles.

Orchestration differs from configuration in temporal function. Configuration establishes or changes the operating arrangement; orchestration coordinates its execution.

4.3 Supervision — S

Supervision is observation directed toward detecting abnormal, risky, or unacceptable system behavior.

It may include:

monitoring live operation;
watching for exceptions;
reviewing alerts;
maintaining situation awareness;
checking whether the system remains within scope;
and preparing to intervene.

Supervision answers: Is something occurring that requires attention?

4.4 Verification — V

Verification is human work directed toward determining whether an output, intermediate state, decision, or action is acceptable.

It may include:

factual checking;
source validation;
code testing;
legal or technical review;
comparison with ground truth;
evaluation of reasoning or provenance;
and approval of consequential action.

Verification answers: Is this result valid enough to accept?

Supervision and verification may occur within the same episode, but they should not be double-counted. Detecting an abnormal output is supervision. Determining that it is false is verification.

4.5 Governance — G

Governance includes human work required to establish, interpret, apply, document, and revise the rules under which the configuration regime operates.

It may include:

risk classification;
policy design;
permission approval;
accountability assignment;
incident review;
audit documentation;
stakeholder consultation;
gate decisions;
and management of legal or institutional obligations.

Governance is not equivalent to all management work. It is included where the activity concerns authorization, constraint, accountability, correction, or continuation of the AI configuration regime.

Governance can become recursive when a new control layer generates further configuration, monitoring, evaluation, and policy work.

4.6 Realized Configuration-Debt Cost — D_r

Configuration debt is represented through a stock-flow distinction. It is narrower than general technical debt but directly informed by research on hidden debt in machine-learning systems, including entanglement, configuration issues, undeclared dependencies, hidden feedback loops, and changes in the external world (Sculley et al., 2015).

The stock evolves as:

Ds(t+1)=Ds(t)+AD(t)-RD(t)

where:

Ds(t) is accumulated debt stock;
AD(t) is debt added;
RD(t) is debt retired.

Realized debt cost, Dr(t), is the human burden incurred during the period because earlier configuration choices must be understood, repaired, bypassed, migrated, or compensated for.

Examples include:

undocumented exceptions;
obsolete prompts;
permission drift;
duplicated policies;
hidden dependencies;
incompatible memories;
model-specific workarounds;
and abandoned integrations that remain operationally relevant.

Debt stock is not added automatically to current burden. Only realized cost enters M(t). Expected future realization should be reported separately.

4.7 Inclusion and Exclusion Rules

Human activity should be included where its dominant purpose is to make the AI configuration regime function, remain acceptable, or remain governable.

It should be excluded where it is ordinary direct work that would have occurred under the baseline.

Primary-code rule

Each time interval receives one primary component code according to its dominant purpose.

Secondary-tag rule

Secondary tags may preserve complexity without increasing counted time.

Temporal-separation rule

Distinct sequential episodes may receive different codes.

For example:

detect abnormal output — S;
determine that the output is false — V;
revise the prompt policy — C;
update the approval rule — G;
repair an undocumented dependency — D_r.

The same thirty minutes cannot be counted five times. Distinct intervals can be coded separately.

Dominant-purpose rule

Where an activity serves several functions simultaneously, the primary code should reflect the immediate operational purpose. Coders should record uncertainty rather than manufacture false precision.

4.8 Incremental Attribution

The relevant burden is:

Δ Ma ∣ b(T)=Ma(T)-Mb(T)

where a is the AI configuration regime and b is the selected baseline.

Negative component values are permitted. For example, the AI regime may reduce orchestration while increasing verification.

Attribution requires burden-accounting symmetry. Equivalent baseline functions must be measured according to the same rule.

4.9 Actor-Level Burden

For actor i:

Mi(T)=(Ci,Oi,Si,Vi,Gi,Dr,i)

The analysis should identify the distribution of:

Benefit ∣ Labor ∣ Risk ∣ Knowledge ∣ Authority

Aggregate burden can conceal transfer to:

expert reviewers;
downstream correction workers;
compliance staff;
operators;
customers;
or affected non-users.

4.10 Measurement Units

Possible units include:

time;
loaded labor cost;
scarce-expert hours;
cognitive interruption;
decision latency;
expected risk-adjusted burden;
and opportunity cost.

No unit should be treated as universal. The same activity may carry different implications according to actor, domain, stakes, and timing.

4.11 Coding Reliability

A measurement program should include:

a coding manual;
worked examples;
independent coders;
uncertainty labels;
adjudication procedures;
and inter-rater reliability reported separately for each component.

If the six components cannot be distinguished reliably, the taxonomy should be revised or simplified rather than protected through interpretive flexibility.

4.12 Meta-Work Coding Rules

Activity	Primary code	Inclusion condition	Frequent confusion
Define objective, prompt, role, permission, or evaluation rule	C	Work changes what the system is expected or allowed to do	Direct domain planning
Route tasks, reconcile agents, synchronize state	O	Work coordinates execution across components	Configuration of the routing design
Watch operation, alerts, exceptions, or boundaries	S	Work detects whether attention or intervention is required	Verification of correctness
Check factual, legal, technical, or functional validity	V	Work determines whether an output or action can be accepted	Supervision or ordinary quality work already in the baseline
Define accountability, approval, risk, policy, or continuation	G	Work governs authorization and consequence	General administration
Repair consequences of accumulated configuration debt	D_r	Earlier configuration creates current repair or workaround work	New configuration or ordinary maintenance

5. Agentic AI as an Integrative Technical Substrate

The AI Configuration Paradox is not a general claim that all complex technologies generate excessive overhead. Nor is it reducible to ordinary organizational dysfunction. Its specific relevance to agentic AI arises from a distinctive combination of technical properties that can connect, accelerate, and reinforce forms of human meta-work that would otherwise remain more limited or loosely coupled.

Current terminology distinguishes AI agents from more complex agentic-AI systems, but no single usage is yet universal. The OECD emphasizes autonomy, tool use, task decomposition, coordination among agents, extended operation, and socio-technical embedding, while NIST’s current standards initiative focuses on agents capable of autonomous action, secure operation, identity, authorization, and interoperability (OECD, 2026; NIST, 2026a).

Agentic AI is used here as a functional continuum rather than as a claim about genuine autonomy or a requirement that the system contain multiple agents. A system is agentic to the extent that it can perform several of the following functions with reduced direct human intervention:

maintain task state over time;
decompose objectives into subtasks;
select tools or models;
invoke external resources;
route work among components;
evaluate intermediate outputs;
revise plans;
retain memory;
trigger actions;
and coordinate multiple specialized roles.

These capabilities may create real value. The theory does not treat agency, autonomy, or complexity as intrinsically undesirable. The relevant issue is that agentic systems are unusually capable of generating new layers of specification, coordination, oversight, verification, governance, and repair.

The technical substrate therefore does not independently cause the paradox. It alters the conditions under which the P1–P4 configuration can form.

Agentic Properties + Human and Institutional Mechanisms → Increased Probability of a Persistent Configuration Regime

The relation is conditional. A tightly bounded agentic system with a clear objective, stable architecture, proportionate verification, and operative stopping authority may produce substantial net benefit without approaching the paradox.

5.1 Radical Extensibility

Agentic systems are often designed as extensible arrangements rather than closed tools. A conventional software application typically exposes a bounded set of functions. Extending it may require formal development, testing, deployment, and organizational approval.

An agentic system can often be extended through comparatively low-friction actions: adding another prompt, creating another role, attaching another tool, granting another permission, introducing another memory source, adding a reviewer or critic, modifying a routing instruction, or placing another agent between an existing input and output.

Each local addition may appear inexpensive. The cumulative regime may nevertheless become costly.

Low Apparent Local Extension Cost ≠ Low Cumulative Configuration Cost

An additional component may create new dependencies, handoffs, permissions, verification obligations, failure modes, state to reconcile, and governance questions.

Radical extensibility can therefore increase C, O, S, V, G, and later Dr. The system can grow through a sequence of locally defensible changes whose total burden was never evaluated as a whole.

5.2 Recursive Configurability

Agentic AI differs from many earlier tools because the system may configure, evaluate, or govern mechanisms that themselves configure, evaluate, or govern other mechanisms.

Examples include an agent that rewrites prompts for another agent, a planner that creates or assigns sub-agents, an evaluator that modifies routing rules, a monitoring agent that decides when another agent must be reviewed, or a governance agent that generates policies for tool use.

This creates recursive configurability:

Configuration of a System → Configuration by the System → Human Oversight of System-Generated Configuration

Humans may need to determine whether a system-generated change is valid, whether it altered the original objective, whether permissions remain appropriate, whether criteria drifted, whether the system is optimizing internal proxies, and whether a rollback state still exists.

Automation of configuration may therefore transform direct configuration work into meta-configuration and governance work rather than eliminate it.

5.3 Cheap Generation and Expensive Acceptance

Generative and agentic systems can produce candidate outputs, plans, code, analyses, decisions, or actions at low marginal production cost. Acceptance remains a different problem.

A candidate output may need factual, legal, technical, security, provenance, or risk verification. The economic and cognitive asymmetry is:

Generation Capacity>Human Acceptance Capacity

A system that generates ten times more material does not create ten times more usable value if the organization can responsibly validate only a fraction of it.

Agentic systems may intensify the problem because outputs are intermediate decisions that shape later planning, tool selection, memory, delegation, and synthesis. Verification may need to occur at the level of objective interpretation, plan validity, tool selection, intermediate state, final output, external action, and retained memory.

The burden depends on error detectability, domain stakes, ground truth, reversibility, and the cost of false acceptance. This is consistent with emerging evidence that practical oversight of software agents includes preventive control, co-planning, real-time monitoring, and post hoc review rather than a single final approval step (Dhanorkar et al., 2026).

5.4 Persistent State

Agentic systems may retain conversational history, task state, preferences, summaries, retrieved documents, tool outputs, permissions, plans, and learned or system-generated instructions.

Persistent state can improve continuity. It can also produce accumulated obligations. State must be selected, stored, updated, reconciled, protected, audited, corrected, and eventually deleted or migrated.

Past Configuration → Retained State → Future Behavior and Future Burden

A false summary, outdated instruction, or persistent permission can shape later behavior. Persistent state is therefore a route through which configuration debt and path dependence accumulate.

5.5 Anthropomorphic Role Proliferation

Agentic systems are frequently organized through labels such as planner, researcher, critic, reviewer, manager, security officer, evaluator, or compliance agent.

Role-based design can improve modularity. It can also conceal technical and institutional complexity behind familiar social language.

A “reviewer agent” does not necessarily possess independent incentives, evidence, legal responsibility, superior epistemic access, or authority to stop the system. Several agents may be generated by the same model, prompt family, data, and system objective.

Role differentiation should not be confused with epistemic or institutional independence. A perceived organizational problem may repeatedly be translated into another technical role, increasing the configuration regime that humans must understand and govern.

5.6 Internal Observability and Activity Abundance

Agentic systems generate abundant internal traces: messages, tool calls, token counts, task completions, planning steps, confidence estimates, reviewer scores, dashboards, and execution graphs.

These traces can improve debugging and accountability. They also create a measurement asymmetry. Internal activity is immediate, countable, visible, and attributable to the system. External benefit may be delayed, multidimensional, distributed, and difficult to attribute.

High Internal Observability + Low External-Effect Observability → Proxy Dependence Risk

Activity indicators can substitute for external-outcome evidence. They become problematic when they acquire decision authority independent of validated external benefit.

5.7 Distributed Ownership and Operation

Different actors may control model selection, infrastructure, data, prompts, deployment, operation, verification, risk acceptance, and continuation decisions.

The beneficiary may not be the operator. The operator may not be the verifier. The verifier may lack authority to restrict the system. The risk bearer may not have access to the relevant evidence.

Design ∣ Operation ∣ Benefit ∣ Verification ∣ Risk ∣ Authority

Distributed ownership is not inherently defective. The problem arises when no actor can assemble the complete decision object.

5.8 Exit Dependence

Agentic systems can become difficult to leave through workflow redesign, loss of manual competence, data-format dependence, model-specific prompts, custom integrations, persistent memory, permission structures, organizational specialization, and disappearance of alternative processes.

Adoption → Integration → Dependence → Higher Exit Cost

An organization may continue not because the current configuration remains superior, but because comparison or rollback is no longer feasible.

Continuation Because of Superior Value ≠ Continuation Because Exit Has Become Costly

Exit dependence is a reinforcing mechanism, not a sufficient condition for P3 or P4.

5.9 Coupling Across the Meta-Work Vector

Agentic property	Primary meta-work effects	Candidate mechanism or risk
Radical extensibility	C, O, V, G, D_r	Exploration lock; commitment accumulation
Recursive configurability	C, S, V, G	Means–ends reversal; governance recursion
Cheap generation	V, S, O	Verification bottleneck; oversight transfer
Persistent state	C, G, D_r	Configuration debt; exit dependence
Role proliferation	C, O, V, G	Fragmented responsibility; false independence
Internal observability	V, G and decision metrics	Progress substitution; proxy capture
Distributed ownership	All components across actors	Functional fragmentation
Exit dependence	G, D_r, intervention capacity	Escalation and persistent continuation

The table identifies plausible routes for empirical testing, not deterministic relationships. The same property may reduce burden in one regime and increase it in another.

5.10 Comparison and Positive Design Conditions

The paradox is possible outside agentic AI. Complex enterprise software and bureaucratic systems can generate lifecycle burden, goal displacement, proxy dependence, escalation, and lock-in.

Agentic AI is theoretically important because it combines open-ended natural-language specification, dynamic planning, tool use, autonomy, state, internal evaluation, inter-agent delegation, and low-friction modification in one substrate.

Agenticity should be treated as a multidimensional architecture variable rather than a binary label. Relevant dimensions include planning depth, action authority, tool access, persistence, dependency count, inter-agent coupling, ability to modify operating rules, and reversibility.

The properties do not make the paradox inevitable. They may be governed through narrow objectives, adequacy thresholds, bounded architecture, reversible components, minimal permissions, external-outcome metrics, preserved human competence, declared baselines, and operative stopping authority.

Agentic properties may also reduce meta-work through automated provenance, constrained orchestration, temporary state, machine-readable policies, and modular rollback.

The theory is therefore not anti-agentic. It asks whether the architecture’s benefits remain greater than the complete burden it creates and whether institutions retain the capacity to act when that relation deteriorates.

6. Classical Human-Sciences Mechanisms

The AI Configuration Paradox does not require a new psychology or sociology for every part of its explanation. Its candidate mechanisms draw on classical research concerning rationalization, bureaucracy, bounded decision-making, indicators, commitment, organizational learning, and human–automation interaction. The contribution lies in disciplined transfer and integration under agentic conditions.

For each mechanism, five questions must remain separate:

What did the source establish?
Which part transfers legitimately to agentic AI?
What does not transfer without new evidence?
What article-specific extension is proposed?
Which part of P2–P4 can the mechanism help explain?

None of the mechanisms is constitutive. P1–P4 define the paradox; mechanisms explain possible formation and persistence.

6.1 Formal Rationalization and Substantive Purpose

Weber’s account of rationalization distinguishes calculable, rule-governed, internally coherent ordering from judgments grounded in substantive values or ends. Formal rationality can be indispensable for predictability and administration, yet internally coherent procedure does not guarantee substantive success (Weber, 1978).

The legitimate transfer is limited:

Internal Formal Coherence ≠ External Substantive Value

Prompts, schemas, roles, permissions, evaluations, memory rules, and governance procedures can become more complete while the evaluative objective remains unchanged or deteriorates. The article does not claim that Weber predicted agentic AI or that formalization is inherently defective. It uses the distinction to deny that architecture maturity, procedural completeness, or activity counts are sufficient evidence for P3.

6.2 Means–Ends Reversal

Merton explains how procedures established as means can acquire independent practical value, producing goal displacement and rigid adherence detached from the circumstance or purpose that justified the rule (Merton, 1940).

Within an AI configuration regime:

Evaluative Objective → Configuration as Means → Maintenance of Configuration → Configuration as Operative Objective

The extension is not limited to bureaucratic rules. Models, agents, memory, metrics, permissions, integrations, and governance layers can become organizationally protected means. The mechanism contributes to P3 where maintaining the means consumes resources required by the end, and to P4 where the means become the principal evidence for their own continuation. Necessary maintenance remains outside the mechanism unless detachment from the objective is demonstrated.

6.3 Bounded Rationality, Search, and Satisficing

Simon rejects models that assume unlimited information and computation. Under bounded rationality, decision makers search under constraints and use aspiration levels or adequacy criteria rather than compute a global optimum (Simon, 1955). Satisficing is not low quality; it is a termination rule for decision-making under real limits.

Agentic extensibility can enlarge the configuration search space without enlarging human decision capacity proportionately:

Bounded Human Rationality + Open-Ended Configurability → Potentially Unbounded Search Burden

The article-specific term satisficing deficit denotes the absence of an operative adequacy condition. The governance derivation follows:

Performance ≥ Declared Adequacy Threshold → Presumption Against Unjustified Additional Complexity

This does not prohibit improvement after adequacy. It shifts the burden of proof to marginal benefit, complete marginal burden, and reversibility.

6.4 Indicators, Progress Substitution, and Proxy Capture

Campbell’s evaluation research shows that consequential indicators can become targets of adaptation, thereby weakening their validity as representations of the condition they were intended to measure (Campbell, 1979).

Agentic systems produce dense internal indicators: tool calls, completed subtasks, reviewer scores, traces, utilization, architecture counts, and confidence values. The theory separates two stages:

perceptual progress substitution — visible internal activity inflates perceived benefit;
behavioral proxy capture — a decision-bearing proxy reorganizes architecture, effort, and continuation around itself.

The article-specific extension is the configuration feedback:

Internal Observability → Decision Dependence → Behavioral Reorganization → Further Configuration

Metrics are not presumed corrupt. The claim requires evidence that proxy pressure changes decisions independently of validated external outcomes.

6.5 Escalation of Commitment

Staw’s experiment shows that negative consequences can increase commitment where decision makers bear responsibility for the original choice (Staw, 1976). Continuation after negative evidence is not automatically irrational; new investment may be prospectively justified. The mechanism concerns the influence of prior commitment on interpretation and resource allocation.

AI configurations can accumulate money, specialized knowledge, professional identity, public claims, and technical dependencies. Weak outcomes may therefore trigger another rescue move—another agent, evaluator, model, memory layer, or governance control—rather than simplification.

Escalation is a candidate mediator between P3 evidence and P4. It is neither necessary nor sufficient: a committed sponsor may still face effective stopping authority, while a system may persist without personal commitment because authority is fragmented or exit is costly.

6.6 Exploration, Exploitation, and Exploration Lock

March distinguishes exploration—search, variation, experimentation, discovery—from exploitation—refinement, implementation, and use of established capabilities (March, 1991). Neither is universally superior; their returns differ across time and organizational location.

Agentic systems lower the apparent cost of local experiments while leaving cumulative comparison, coordination, adaptation, and debt costly. Exploration lock denotes a regime in which new configuration possibilities repeatedly delay stable operation without a bounded learning rationale.

Legitimate exploration requires a declared learning objective, hypotheses, budget, evidence, duration, and decision rule. The mechanism contributes to P4 only where “learning” becomes an indefinitely renewable justification rather than a bounded program.

6.7 Articulation, Patchwork, and Distributed Non-Decision

Articulation-work research explains how cooperative arrangements require additional work to align actors, resources, sequences, and contingencies (Schmidt & Bannon, 1992). Patchwork research shows that AI integration relies on situated bridging and repair labor often omitted from formal system descriptions (Fox et al., 2023).

The article-specific institutional synthesis is distributed non-decision. Benefit, labor, risk, knowledge, and authority can be distributed so that no actor possesses the complete decision object:

Benefit ∣ Labor ∣ Risk ∣ Knowledge ∣ Authority

Every actor may perform a locally defined role while no actor is responsible for total net value or empowered to change the system state. Division of labor is not itself dysfunctional. Fragmentation becomes explanatory where it prevents burden evidence from reaching an actor capable of imposing a consequence.

6.8 Automation Irony, Oversight Transfer, and Exit Dependence

Bainbridge’s automation irony, Endsley and Kiris’s out-of-the-loop findings, and Parasuraman and Riley’s differentiated account of automation use show why reduced direct execution can preserve or intensify human responsibility for exceptions, monitoring, acceptance, and recovery (Bainbridge, 1983; Endsley & Kiris, 1995; Parasuraman & Riley, 1997).

The transfer to agentic AI is supported—but not proven universally—by current evidence that developers perform proactive and reactive oversight around software agents (Dhanorkar et al., 2026).

The article extends momentary takeover into institutional exit dependence:

Delegation → Competence or Alternative-Process Decay → Higher Exit Cost → Continuation Dependence

Automation need not cause competence decay; training, rehearsal, transparency, bounded autonomy, and maintained alternatives can preserve intervention capacity. Exit dependence is an empirically testable reinforcing mechanism, not a presumption.

6.9 Pseudo-Automation and Burden Substitution

Pseudo-automation shows that reported automation can shift tasks onto customers while making the residual frontline work more parallel, problem-oriented, and relationally demanding (Moradi et al., 2025). The present theory generalizes the distributional diagnostic without claiming that every burden transfer is pseudo-automation.

Configuration-burden substitution requires four elements:

a principal beneficiary receives the reported gain;
another actor absorbs substantial configuration, verification, correction, governance, or risk burden;
that burden is excluded or undervalued in the governing metric;
the burden-bearing actor lacks proportionate correction authority.

This mechanism can reveal that an apparently positive aggregate result depends on an incomplete system boundary.

6.10 Configuration Debt as a Human-Cost Pathway

Research on hidden technical debt in ML systems identifies entanglement, data and configuration dependencies, hidden feedback loops, and changes that make local modification expensive (Sculley et al., 2015).

The present theory does not claim originality for technical debt. Its extension is to distinguish accumulated stock from realized human cost, then connect that cost to actor incidence, materiality, exit, and continuation. Debt contributes to P2 only when it realizes as current human work; it contributes to P3 only under a materiality rule; and it contributes to P4 only when evidence fails to alter the regime.

6.11 Mechanism Interaction

The mechanisms may reinforce one another:

formalization can support means–ends reversal;
absent adequacy can support exploration lock;
internal observability can support proxy capture;
accumulated investment can support escalation;
articulation burden can be hidden by distributed ownership;
delegation can increase exit dependence;
and governance introduced to control these processes can itself become recursive.

A possible sequence is:

Formalization → More Configuration → More Internal Indicators → Proxy-Based Continuation → Commitment → Exit Dependence → Weak Termination

This is one candidate pathway, not a universal causal chain.

6.12 Source-Fidelity and Originality Boundary

The article inherits mechanisms from Weber, Merton, Simon, Campbell, Staw, March, CSCW, technical-debt research, and Human Factors. It does not attribute the following article-specific constructions directly to those sources:

AI Configuration Paradox;
configuration displacement;
configuration progress substitution;
terminal configuration failure;
configuration equilibrium;
configuration-burden substitution;
satisficing deficit;
exploration lock;
stopping authority;
or the SHIP / RESTRICT / HOLD / ROLLBACK gate.

The correct scholarly relation is:

Established traditions supply component mechanisms and adjacent empirical objects; the article re-specifies and integrates them into a conjunctively defined, baseline-relative, and governance-linked candidate theory.

7. Integrated Causal Architecture

The AI Configuration Paradox is defined by P1–P4, but those conditions do not specify a single causal sequence.

A case may reach the same paradoxical state through different combinations of technical properties, organizational conditions, cognitive mechanisms, institutional structures, and domain constraints.

The causal model therefore distinguishes:

constitutive conditions — what must be present for classification;
antecedents — conditions that increase the probability of formation;
mediators — mechanisms through which antecedents affect burden, displacement, or continuation;
moderators — variables that strengthen, weaken, or redirect effects;
outcomes — observable states associated with P2–P4;
feedback loops — processes through which outcomes generate further configuration.

This prevents definitional inflation and causal overcompression.

7.1 Constitutive Conditions

The classification remains:

AICPd,a ∣ b,T ⇔ P1d,T ∧ P2a ∣ b,T ∧ P3d,a ∣ b,T ∧ P4a,T

P1 establishes the instrumental reference point. P2 identifies incremental human burden. P3 establishes material displacement. P4 establishes continuation without bounded learning or an effective stopping path.

A causal explanation must answer separately:

Why did substantial meta-work arise?
Why did that burden become materially displacing?
Why did the regime continue after relevant evidence emerged?

7.2 Candidate Antecedents

Candidate antecedents include:

goal ambiguity — the evaluative objective is abstract, multidimensional, contested, or under-operationalized;
capability-first adoption — the organization begins with what the AI can do rather than which objective requires improvement;
weak baseline specification — current performance and burden are not credibly measured;
high verification asymmetry — outputs are cheap to produce but difficult to validate;
distributed ownership — architecture, burden, benefit, risk, and authority are divided;
high reversibility cost — integration makes later rollback expensive;
environmental instability — repeated adaptation is required and stabilization becomes difficult to observe.

These conditions are neither necessary nor sufficient.

7.3 Enabling Technical Properties

The agentic properties described in Section 5 act as enabling and amplifying conditions.

The disciplined relation is:

Antecedent + Agentic Enabling Property + Human or Institutional Mechanism → Candidate P2–P4 Outcome

7.4 Candidate Mediators

Candidate mediators include:

open-ended formalization;
means–ends reversal;
satisficing deficit;
verification transfer;
progress substitution;
behavioral proxy capture;
escalation of commitment;
exploration lock;
fragmented responsibility;
and exit dependence.

These mechanisms explain how antecedents and technical properties translate into burden, displacement, or continuation.

7.5 Moderators

Moderators include:

objective clarity;
task decomposability;
error detectability;
reversibility;
domain stakes;
ground-truth availability;
burden visibility;
distribution of authority;
preserved human competence;
and configuration budgets.

The same factor can operate differently across pathways. High stakes may justify more verification while also increasing the need for effective stopping authority.

7.6 Five Canonical Causal Pathways

The five pathways are canonical but not exhaustive. They may overlap, operate in different orders, or remain absent.

Path A — Goal-Ambiguity Path

Underdefined Objective → No Stable Adequacy Threshold → Open-Ended Formalization → Persistent Reconfiguration → P2 and Candidate P3

Principal moderators include objective clarity, external metrics, aspiration levels, and configuration budgets.

Path B — Verification Path

Cheap Generation → Increased Candidate Output → Expensive Acceptance → Oversight and Verification Transfer → Material Displacement

The pathway is strongest where errors are hard to detect, consequences are high, and ground truth is weak.

Path C — Proxy Path

High Internal Observability → Internal Activity Indicators → Decision Dependence → Behavioral Reorganization → Proxy-Based Continuation

The pathway should be distinguished from legitimate use of validated leading indicators.

Path D — Institutional Path

Fragmented Responsibility → Partial Evidence Across Actors → Incomplete Decision Object → Weak Stopping Authority → Persistent Continuation

The system continues through distributed non-decision.

Path E — Dependence Path

Delegation → Workflow Integration → Competence or Alternative-Process Decay → Higher Exit Cost → Reduced Responsiveness to Comparative Value

The regime may continue because alternatives have become difficult to restore rather than because current value remains superior.

7.7 Reinforcing Feedback Loops

Architecture–burden loop

New Component → New Coordination and Verification Burden → New Control Mechanism → Additional Component

Proxy–expansion loop

Internal Metric Improvement → Perceived Progress → Further Investment → More Internal Activity

Commitment–complexity loop

Investment → Specialization and Identity → Resistance to Simplification → Further Investment

Dependence–continuation loop

Continued Use → Competence Decay and Integration → Higher Exit Cost → Continued Use

Governance-recursion loop

Observed Risk → Additional Governance Layer → New Governance Burden → Need for Governance of Governance

These loops are hypotheses for process tracing, not assumptions.

7.8 Formation, Displacement, and Persistence

The model should not collapse all outcomes into one variable.

Burden Formation ≠ Material Displacement ≠ Persistent Continuation

P2 may arise through architecture growth, verification transfer, persistent state, coordination density, or governance expansion. P3 depends on the relation between burden, external benefit, baseline, materiality, time horizon, and distribution. P4 depends on learning rationale, evidence visibility, thresholds, authority, consequence, and exit cost.

7.9 Negative Cases and Alternative Orders

The model must represent:

high complexity without paradox;
P2 and P3 followed by effective correction;
persistent continuation without configuration-regime-attributable displacement;
paradox under a clearly defined objective;
fragmentation that precedes technical complexity;
and commitment that precedes proxy selection.

The order of causes need not be fixed, and reciprocal causation is possible.

7.10 Intervention Points

Causal problem	Candidate intervention
Goal ambiguity	Declare objective, baseline, beneficiary, and adequacy threshold
Open-ended formalization	Impose configuration budget and terminal-acceptance rule
Verification transfer	Measure incremental verification and restrict generation volume or scope
Proxy capture	Require external-outcome evidence and prohibit activity-only continuation
Escalation	Use prospective justification and exclude sunk investment
Exploration lock	Use expiration, hypothesis registers, and transition criteria
Fragmented responsibility	Create a unified decision object and named stopping authority
Exit dependence	Preserve alternatives, portability, manual competence, and tested rollback
Governance recursion	Apply metric minimality, governance budgets, and self-application

These are theoretically derived intervention points, not universal prescriptions.

7.11 Terminal Configuration Failure and Configuration Equilibrium

Terminal Configuration Failure is the article-specific name for the P4-level failure in which relevant burden or displacement evidence cannot produce terminal acceptance, simplification, restriction, suspension, or rollback. It is not merely the absence of a final project date. It is failure of the operative correction chain:

Signal → Threshold → Authority → Consequence

The chain may fail because a signal is not measured, a threshold is absent, authority is divided, or the authorized consequence cannot be implemented. Terminal configuration failure is therefore observable independently from the mechanisms proposed to explain it.

Configuration Equilibrium describes a persistent regime in which architecture, human meta-work, internal metrics, commitments, dependencies, and authority relations mutually stabilize continuation. The term does not denote a welfare optimum, a Nash equilibrium, or a formally solved game. It denotes an equilibrium-like socio-technical persistence pattern:

Configuration Activity + Continuation Incentives + Exit Dependence + Weak Correction → Self-Stabilizing Regime

Configuration equilibrium is not an additional constitutive condition. A case satisfies the paradox through P1–P4. The equilibrium concept describes a developed persistence state that may emerge after P4, and it remains a candidate explanatory description requiring longitudinal evidence.

7.12 Integrated Causal Model

The complete map can be represented as:

Evaluative Objective (P1) + Antecedents + Agentic Enabling Properties + Mediators → Configuration-Regime-Attributable Meta-Work (P2) → Material Displacement (P3) + Ineffective Learning or Stopping Structure → Persistent Continuation Failure (P4)

This is a disciplined conceptual map, not a statistical equation.

8. Formal Conceptual Framework

The formal framework serves three limited purposes. It preserves distinctions that ordinary claims about automation, productivity, or cost tend to collapse; makes P2 and P3 empirically inspectable; and supports comparison among alternative configuration regimes without pretending that every benefit, burden, risk, or human value can be reduced automatically to one number.

The equations are measurement structures rather than universal laws. They do not predict outcomes without empirical parameterization and do not replace substantive judgment.

The governing hierarchy is:

Multidimensional Profiles → Domain-Specific Measurement → Optional Normalization → Optional Scalar Valuation → Decision Rule → Governance Consequence

8.1 Unit of Analysis and Notation

Let:

d denote the external decision domain or activity;
T denote the evaluation period;
a denote an alternative configuration regime;
b denote a selected baseline;
i denote an affected actor;
n denote an architecture containing n components.

The complete unit of analysis is not the model alone. To avoid symbol collisions, the article uses the following object:

ℛd,a ∣ b,T={Ωd,𝒜a,Ma,Ya,βa ∣ b,Λa,Γa,Ξa}

where:

Ωd = identifiable evaluative objective;
𝒜a = AI and workflow architecture;
Ma = human meta-work profile;
Ya = observed external-outcome profile;
βa ∣ b = relation between alternative a and selected baseline b;
Λa = distribution of benefit, burden, risk, knowledge, and authority;
Γa = decision and stopping structure;
Ξa = exit and reversibility profile.

The symbols identify distinct analytical objects; they do not imply commensurability.

8.2 External-Benefit Profile

The external-benefit profile is:

Bd,a(T)=(Bq,Bt,Bc,Br,Bu)d,a,T

Possible dimensions are:

Bq — quality improvement;
Bt — time or latency improvement;
Bc — resource or direct-execution cost improvement;
Br — risk reduction;
Bu — user or stakeholder utility.

Quality may include accuracy, completeness, relevance, consistency, legal or technical correctness, explanatory value, or domain-specific performance.

Time benefit may include reduced completion time, lower decision latency, faster response, or earlier availability of a useful outcome. Time savings should not include delay transferred to downstream verification or approval.

Resource benefit may include lower direct-execution labor, reduced non-human operating cost, fewer external services, lower rework, or released scarce expertise. The same saving should not be counted twice.

Risk benefit may include lower probability of material error, lower expected loss, better coverage, stronger resilience, or improved detection of rare conditions. Expected-value treatment is appropriate only where probability and consequence estimates are defensible.

User or stakeholder utility may include usability, accessibility, autonomy, satisfaction, perceived control, or reduced frustration.

Raw outcomes should be recorded before conversion into relative benefit. If subtraction is meaningful for a dimension:

Bd,a ∣ b(T)=Yd,a(T)-Yd,b(T)

Ordinal, categorical, or non-linear outcomes require another declared comparison rule.

8.3 Human Meta-Work Profile

The burden profile is:

Ma(T)=(C,O,S,V,G,Dr)a,T

The incremental burden relative to baseline b is:

Δ Ma ∣ b(T)=Ma(T)-Mb(T)

A negative component value is permitted. The vector should remain disaggregated because a decrease in one component may be accompanied by an increase in another with different implications.

8.4 Profile Comparison Before Scalarization

Benefit and burden profiles can often be compared without reducing them to one number.

Dominance

Alternative a dominates baseline b where it is no worse on every declared benefit and burden dimension and materially better on at least one.

Threshold comparison

An alternative may be accepted where it satisfies declared thresholds:

qa ≥ qmin

ra ≤ rmax

τa ≤ τmax

Ma,k ≤ Mk,max

This approach is appropriate where adequacy matters more than optimization.

Multi-criteria decision rule

A domain may specify minimum conditions, priority ordering, veto dimensions, and trade-off rules.

Non-compensatory constraints

Some dimensions should not be tradable against others. Examples include legal prohibitions, rights, minimum safety requirements, unacceptable discrimination, absence of valid consent, or irreversible high-consequence risk.

The valid structure is:

Eligibility Constraint → Profile Comparison → Optional Scalar Valuation

not automatic authorization from scalar net value.

8.5 Normalization and Valuation

Where a common decision unit is justified:

Bd,a ∣ b*(T)=Vd(Bd,a ∣ b(T))

Ma*(T)=Wd(Ma(T))

The functions may use monetary value, equivalent labor cost, normalized utility, expected loss, or another defensible unit. They must disclose unit, weights, empirical source, uncertainty, distributional assumptions, time horizon, discounting, exclusions, and non-compensatory constraints.

Scalarization is invalid where it assigns arbitrary weights, combines incompatible units without normalization, treats ordinal scores as interval measures without justification, hides burden transfer, monetizes veto constraints as ordinary costs, or double-counts effects.

Sensitivity analysis should report the result under plausible valuations and identify which assumptions change the decision.

8.6 Scalar Net Value

Where valid common valuation exists:

Nd,a ∣ b*(T)=Bd,a ∣ b*(T)-Δ Ma ∣ b*(T)

where:

Δ Ma ∣ b*(T)=Ma*(T)-Mb*(T)

A positive result indicates that valued external advantage exceeds valued incremental meta-work relative to the selected baseline. It does not establish ethical acceptability, legal validity, fairness, safety, or superiority under another baseline.

For repeated periods:

Nd,a ∣ b*(0:T)=∑t=0Tδt[Bd,a ∣ b*(t)-Δ Ma ∣ b*(t)]

with 0<δ ≤ 1. Discounting should be used only where the domain requires intertemporal comparison and should not conceal delayed debt or future burden transfer.

8.7 Configuration Debt

The debt stock evolves as:

Ds(t+1)=Ds(t)+AD(t)-RD(t)

Only realized debt cost Dr(t) enters current burden. Expected future debt should be reported separately:

𝔼[Dr(t+1:T)]

A useful report distinguishes current burden, debt stock, expected future realization, and uncertainty.

8.8 Material Displacement

Where burden and baseline resources can be expressed in the same unit u:

DRa ∣ bu(T)=Δ Ma ∣ bu(T)Rd,b0,u(T)

Possible units include expert-hours, total labor, budget, or another bounded resource.

P3 is satisfied only where at least one predeclared materiality criterion is met. Examples include a resource-share threshold, materially negative net value, failure against a credible simpler baseline, delay that undermines the objective, burden on a defined actor beyond an accepted threshold, or displacement of another activity of equal or greater value.

No universal numerical threshold is proposed.

8.9 Configuration Progress Substitution and Proxy Capture

Perceptual progress substitution can be represented conceptually as:

Bd,a*(T)=Bd,a*(T)+εP(T), εP(T) ≥ 0

where εP is perceived-benefit inflation associated with internal activity or sophistication.

Behavioral proxy capture is not merely perception error. It occurs when a proxy becomes decision-bearing and changes resource allocation, architecture, human behavior, or continuation.

A general research form is:

𝒜t+1=f(𝒜t,Zt,Yt,Kt)

where Zt is an internal proxy and Yt is external outcome evidence. This is an empirical research form, not a calibrated universal function.

8.10 Marginal Component Test

For a proposed component n+1, where valid valuation exists:

Δ Bd,n+1*>Δ Mn+1*

The comparison should include direct external effect, additional configuration, orchestration, supervision, verification, governance, expected debt, and effect on reversibility.

The relevant baseline is the architecture without the proposed component. Interaction effects may require factorial comparison, ablation, staged deployment, or architecture-bundle analysis.

The same logic applies to removal. A component should be considered for removal where burden reduction exceeds benefit loss.

8.11 Actor-Level Distribution

For actor i:

Ma,i(T)=(Ci,Oi,Si,Vi,Gi,Dr,i)

Where actor-specific scalar valuation is legitimate:

Nd,a ∣ b,i*(T)=Bd,a ∣ b,i*(T)-Δ Ma ∣ b,i*(T)

The analyst should report benefit, labor, risk, knowledge, and authority separately. A system may have positive aggregate value while senior experts absorb disproportionate verification burden, downstream workers perform unrecorded correction, or non-users bear risk.

8.12 Worked Example: Quarterly Regulatory Evidence Brief

The following example is synthetic. A policy unit must produce a quarterly evidence brief within ten working days, with acceptable analytical quality, a critical-error probability no greater than 1.5%, and a defensible audit trail.

Three regimes are compared:

A — Direct human workflow
B — Bounded AI assistant
C — Multi-agent configuration

Raw external outcomes

Dimension	A	B	C
Quality score	86	89	91
Cycle time	8 days	6 days	5 days
Direct-execution human labor	100 h	55 h	30 h
Estimated critical-error probability	1.6%	1.3%	1.0%
User-utility score	78	83	87

Alternative A fails the 1.5% safety floor. B and C qualify for further comparison.

Meta-work profiles

Component	A	B	C
C	4	8	24
O	10	4	16
S	0	5	12
V	35	24	38
G	6	6	14
D_r	0	1	8
Total	55	48	112

The multi-agent regime reduces direct execution most strongly but generates the largest meta-work burden.

Assume loaded rates of 160/h for C,120/h for O, 100/h for S,150/h for V, 180/h for G, and140/h for Dr.

Then:

MA*=$8,170

MB*=$7,080

MC*=$16,300

Thus:

Δ MC ∣ A*=$8,130

Δ MC ∣ B*=$9,220

Assume each quality point is valued at 500, each working day saved at1,500, each direct-execution hour saved at 100, and each percentage-point reduction in critical-error probability at8,000. User utility is reported separately.

Relative to A:

BC ∣ A*=$18,800

NC ∣ A*=$10,670

Relative to B:

BC ∣ B*=$7,400

NC ∣ B*=-$1,820

The conclusion changes because the baseline changes. C is strongly favorable against direct human execution but unfavorable against the simpler AI baseline.

Suppose the unit declares a quarterly resource base of 120 expert-hours and a materiality threshold of 25%. The incremental meta-work of C relative to B is 64 hours:

DRC ∣ Bhours=64120=0.533

Both the resource-share threshold and negative incremental net-value rule are met. Thus P3 is satisfied relative to B.

The appropriate gate is RESTRICT, not rollback to the manual process. The bounded assistant already meets the safety constraint and creates lower burden. P4 is not yet satisfied. If the organization restricts the architecture, the paradox is prevented. If it continues C indefinitely because activity is high, architecture appears sophisticated, prior investment is large, or no actor can simplify it, P4 may become satisfied.

The example demonstrates that a system can reach P3 without yet constituting the full paradox.

9. Baselines, Materiality, Distribution, and Exit

The classification depends on comparison. Human meta-work is incremental only relative to a baseline. External benefit is meaningful only relative to an alternative state. Material displacement cannot be established without specifying which resources, actors, and opportunities are displaced. Exit dependence cannot be evaluated without identifying what the regime could realistically be replaced by.

The analysis must answer:

Compared with what?
Material according to which rule?
Benefit and burden for whom?
How costly is it to restrict, replace, or leave the regime?

9.1 Baseline as Decision Comparator

The baseline is not necessarily the state immediately before adoption. It should correspond to the decision being evaluated.

To evaluate adoption, direct human or pre-adoption work may be relevant. To evaluate an agentic extension, bounded AI assistance may be relevant. To evaluate a component, the baseline is the same architecture without it. To evaluate rollback, the comparator must be a state that can be restored or reconstructed.

The correct question is not whether the system is better than nothing, but whether it is better than the most decision-relevant credible alternative under the same objective, quality floor, time horizon, and accounting rule.

9.2 Baseline Hierarchy

Candidate baselines include:

Non-adoption baseline

Appropriate only where the activity itself is optional.

Direct human-execution baseline

Useful for identifying changes in execution, quality, risk, and transfer into meta-work. It may become operationally unavailable after competence decay.

Pre-adoption organizational baseline

Includes existing review, management, governance, delays, errors, and coordination. It is often best for incremental attribution but may be artificially weak.

Simple AI-assistance baseline

A bounded tool supports a human-led workflow without broad autonomy, persistent state, or multi-agent coordination. This is often the most important comparator for agentic expansion.

Bounded single-agent baseline

Tests whether multi-agent or recursively configurable architecture adds value beyond bounded agency.

Ablated architecture baseline

The current system is compared with the same system after removal of one agent, evaluator, memory layer, approval stage, or integration.

Credible external alternative

Another vendor, model, conventional system, service, redesigned human workflow, or hybrid process that satisfies the same objective and eligibility constraints.

Current extended configuration

The present regime serves as baseline when evaluating restriction, simplification, or rollback.

9.3 Baseline-Selection Rules

A valid baseline should satisfy:

decision relevance;
functional comparability;
common eligibility constraints;
temporal comparability;
realistic availability;
burden-accounting symmetry.

It is invalid to count AI governance but not human governance, AI verification but not baseline review, or AI setup while ignoring baseline training.

9.4 Multiple-Baseline Robustness

Let:

ℬ={b1,b2,…,bk}

be the credible baseline set.

A configuration is robustly favorable where it satisfies non-compensatory constraints, produces positive or dominant comparative value across principal baselines, and does not depend on excluding material actor-level burden.

A configuration is baseline-sensitive where the decision changes by comparator. This may show that the value lies in AI adoption generally rather than in the added complexity of the current architecture.

A justification is baseline-fragile where favorability appears only against a deliberately weak, unrealistic, obsolete, or incompletely measured comparator.

9.5 Baseline Contamination

A baseline can be altered by the configuration regime. Human workers change processes, manual competence declines, previous workflows disappear, downstream teams are redesigned around AI output, and governance is created specifically for the system.

Where possible, preserve pre-adoption data, parallel comparisons, and archival documentation. A reconstructed baseline should disclose uncertainty.

The disappearance of an alternative should not be treated as proof that the current regime is superior. It may be evidence of endogenous baseline decay and exit dependence.

9.6 Materiality Rules

Materiality is not statistical significance, subjective irritation, or visible complexity. A burden is material where it affects the evaluative objective, resource allocation, stakeholder condition, or comparative value enough to alter the decision.

Resource-share materiality

DRa ∣ bu(T)=Δ Ma ∣ bu(T)Rd,b0,u(T)

A domain may declare a threshold based on expert capacity, labor, budget, attention, or another bounded resource.

Net-value materiality

Material displacement may be established where scalar net value becomes negative, expected benefit falls below a declared minimum, or the configuration loses a material share of the benefit that justified adoption.

Delay materiality

A useful result delivered after the relevant deadline may fail the objective:

τa>τmax

Opportunity-cost materiality

Meta-work may displace another activity of equal or greater value.

Stakeholder-burden materiality

An actor-specific burden threshold may be non-compensatory even where aggregate value is positive.

Multi-rule materiality

A domain may use several sufficient or conjunctive criteria. It must state whether criteria are additive, vetoing, or uncertainty-sensitive.

9.7 Actor-Level Incidence

The distributional structure is:

ℒa(T)={Bi,Mi,Ri,Ki,Ai}i=1n

where benefit, meta-work, risk, knowledge, and authority may be concentrated or separated.

Relevant separations include beneficiary–configurer, operator–verifier, risk–authority, knowledge–authority, and authority–burden separation.

These separations are not inherently illegitimate. They become analytically important when they affect visibility, incentives, or correction.

9.8 Configuration-Burden Substitution

Configuration-burden substitution occurs where:

one actor or group receives a principal benefit;
another absorbs substantial configuration, verification, correction, governance, or risk burden;
the transferred burden is hidden, excluded, or undervalued by the governing metric;
the burden-bearing actor lacks proportionate correction authority.

The concept is not equivalent to inequality. Specialized burden may be legitimate where it is visible, resourced, authorized, and externally valuable.

Evidence may include time-use data, downstream correction logs, incident reports, interview divergence, unrecorded workarounds, and differences between management metrics and worker experience.

9.9 Exit and Reversibility

Exit is multidimensional:

technical reversibility — components, permissions, memory, or tools can be removed;
data reversibility — state and provenance can be exported and reused;
operational reversibility — the organization can continue the activity after rollback;
cognitive reversibility — humans retain competence and situation awareness;
institutional reversibility — an authorized actor can order and implement rollback;
vendor reversibility — contracts and dependencies permit migration.

The exit profile is:

Xa(T)=(Xtech,Xdata,Xops,Xcog,Xinst,Xvendor)

Possible measures include rollback time, service interruption, migration effort, functions lacking manual alternatives, takeover performance, dependency count, baseline-reconstruction cost, and authority latency.

Where compatible valuation exists:

Xa*(T)=Qd(Xa(T))

Exit cost should remain separate from current burden unless and until incurred or explicitly expected.

9.10 Exit Dependence and Preservation

Continuation because net value remains positive is different from continuation because exit cost is high.

A regime with weak comparative value may continue because alternatives have decayed. High exit cost can justify temporary continuation during transition but does not prove indefinite superiority.

Exit dependence can be reduced through modular architecture, bounded permissions, state export, documented dependencies, maintained manual procedures, periodic human practice, tested rollback, parallel fallback systems, time-limited integrations, and contractual portability.

Preservation itself creates burden and should be counted in the appropriate meta-work component.

9.11 Baseline and Materiality Selection Matrix

Decision question	Principal baseline	Secondary robustness baseline	Materiality focus	Exit requirement
Should AI be adopted?	Pre-adoption process or direct human workflow	Non-adoption or external alternative	Net value, quality, risk, resource share	Preserve non-AI process during pilot
Should simple assistance become agentic?	Bounded AI assistant	Direct human workflow	Incremental burden and marginal benefit	Preserve simpler architecture as fallback
Should multi-agent architecture be retained?	Bounded single-agent system	Ablated variants	Marginal component and relational burden	Measure rollback and dependency cost
Should a component be added?	Same architecture without component	Alternative component	Complete marginal burden	Require component-level removability
Should the system be restricted?	Current extended configuration	Simpler qualifying state	Burden reduction with preserved adequacy	Confirm continuity after restriction
Should the system be rolled back?	Current configuration	Prior stable state or external alternative	Current value, risk, transition cost	Test technical, data, cognitive, institutional reversibility
Should experimentation continue?	Current experiment	Stop, stabilize, or alternative experiment	Learning value, budget, duration	Require expiration and recoverable state

10. Empirical Hypotheses and Research Program

The AI Configuration Paradox is a conceptual theory until its constructs can be measured reliably, distinguished from adjacent explanations, and used to predict outcomes beyond established variables. The empirical program must test four levels of claim:

measurement validity — whether P1–P4 and the six meta-work components can be observed and coded reliably;
mechanism validity — whether proposed antecedents and mediators predict burden, displacement, or continuation;
incremental validity — whether the framework adds information beyond TCO, articulation work, technical debt, workload, sunk cost, and general organizational dysfunction;
intervention validity — whether correction-sovereignty mechanisms alter system state without destroying justified value, learning, or safety.

No single study can establish all four levels. The preferred unit is a configuration episode or decision period, not a model or organization treated as one undifferentiated case.

10.1 Empirical Objects and Core Measures

For each regime, research should identify:

the evaluative objective Ωd;
architecture 𝒜a;
credible baseline b;
meta-work profile Ma;
external outcomes Ya;
actor distribution Λa;
lifecycle phase;
stopping structure Γa;
and exit profile Ξa.

P1 is measured through objective specificity, beneficiary, baseline, adequacy, period, and objective stability. P2 is measured through incremental component burden. P3 requires a predeclared materiality rule. P4 requires longitudinal evidence that a material signal did not produce an effective, bounded learning or stopping response.

10.2 H1 — Relational Complexity–Burden Hypothesis

As relational architecture complexity increases, marginal human meta-work will eventually grow faster than marginal external benefit in at least some task and domain classes.

Theoretical derivation

Low-friction local extensions can create additional handoffs, shared state, permissions, exception pathways, and validation obligations. The relevant explanatory variable is relational complexity rather than raw component count: a small but tightly coupled architecture may impose more meta-work than a larger modular system.

Variables

Independent variable: dependency density, handoffs, state synchronization, permission interactions, tool overlap, heterogeneity, feedback depth, and cross-component correction.

Dependent variables: incremental component burden, configuration latency, incident-repair time, verification burden, and marginal net value.

Mediators: handoff failure, state inconsistency, exception frequency, duplicate verification, and governance expansion.

Moderators: modularity, decomposability, interface standardization, observability, automated validation, and reversibility.

Decision-relevant baselines and study design

Use bounded single-agent, ablated, and lower-coupling equivalent workflows. Suitable designs include controlled architecture experiments, longitudinal field studies, graph-based dependency analysis, and staged ablation.

Predicted result

Relational complexity should predict meta-work more strongly than component count, particularly for poorly decomposable tasks with shared state and difficult validation.

Strongest alternative explanation

The apparent effect may be produced by task difficulty, immature tools, team inexperience, or generally poor implementation rather than relational architecture itself.

Falsification or narrowing condition

H1 should be reduced if the association disappears after task difficulty, maturity, and implementation quality are controlled, or if added relational complexity consistently creates proportionally greater external benefit without corresponding burden.

10.3 H2 — Goal-Ambiguity Hypothesis

More abstract, multidimensional, contested, or under-specified objectives will produce longer configuration periods, more objective revision, greater architecture change, and lower probability of terminal acceptance.

Theoretical derivation

Where adequacy cannot be recognized, further configuration remains locally defensible:

Goal Ambiguity → Satisficing Deficit → Open-Ended Formalization → Persistent Reconfiguration

Variables

Independent variable: competing objectives, absent adequacy criteria, stakeholder disagreement, revision frequency, and distance between stated goals and observable outcomes.

Dependent variables: duration of configuration, number of architecture and metric revisions, C burden, and probability of transition to routine operation.

Mediators: metric proliferation, repeated reclassification of failure, and search-space expansion.

Moderators: domain expertise, external ground truth, configuration budgets, and named stopping authority.

Decision-relevant baselines and study design

Compare projects with clear and ambiguous objectives, use randomized task framing, or introduce objective clarification in an interrupted field design.

Predicted result

Ambiguity should predict longer stabilization, more revision, and lower terminal-acceptance probability.

Strongest alternative explanation

Environmental instability, genuine research novelty, or changing stakeholder needs may require repeated configuration even where the original objective was adequately specified.

Falsification or narrowing condition

H2 should be weakened if ambiguity adds no information after environmental instability and task novelty are controlled, or if ambiguous projects converge as reliably as precise projects.

10.4 H3 — Oversight-Transfer Hypothesis

Increasing technical autonomy will reduce direct execution work while increasing the proportion of total human labor devoted to orchestration, supervision, verification, governance, and maintenance of intervention capacity.

Theoretical derivation

Automation may remove routine production while preserving responsibility for objective setting, acceptance, exceptions, and recovery. Classical Human Factors research predicts transfer into monitoring and takeover obligations; current developer evidence indicates proactive as well as reactive oversight around software agents (Bainbridge, 1983; Dhanorkar et al., 2026).

Variables

Independent variable: action authority, planning depth, tool access, persistence, execution horizon, and approval frequency.

Dependent variables: direct-execution hours, O/S/V/G hours, meta-work share, intervention frequency, takeover performance, and total labor.

Mediators: output volume, error opacity, exception complexity, reduced direct participation, and action consequences.

Moderators: error detectability, stakes, validation quality, training, transparency, and reversibility.

Decision-relevant baselines and study design

Compare manual work, suggestion-only AI, approval-gated agency, and higher-autonomy execution using within-task experiments, time-allocation field studies, and takeover tests.

Predicted result

Direct execution should decline while the share of remaining labor devoted to meta-work rises. Total labor may decrease, remain stable, or increase.

Strongest alternative explanation

Observed oversight growth may be a temporary learning or interface-maturity effect rather than a durable consequence of technical autonomy.

Falsification or narrowing condition

H3 fails as a broad claim if autonomy consistently reduces direct work and all meta-work without weakening intervention capacity, or if oversight transfer disappears after stabilization across domains.

10.5 H4 — Verification-Bottleneck Hypothesis

As generation cost falls, verification will become a larger share of total human task work where errors are difficult to detect, consequences are high, or reliable ground truth is unavailable.

Theoretical derivation and measurement rule

Low-cost generation can increase candidate-output volume faster than responsible acceptance capacity. Verification share must use one common scalar unit u rather than combining a scalar with a burden vector:

VSau(T)=Vau(T)Ha,execu(T)+Cau(T)+Oau(T)+Sau(T)+Vau(T)+Gau(T)+Dr,au(T)

where Ha,execu is direct human execution converted into the same unit u.

Variables

Independent variables: generation rate, error detectability, stakes, ground-truth availability, and task openness.

Dependent variables: verification hours and share, review backlog, acceptance latency, false acceptance, correction cycles, and accepted usable output.

Mediators: candidate-output volume, error plausibility, provenance difficulty, and downstream propagation.

Moderators: automated tests, structured output, traceability, constrained generation, and sampling.

Decision-relevant baselines and study design

Use manual generation, bounded assistance, and matched output-volume conditions. Suitable studies include controlled volume experiments, high- versus low-ground-truth domain comparisons, and longitudinal production observation.

Predicted result

Verification share should rise most sharply in opaque, consequential domains where plausible errors are difficult to detect.

Strongest alternative explanation

The burden may reflect low model quality, redundant review policy, conservative institutional culture, or temporary unfamiliarity rather than a general production–acceptance asymmetry.

Falsification or narrowing condition

H4 should be weakened if accepted output scales without proportional verification growth under opaque and high-stakes conditions, or if the burden is fully explained by temporary unfamiliarity or poor system quality.

10.6 H5a — Perceptual Progress-Substitution Hypothesis

Visible indicators of internal activity and sophistication will increase perceived benefit when external performance is held constant.

Theoretical derivation

Agentic systems expose abundant internal activity—agent counts, traces, tool calls, dashboards, and internal scores—while external benefit may be delayed, multidimensional, or difficult to attribute. Evaluators may therefore use visible technical activity as evidence of progress even when the externally relevant outcome has not changed.

Variables

Independent variables: number of agents, workflow diagrams, traces, task counts, dashboards, internal scores, and model-brand information.

Dependent variables: perceived value, continuation support, expansion support, confidence, and perceived organizational progress.

Mediators: perceived sophistication, system effort, completeness, and control.

Moderators: evaluator expertise, external-outcome visibility, metric literacy, and downstream accountability.

Decision-relevant baseline and study design

The primary baseline is the same output, reliability information, and external performance presented without internal-activity or sophistication cues. Randomized experiments should vary architecture descriptions and dashboards while holding externally relevant evidence constant.

Predicted result

Activity and sophistication cues should increase perceived value and support for continuation or expansion, especially where external outcomes are weakly visible.

Strongest alternative explanation

The cues may communicate legitimate information about resilience, coverage, auditability, or redundancy that is not captured by the immediate external-performance measure.

Falsification or narrowing condition

H5a should be rejected if internal cues have no independent effect after legitimate information about coverage, resilience, reliability, and auditability is controlled, or if responses track only externally relevant information contained in the cues.

10.7 H5b — Behavioral Proxy-Capture Hypothesis

When internal activity indicators become decision-bearing, architecture and human effort will shift toward improving those indicators even where external outcomes remain stable or deteriorate.

Theoretical derivation

Perceptual substitution becomes behavioral proxy capture when an internal indicator influences funding, renewal, evaluation, or architecture authorization. Once consequential, the metric can reorganize human effort and technical design around its own production rather than the evaluative objective.

Variables

Independent variable: degree to which an internal metric affects funding, renewal, performance evaluation, or architecture authorization.

Dependent variables: configuration effort, architecture expansion, time devoted to proxy improvement, proxy–outcome divergence, and continuation probability.

Mediators: incentive pressure, reporting visibility, managerial dependence, and manipulability.

Moderators: independent external validation, metric plurality, expiration, audit, and authority to reject the metric.

Decision-relevant baselines and study design

Compare the same metric when it is merely descriptive with periods or units in which it becomes decision-bearing. Use difference-in-differences when metrics become tied to renewal, laboratory resource-allocation studies, and documentary process tracing of later architecture decisions.

Predicted result

Decision-bearing proxies should predict later architecture expansion and resource allocation after external outcomes are controlled.

Strongest alternative explanation

The internal metric may be a valid leading indicator of delayed external benefit, making apparently proxy-directed adaptation rational rather than corrupting.

Falsification or narrowing condition

H5b should be rejected if decision dependence does not change architecture or effort after external outcomes are controlled, or if the indicator consistently predicts later external benefit without evidence of strategic adaptation that degrades its validity.

10.8 H6 — Correction-Sovereignty Package Hypothesis

Configuration regimes with a complete and operative stopping structure will exhibit lower persistent burden and faster correction after material-displacement evidence than review-only regimes.

The operative structure is:

Signal → Threshold → Authority → Consequence

Theoretical derivation

P3 evidence cannot prevent P4 unless it reaches an authorized decision process capable of changing system state. Review-only structures can document burden while allowing continuation by default; thresholds, named authority, expiration, and rollback capacity should make correction operational.

Variables

Independent variable: completeness of thresholds, named authority, enforceability, stakeholder access, expiration, rollback readiness, and decision latency.

Dependent variables: time from signal to decision, simplification probability, restriction or rollback probability, later burden, continuation beyond threshold, and preserved external benefit.

Mediators: evidence visibility, decision ownership, lower coordination cost for correction, and reduced sunk-cost influence.

Moderators: stakes, uncertainty, reversibility, regulation, and organizational power.

Decision-relevant baselines and study design

Compare complete packages with advisory review, ordinary governance, or the same regime before implementation. Use factorial designs to separate threshold, authority, expiration, representation, and rollback components; use staged implementation to observe process change.

Predicted result

Complete stopping structures should reduce continuation beyond declared thresholds, shorten signal-to-consequence latency, and accelerate proportionate correction while preserving justified external benefit.

Strongest alternative explanation

Organizations adopting complete stopping structures may already possess clearer objectives, stronger management, more resources, or lower tolerance for weak systems; these background factors may explain the observed correction.

Falsification or narrowing condition

H6 should be rejected or sharply reduced if the package merely adds governance burden, does not change system state, fails to preserve justified benefit, or adds no explanatory value beyond general management quality.

10.9 Classification Validation

Before estimating prevalence, the P1–P4 classification requires:

content validity from AI governance, Human Factors, organizational research, measurement, and domain experts;
inter-rater reliability for each P condition and the full conjunction;
convergent validity with workload, articulation, lifecycle, and technical-debt measures;
discriminant validity from high workload, poor implementation, technical debt, and inertia;
predictive validity for persistent architecture expansion, burden, and delayed simplification.

High agreement on P2 with weak agreement on P3 or P4 would indicate that the framework remains a workload taxonomy rather than a distinct paradox theory.

10.10 Incremental-Validity Tests

Comparison models should include:

total cost of ownership;
articulation work and patchwork labor;
technical debt;
perceived workload;
task complexity;
organizational maturity;
automation reliance;
sunk cost;
implementation quality;
authority fragmentation.

Only after reliable measurement should researchers use multilevel models, survival analysis of time to restriction or rollback, structural models, or out-of-sample prediction. The central question is whether P1–P4 improves prediction of persistent continuation, architecture expansion, failure to simplify, or burden transfer.

10.11 Negative Cases and Longitudinal Requirement

Deliberate negative cases should include:

complex agentic systems with low meta-work;
high-burden systems with greater justified benefit;
P3 episodes followed by rapid correction;
long but bounded research programs;
extensive governance without recursion;
and organizations preserving credible exit capacity.

P4 is inherently temporal. The preferred event history is:

Adoption → Burden Formation → Materiality Signal → Decision → Consequence or Non-Consequence

Cross-sectional dissatisfaction cannot establish persistent continuation failure.

10.12 Research Sequence and Quality Requirements

Construct and coding development
Controlled mechanism studies
Longitudinal field studies
Comparative process tracing
Governance intervention studies
Replication and domain transfer

Quality requirements include preregistration where feasible, transparent baseline selection, lifecycle-phase control, actor-level reporting, active search for informal work, sensitivity analysis, and publication of null findings.

10.13 Program-Level Rejection Conditions

The theory should be narrowed, reclassified, or rejected if P1–P4 cannot be coded reliably; P2 cannot be separated from ordinary workload or articulation; P3 cannot be operationalized independently; P4 adds no information beyond inertia or authority failure; the main mechanisms fail; correction sovereignty does not change behavior; and the complete framework adds no out-of-sample value beyond established constructs.

11. Governance Derivation and the Configuration Continuation Gate

Governance should not enter the theory as an external policy layer appended after the analysis. The governance architecture must be derived from the mechanisms that generate and sustain the configuration regime.

The distinctive failure is that evidence of burden, displacement, or weak comparative value may remain disconnected from an operative consequence.

The required structure is:

Signal → Metric → Threshold → Authority → Gate → Consequence

A signal without a metric remains anecdotal. A metric without a threshold remains descriptive. A threshold without authority remains advisory. Authority without a defined gate produces inconsistency. A gate without an enforceable consequence remains symbolic.

11.1 Governance Objective

The purpose of the Configuration Continuation Gate is:

Preserve the instrumental relation between the configuration regime and the evaluative objective by ensuring that evidence concerning benefit, burden, displacement, uncertainty, and reversibility can produce a proportionate change in system state.

This lifecycle orientation is consistent with the NIST AI Risk Management Framework and its generative-AI profile; the specific four-state gate remains an article-derived candidate rather than a NIST requirement (NIST, 2023; NIST, 2024).

The gate decides whether a particular regime should continue as defined, continue under restriction, pause pending evidence or correction, or return to a simpler or previously validated state.

The four authorized states are:

SHIP
RESTRICT
HOLD
ROLLBACK

No fifth canonical state is required.

11.2 Minimum Decision Object

The gate should evaluate:

𝒟gated,a ∣ b,T={Ωd,Bd,a ∣ b,Ma ∣ b,P,U,Ξa,Λa,ℬ}

where:

Ωd = evaluative objective and adequacy condition;
Bd,a ∣ b = external-benefit profile relative to baseline;
Ma ∣ b = incremental human meta-work profile;
P = P1–P4 evidence;
U = uncertainty and learning value;
Ξa = exit and reversibility profile;
Λa = distribution of benefit, burden, risk, knowledge, and authority;
ℬ = credible alternatives and baselines.

This is a decision record, not a scalar formula. Scalar net value may be included where valid, but it does not replace the full object.

11.3 Derived Responses to the Mechanisms

Means–ends reversal

At each major decision, the record should restate the evaluative objective, external outcome evidence, baseline, necessary architecture, and distinction between internal indicators and external effects.

Internal Mechanism Improvement ⇏ SHIP

Satisficing deficit

The governance architecture should require adequacy and quality or safety thresholds. After adequacy, additional complexity carries a new burden of proof.

Performance ≥ Adequacy Threshold → Presumption Against Unjustified Expansion

Proxy capture

Every decision-bearing metric should be classified as external outcome, validated leading indicator, operational diagnostic, or unvalidated proxy.

Internal Indicator Alone ⇏ Continuation Authorization

Escalation of commitment

Continuation decisions should be prospective. Sunk cost should be separated from avoidable future cost, expected future benefit, transition cost, and expected exit cost.

Prior Investment ⇏ Continuation Justification

Exploration lock

A valid experimental state specifies a learning objective, hypothesis, budget, time limit, configuration limit, evidence, stopping rule, and transition criterion. HOLD can permit bounded experimentation without granting ordinary SHIP authorization.

Fragmented responsibility

The gate should assemble the complete distribution of benefit, labor, risk, knowledge, and authority. Stopping authority must be named, enforceable, informed, and capable of imposing a consequence.

Exit dependence

High rollback cost does not establish current superiority.

High Exit Cost ⇏ SHIP

Temporary RESTRICT may be justified while alternatives and exit capacity are restored.

11.4 SHIP

SHIP authorizes continued operation under defined scope and conditions.

It is justified where:

the evaluative objective remains clear;
eligibility constraints are satisfied;
benefit is positive or adequate against credible baselines;
incremental meta-work is proportionate;
no material displacement threshold is crossed;
burden distribution is visible and acceptable;
uncertainty is bounded;
and restriction or rollback remains feasible.

SHIP should specify scope, architecture version, permissions, monitoring obligations, reassessment date, and expiration or renewal conditions. It is conditional authorization, not permanent approval.

11.5 RESTRICT

RESTRICT authorizes continued operation under reduced scope, complexity, autonomy, permission, duration, or scale.

It is appropriate where the system retains useful external benefit but a component or dimension creates disproportionate burden or risk, a simpler baseline performs nearly as well, material displacement can be reduced through narrowing, or exit dependence requires staged simplification.

Restrictions may remove agents, reduce planning depth, limit tool access, shorten memory, require approval for external actions, reduce deployment scale, sample verification, or confine the system to task classes with strong evidence.

The logic is:

Preserve Valid Benefit − Remove Unjustified Burden

11.6 HOLD

HOLD suspends ordinary continuation, expansion, or authorization pending specified evidence or correction.

It is appropriate where evidence is insufficient, materiality cannot be assessed, baseline data are missing, a safety or rights question is unresolved, stopping authority is unclear, proxy validity is uncertain, experimentation lacks a valid rule, or rollback readiness is inadequate.

HOLD should specify what is paused, what may continue, which evidence is required, who must produce it, by when, and which decision follows if the evidence remains unavailable.

An indefinite HOLD without a decision rule reproduces continuation failure.

11.7 ROLLBACK

ROLLBACK returns the regime to a prior, simpler, safer, more reversible, or better-validated state.

It is appropriate where P3 is established, continuation lacks a valid learning rationale, a credible simpler alternative is superior, restrictions cannot reduce burden sufficiently, the objective is no longer served, a non-compensatory constraint is violated, or continued operation materially increases irreversible risk or dependence.

Rollback need not mean abandonment of AI. It may mean movement from multi-agent to single-agent operation, autonomous action to suggestion-only use, persistent memory to temporary state, broad to limited tools, or AI assistance to a validated manual process.

11.8 Evidence Pattern to Gate Outcome

Evidence pattern	Default tendency	Required reasoning
Strong external benefit, proportionate burden, valid baseline, bounded uncertainty, preserved exit	SHIP	Continue under defined scope, monitoring, and reassessment
Positive benefit but one component or layer creates disproportionate burden	RESTRICT	Preserve justified function while removing or narrowing the burden source
Missing baseline, unresolved proxy validity, unclear materiality, incomplete authority, or unbounded experiment	HOLD	Pause expansion or ordinary authorization until specified evidence exists
Negative comparative value against a credible simpler baseline, material displacement, no valid learning rationale	ROLLBACK or RESTRICT	Roll back where simplification cannot preserve adequate value; otherwise restrict
High immediate exit cost but weak current comparative value	RESTRICT with transition plan	Continue only as required to restore alternatives and reduce dependence
Non-compensatory safety, legal, or rights violation	HOLD or ROLLBACK	Scalar benefit cannot override the constraint
P3 signal followed by rapid effective correction	RESTRICT, then reassess	P4 may be prevented if consequence is operative
Internal activity rises while external outcomes remain flat and proxy validity is weak	HOLD or RESTRICT	Remove proxy dependence and require external evidence
Adequacy achieved and further complexity lacks marginal justification	RESTRICT or deny expansion	Post-adequacy complexity carries a new burden of proof

The table states default tendencies, not automatic decisions.

11.9 Uncertainty

Uncertainty should not automatically justify continuation or termination.

The gate should distinguish uncertainty about benefit, burden, risk, reversibility, and learning value.

Continued bounded experimentation may be justified where uncertainty can be reduced, information value is material, the experiment is limited, and failure is reversible.

HOLD or restriction is more appropriate where consequences are high, evidence cannot be independently verified, rollback is difficult, burden is already material, or experimentation increases dependence.

11.10 Governance Anti-Recursion

The governance system can itself become part of the paradox through new metrics, approval layers, documentation, governance agents, and recurring review.

The gate should therefore apply the theory to itself through:

metric minimality — every metric supports a defined decision;
rule transparency — every rule states purpose, trigger, owner, consequence, review date, and expiration;
governance budget — time, cost, latency, approval layers, and scarce expertise are bounded;
no automatic governance-agent proliferation;
independent simplification authority;
expiration by default for temporary rules, permissions, exceptions, and pilots;
self-application — the gate’s own burden and marginal value are periodically evaluated.

11.11 Procedural Sequence

A practical procedure can follow eight steps:

restate the evaluative objective;
identify the current architecture;
measure external benefit and meta-work;
apply baseline and materiality rules;
inspect distribution and exit;
assess uncertainty and learning rationale;
verify stopping authority;
issue a time-bounded gate decision.

The procedure should remain proportional. A low-stakes bounded assistant should not require the same apparatus as a consequential autonomous system.

11.12 LoopGuard-AI Boundary

LoopGuard-AI may serve as a candidate reference architecture for structuring evidence, connecting thresholds to authority, recording decisions, and implementing the four states.

It does not establish construct validity, prevalence, optimality, or governance effectiveness.

The relation is:

AI Configuration Paradox → Governance Requirements → LoopGuard-AI as Candidate Implementation

not validation of the theory by implementation.

12. Strongest Objections and Replies

The theory should be tested against interpretations that can explain the same observations without adopting P1–P4. Each objection below is treated in three steps: its valid force, the precise reply, and the discriminating test.

12.1 “This Is Only Total Cost of Ownership”

Valid force. Rigorous TCO already includes implementation, integration, training, support, maintenance, migration, governance, and replacement. If the theory merely states that AI has hidden expenses, it adds little.

Reply. P1–P4 adds a classificatory and institutional claim. High complete cost can remain consciously justified, bounded, and responsive to evidence. The paradox requires attributable burden, material displacement, and persistence despite an ineffective learning or stopping path. It can also arise through scarce expertise, delay, or burden transfer not captured well in accounting.

Discriminating test. After complete TCO is measured, does P1–P4 improve prediction of continued architecture expansion or failure to simplify? If not, the theory should be reduced to a TCO extension.

12.2 “This Is Articulation Work or Patchwork Under a New Name”

Valid force. CSCW and recent AI-integration research already show that cooperative systems require alignment, contextualization, repair, and hidden bridging labor (Schmidt & Bannon, 1992; Fox et al., 2023).

Reply. Human meta-work is deliberately adjacent to those constructs. The residual claim is not discovery of integration labor. It is the separation of six functions, incremental attribution against a baseline, materiality, actor incidence, exit, and P4. Articulation may be productive and proportionate; patchwork may repair a valuable system. Neither condition is automatically paradoxical.

Discriminating test. Can articulation and patchwork measures distinguish corrected high-burden cases from persistent materially displacing regimes as well as P1–P4?

12.3 “This Is Pseudo-Automation or Labor Offsetting”

Valid force. Reported automation can shift labor to customers or other workers and intensify the work that remains (Moradi et al., 2025). Configuration-burden substitution overlaps directly with this problem.

Reply. Pseudo-automation is a distributional mechanism, not the complete configuration theory. The paradox can arise without customer labor transfer through verification, governance, debt, or exit dependence. Pseudo-automation can also occur without agentic AI or P4.

Discriminating test. Does the full classification explain persistence and correction beyond labor-offset measures?

12.4 “This Is Ordinary Technical Debt”

Valid force. ML systems already exhibit entanglement, hidden feedback, configuration dependencies, and costly maintenance (Sculley et al., 2015).

Reply. The article does not claim originality for debt. It distinguishes debt stock from realized human cost and connects that cost to objective, baseline, materiality, distribution, exit, and continuation. Debt that is visible, proportionate, and retired effectively does not satisfy the paradox.

Discriminating test. After technical-debt measures are included, does P4 improve prediction of non-correction?

12.5 “This Is Ordinary Organizational Dysfunction”

Valid force. Goal ambiguity, fragmented authority, weak metrics, sunk cost, and inertia predate AI.

Reply. The theory is interactional rather than technologically deterministic. It asks whether agentic properties make those mechanisms easier to instantiate, couple, and expand, and whether the resulting incremental meta-work satisfies P2–P4. General dysfunction alone is insufficient.

Discriminating test. Compare AI and non-AI projects within the same organizational environment and control for pre-existing dysfunction.

12.6 “Configuration Is Productive Work”

Valid force. Configuration may create reusable infrastructure, learning, resilience, technical capacity, or intrinsic craftsmanship. Treating it as waste would be conceptually defective.

Reply. P2 is non-evaluative. P3 requires material displacement under a declared objective and rule. Configuration may itself be a legitimate objective if success is not defined merely as continuing to configure. Objective revision must be explicit and cannot retroactively prove the original claim.

Discriminating test. Was configuration value declared, evidenced, bounded, and assigned to an identifiable beneficiary before it was used to defend continuation?

12.7 “All New Technologies Require Learning and Stabilization”

Valid force. Setup, training, workflow redesign, and temporary inefficiency are normal. Early evaluation can bias analysis against innovation.

Reply. The framework separates setup, stabilization, routine maintenance, and persistent reconfiguration. P4 protects bounded learning through objective, hypothesis, budget, duration, evidence, stopping criteria, and reauthorization.

Discriminating test. Does burden decline and architecture converge according to the learning plan, or does learning become an indefinite justification without cumulative decision value?

12.8 “High Oversight Is the Price of Safety”

Valid force. High-stakes systems may require extensive verification and governance. Reducing these merely to improve productivity could be unsafe or unlawful.

Reply. Oversight is counted because it is part of the complete regime, not because it is presumptively excessive. Risk reduction belongs in the benefit profile, and safety or rights floors may be non-compensatory. The relevant comparison applies the same requirements to all alternatives.

Discriminating test. Does the complete system, including required safeguards, outperform credible safer or simpler alternatives?

12.9 “The User Can Simply Stop”

Valid force. Low-dependency individual tools may be genuinely optional and reversible; such cases are unlikely to satisfy P4.

Reply. Stopping is an institutional and technical capacity. The burden bearer may lack authority; workflows, memory, data, contracts, competence, and reporting may be dependent on the system. Clicking close is not equivalent to restoring the activity under evaluation under a viable alternative.

Discriminating test. Who can stop which component, how quickly, with what data, competence, operational continuity, and authority?

12.10 “The Theory Is Anti-Complexity”

Valid force. Complexity can produce robustness, specialization, safety, and capability unavailable to simple systems.

Reply. The theory does not rank simplicity morally. It requires comparative justification for relational complexity. A complex modular system may impose less meta-work than a smaller tightly coupled one. RESTRICT preserves justified complexity while removing components lacking marginal value.

Discriminating test. Does complexity create external value greater than its complete relational burden against a qualifying simpler baseline?

12.11 “Subjective Benefit Cannot Be Measured Reliably”

Valid force. Autonomy, intellectual stimulation, confidence, creativity, and satisfaction may resist monetization and vary across people and time.

Reply. The benefit model is vector-first. Subjective utility can remain separate and uncertain. The framework permits thresholds, dominance, and multi-criteria rules rather than forced scalarization. Measurement difficulty weakens inference; it does not justify ignoring benefit or treating it as unlimited.

Discriminating test. Whose utility is measured, by what instrument, over what period, and is it being used to replace a different failed objective?

12.12 “P4 Makes the Definition Circular”

Valid force. A theory that defines the paradox through continuation and then explains continuation through the paradox would be circular.

Reply. P4 is observed independently through signal, threshold, authority, consequence, and actual system-state change. Fragmentation, proxy capture, commitment, and exit dependence are explanatory variables, not evidence sufficient for P4.

Discriminating test. Can independent coders identify continuation failure without coding the proposed mechanisms?

12.13 “The Governance Mechanism Can Become Its Own Paradox”

Valid force. Metrics, review boards, documentation, monitoring agents, and repeated audits can create another expanding burden regime.

Reply. The objection is accepted as a core design risk. Governance enters G and other relevant components; it is subject to metric minimality, budgets, expiration, marginal-component testing, simplification authority, and self-application. LoopGuard-AI is not presumed effective.

Discriminating test. Does the gate reduce persistent burden and improve correction compared with its own complete cost?

12.14 “P1–P4 Is Too Demanding to Be Useful”

Valid force. Real organizations often lack reliable objectives, baselines, and burden data. A broad hidden-labor concept is easier to apply.

Reply. The demanding conjunction prevents every difficult project from becoming “the paradox.” Partial findings remain reportable: P2 burden, P3 displacement, or weak authority. Insufficient evidence is a valid result.

Discriminating test. Does the stricter classification improve inter-rater reliability and reduce false positives enough to justify its evidentiary cost?

12.15 “The Theory Is Normative Rather Than Empirical”

Valid force. Objectives, adequacy, materiality, distribution, and stopping authority involve values and institutional choices.

Reply. The framework separates descriptive measurement from evaluative rules. Researchers can observe the declared objective, burden, threshold, authority, and consequence while also documenting who selected the rule and whose interests it represents. The theory provides analytical visibility, not value-free resolution.

Discriminating test. Are descriptive findings stable across evaluators even when normative gate decisions differ?

12.16 “Better Alternatives Always Exist”

Valid force. Continuous comparison can recreate the satisficing deficit and force endless optimization.

Reply. The framework requires credible decision-relevant alternatives, not proof of global optimality. Once adequacy is achieved and no material displacement is established, SHIP can be justified until burden, risk, architecture, alternatives, or authorization materially changes.

Discriminating test. Is the proposed alternative feasible, qualifying, and decision-relevant rather than hypothetical?

12.17 “The Paradox Is Temporary and Will Vanish as AI Matures”

Valid force. Reliability, standards, interfaces, validation, and organizational experience may reduce configuration and oversight burden.

Reply. The theory permits a transitional or domain-specific result. Technical maturity may reduce some mechanisms while leaving goal displacement, proxy pressure, authority fragmentation, and exit dependence. Maturity is a moderator, not a protected assumption.

Discriminating test. Do P2–P4 reliably disappear across mature cohorts and task classes, or does burden migrate into higher-order governance and dependency?

12.18 “Benefits Appear Only at Scale”

Valid force. Fixed setup cost can be amortized through reuse, standardization, and volume. Short studies may overcount setup and undercount future value.

Reply. Longitudinal analysis separates fixed burden, variable meta-work, recurring debt, and scale-dependent benefit. Scale is not presumed: output, verification, governance, and incident exposure may grow differently.

Discriminating test. How do marginal external benefit and complete marginal meta-work change with actual—not merely projected—scale?

12.19 “Stopping Authority Creates Governance Domination”

Valid force. A stopping authority can suppress experimentation, centralize power, protect incumbents, or overreact to uncertainty.

Reply. Stopping authority is evidence-bound, proportionate, reviewable, and connected to declared thresholds. Four states permit narrower responses than binary approval or prohibition. Authority without constraints is arbitrary; review without authority is symbolic.

Discriminating test. Does the institution preserve appeal, reasons, expiration, stakeholder evidence, and proportionality while still producing operative consequences?

12.20 Integrated Reduction Test

The strongest objection is explanatory redundancy. The theory should not be defended by the coherence of its vocabulary. Its independent status depends on whether the conjunction of objective, attributable meta-work, material displacement, and continuation failure improves classification, prediction, or intervention after adjacent constructs are controlled.

If it does not, the work should be simplified or reclassified as a synthesis of TCO, articulation work, technical debt, Human Factors, and organizational persistence. If it does, it identifies a distinct socio-technical failure mode.

13. Falsification, Boundary Conditions, and Maturity

A theory of the AI Configuration Paradox is useful only if it can fail. It must be capable of losing explanatory territory.

Failure may occur because the constructs cannot be measured, the classification is unreliable, mechanisms fail, the framework adds no value beyond existing concepts, governance fails to alter behavior, or the theory applies only to a narrower class of systems.

13.1 Levels of Falsification

Descriptive claim

AI adoption can generate substantial human meta-work in M(t). This claim fails where the taxonomy captures nothing meaningful beyond ordinary lifecycle categories or cannot be coded reliably.

Classificatory claim

P1–P4 identifies a coherent condition distinct from high workload, implementation failure, or continuation inertia. This fails where the conditions cannot be coded reliably or do not form a stable discriminant pattern.

Causal claim

The candidate mechanisms explain formation and persistence. This fails where they do not predict burden, displacement, or continuation after controls.

Incremental explanatory claim

The framework adds value beyond TCO, workload, debt, reliance, sunk cost, inertia, and authority fragmentation. This fails where established constructs perform equally well.

Governance claim

A complete stopping structure reduces persistent burden while preserving value. This fails where the package does not alter continuation, merely adds administration, or performs no better than ordinary review.

The theory may survive at one level while failing at another.

13.2 Construct-Failure Conditions

The framework should be revised if:

the six meta-work components cannot be distinguished;
incremental attribution is not feasible;
materiality cannot be operationalized beyond retrospective preference;
P4 cannot be observed independently through signal, threshold, authority, consequence, and actual continuation.

Where causal attribution is unavailable, the report should use AI-associated burden rather than configuration-regime-attributable incremental burden.

13.3 Classification-Failure Conditions

The structure should be revised if it cannot distinguish:

P2 without P3;
P3 followed by effective correction;
P4-like organizational continuation without P3;
cases without an identifiable objective;
and borderline cases without excessive evaluator dependence.

If P4 alone dominates, the framework becomes a general theory of persistence. If P2 automatically implies P3, materiality has failed.

13.4 Mechanism-Failure Conditions

Each mechanism can be removed or narrowed independently.

H1 fails if relational complexity does not predict burden beyond count and task difficulty.
H2 fails if goal ambiguity does not predict persistent reconfiguration.
H3 narrows if oversight transfer is only a short stabilization effect.
H4 weakens if verification does not become a bottleneck in opaque high-stakes settings.
H5a fails if internal activity cues have no independent perceptual effect.
H5b fails if decision-bearing proxies do not alter architecture or effort.
H6 fails if correction sovereignty does not change continuation or only adds burden.

13.5 Incremental-Validity Failure and Reduction

The theory should be reduced if established constructs predict persistent expansion, failure to simplify, continuation beyond thresholds, burden transfer, and delayed rollback as well as or better than the P1–P4 framework.

Possible reduction outcomes include reclassification as:

a TCO measurement extension;
a subtype of automation irony;
an organizational goal-displacement framework;
a governance diagnostic;
or a descriptive taxonomy of configuration-regime-attributable meta-work.

Reduction would narrow theoretical status without eliminating practical value.

13.6 Strong Rejection Conditions

The central theory should be rejected as a distinct socio-technical theory if a substantial replicated, cross-domain, multi-method program finds that:

P1–P4 cannot be coded reliably;
P2 cannot be separated from ordinary workflow and lifecycle cost;
P3 cannot be operationalized independently of evaluator preference;
P4 cannot be observed independently from proposed mechanisms;
corrected and persistent cases cannot be distinguished;
mechanisms do not predict burden or continuation;
the framework adds no explanatory or predictive value;
positive and negative cases cannot be distinguished prospectively;
the gate does not alter continuation or burden;
contradictory evidence is absorbed only through post hoc reinterpretation.

13.7 Primary Scope

The primary domain is professional and organizational activity involving identifiable evaluative objectives, assessable outcomes, human responsibility, configuration over time, and an institutional continuation decision.

Candidate domains include software engineering, research workflows, administration, professional decision support, compliance, knowledge management, and multi-agent operational systems.

13.8 Personal-Life Boundary

Personal-life applications are a secondary extension. Personal objectives may be plural, unstable, experiential, identity-dependent, or resistant to external measurement. Configuration itself may be intrinsically valuable.

Personal-life research therefore requires separate treatment of welfare, attention, autonomy, preference change, consent, identity, and subjective utility.

13.9 High-Stakes Safety Boundary

High verification, supervision, and governance do not establish P3 by themselves. The relevant question is whether the complete regime reduces risk and outperforms credible alternatives under the same constraints.

The theory must not encourage removal of necessary safeguards. Non-compensatory safety and legal constraints remain prior to scalar value.

13.10 Research and Experimentation Boundary

Exploration can be a legitimate objective. Negative short-term value and repeated configuration may be rational where learning is explicit, hypotheses exist, budgets and duration are bounded, outcomes inform a decision, and continuation requires reauthorization.

The boundary is crossed where learning is invoked without a research question, experiments are not compared, knowledge does not accumulate, or the label “research” protects indefinite activity.

13.11 Intrinsic Configuration Boundary

Where the declared objective is to configure, build, explore, learn, or create, configuration work may be part of the benefit. Intrinsic value should be declared and distinguished from the original adoption claim.

A system introduced to save time cannot be declared successful solely because configuration later became interesting.

13.12 Strategic Option-Value Boundary

Organizations may retain a low-current-value system to preserve future capability, resilience, regulatory readiness, or access to emerging infrastructure.

A valid option-value claim should identify the future scenario, plausibility, decision horizon, capability preserved, preservation cost, and expiration. “AI may be important someday” is insufficient.

13.13 Environmental-Instability Boundary

Persistent change is not automatically exploration lock. Reconfiguration may be necessary where regulations, threats, models, data, or objectives change. The distinction depends on necessity, evidence, proportionality, and comparison with alternatives.

13.14 Scale Boundary

Evaluation should distinguish fixed setup burden, variable meta-work, marginal benefit, and scale-dependent debt.

The relevant question is how benefit and complete meta-work scale, not merely whether output volume scales.

13.15 Technology-Maturity Boundary

The theory may prove transitional, domain-specific, or structural.

Transitional: burden fades with mature tools and standards.
Domain-specific: persistence is concentrated in weak-ground-truth, high-stakes, open-ended, or fragmented domains.
Structural: proxy pressure, goal displacement, commitment, fragmentation, and exit dependence persist despite technical maturity.

Empirical research must determine which interpretation is supported.

13.16 Non-Determinism and Normative Boundary

Agentic properties can support bounded delegation, lower coordination, provenance, automated verification, reversibility, and genuine release of human capacity.

The theory predicts conditional variation:

Agentic Properties + Specific Human and Institutional Conditions → Higher Probability of P2–P4

It does not claim that agentic AI necessarily produces the paradox.

The theory also cannot resolve every issue of rights, dignity, autonomy, justice, equality, legality, democratic legitimacy, and distribution through net-value calculation.

Its contribution is analytical visibility, not complete normative resolution.

13.17 CEP and LoopGuard-AI Boundaries

CEP may provide an optional interpretation of equilibrium-like persistence. It is not required for defining, measuring, testing, or governing the paradox and cannot protect the theory from empirical failure.

LoopGuard-AI is a candidate implementation architecture. Failure of the implementation does not by itself refute the theory, and implementation cannot validate a failed theory.

13.18 Maturity Statement

Conceptual maturity — Advanced candidate theory

The framework contains a canonical definition, P1–P4 classification, seven core concepts, six-component taxonomy, causal architecture, formal model, hypotheses, governance derivation, objections, and falsification conditions.

Measurement maturity — Preliminary to intermediate

Coding rules and candidate metrics exist, but reliability, construct validity, thresholds, and valuation rules remain unvalidated.

Empirical maturity — Unvalidated

No controlled, longitudinal, comparative, negative-case, or out-of-sample program has established the theory.

Formal maturity — Conceptual formalization

The equations preserve distinctions and define valid comparison conditions. They are not estimated models, predictive laws, or calibrated decision functions.

Governance maturity — Candidate architecture

The gate is theoretically derived but unvalidated. H6 remains open.

Publication maturity — Final publication version

The expanded academic–professional theory has completed integration and final publication QA. Publication does not alter its empirical status: the theory remains unvalidated.

13.19 Permitted Public Claims

The article may claim that it develops and defines the AI Configuration Paradox, proposes a unified socio-technical research object, integrates mechanisms under agentic conditions, introduces P1–P4 and the meta-work taxonomy, presents hypotheses, and derives a candidate governance architecture.

It should not claim demonstrated prevalence, general productivity reduction, validated constructs, established causality, universal thresholds, validated LoopGuard-AI, or proven governance effectiveness.

13.20 Revision Logic

Evidence may lead to:

CONFIRM — reliable, discriminant, predictive, and incrementally useful;
NARROW — limited to specified systems or domains;
SIMPLIFY — remove mechanisms or distinctions that add no value;
RECLASSIFY — treat the framework as an extension or diagnostic;
REJECT — classification is unreliable and incremental validity absent.

The current status is:

an advanced candidate socio-technical theory with substantial conceptual development, preliminary measurement architecture, explicit falsification conditions, and no completed empirical validation.

14. Conclusion: When the Improvement Mechanism Becomes the Work

Agentic AI changes the form of human participation before it necessarily changes its total magnitude.

A system may reduce direct execution while increasing specification, orchestration, supervision, verification, governance, correction, and configuration-debt repair.

This transfer does not by itself constitute failure. Human meta-work may be the rational condition for greater quality, faster completion, lower risk, expanded capability, accessibility, or strategically valuable learning.

The theory begins where a simpler automation narrative ends. The relevant question is not merely how much work the AI system performs. It is what complete human and institutional regime is required to make the system useful, acceptable, correctable, and governable—and whether that regime remains subordinate to the evaluative objective that justified it.

The AI Configuration Paradox is the conjunction:

AICPd,a ∣ b,T ⇔ P1d,T ∧ P2a ∣ b,T ∧ P3d,a ∣ b,T ∧ P4a,T

Without P1 there is no external purpose against which displacement can be judged. Without P2 the relevant burden has not been attributed to the AI configuration regime. Without P3 high burden may remain proportionate to greater benefit. Without P4 material displacement may represent a correctable episode rather than a persistent paradoxical regime.

The central reversal is:

Automation of Execution → Expansion of Meta-Work → Potential Displacement of the Evaluative Objective

The reversal becomes institutionally persistent where the route from evidence to consequence is broken:

Signal → Threshold → Authority → Consequence

The paradox therefore concerns not only workload but correction sovereignty.

14.1 Why Agentic AI Matters

The classical mechanisms predate AI. Agentic AI matters because it provides an integrative substrate through which formalization, goal displacement, bounded search, proxy capture, commitment, exploration imbalance, fragmented responsibility, and loss of intervention capacity may become easier to instantiate, faster to modify, more recursive, more observable, more distributed, and harder to exit.

The relation remains conditional, not deterministic.

14.2 Complete Burden, Baselines, and Distribution

Human meta-work is represented as:

M(T)=(C,O,S,V,G,Dr)

The vector remains disaggregated before valuation because efficiency in one component can conceal burden elsewhere.

The same architecture may appear favorable against direct human execution and unfavorable against a bounded AI assistant. The relevant decision is often not AI versus no AI, but current configuration versus a credible simpler alternative.

Aggregate benefit is also incomplete where benefit, labor, risk, knowledge, and authority are separated. Configuration-burden substitution prevents hidden labor from being excluded by the accounting boundary of the principal beneficiary.

Continuation because exit has become expensive must remain distinct from continuation because value remains superior.

14.3 From Diagnosis to Governance

The four gate states are SHIP, RESTRICT, HOLD, and ROLLBACK.

The gate does not seek maximal oversight. It seeks a proportionate and operative relation between evidence and system state.

Review is not corrective unless it can produce an enforceable consequence. At the same time, the governance layer must be subject to its own burden accounting, metric minimality, budgets, expiration, and simplification authority.

14.4 Empirical Burden and Reduction Test

The theory should not be accepted because its reversal appears plausible. It must show that P1–P4 can be coded reliably, the meta-work components can be distinguished, attribution and materiality can be operationalized, P4 can be observed independently, and the framework adds explanatory or predictive value beyond adjacent constructs.

If total cost of ownership, workload, technical debt, automation irony, task complexity, sunk cost, inertia, and authority fragmentation explain the same cases equally well, the theory should be narrowed, simplified, reclassified, or rejected.

14.5 Practical Decision Rule

An AI system cannot be said to improve human work or life merely because it automates execution. Its complete incremental human burden must remain instrumentally subordinate to the evaluative objective and proportionate to the benefit it creates.

Where common valuation exists:

Bd,a ∣ b*(T)>Δ Ma ∣ b*(T)

may support continuation, subject to eligibility constraints, materiality, distribution, uncertainty, and exit conditions.

Where scalar valuation is invalid, the decision should use thresholds, dominance, multi-criteria comparison, and non-compensatory constraints.

14.6 Final Proposition

A system introduced as a means of improving an identifiable evaluative objective becomes paradoxical when the higher-order human work required to configure, coordinate, supervise, verify, govern, and repair it materially displaces that objective, while the regime lacks an effective path for the evidence of displacement to alter its continuation.

The theory’s deepest implication is that the distinction between work performed by the system and work required because of the system must become central to evaluating agentic automation.

Without that distinction, automated execution can be reported as improvement while the complete human regime becomes more demanding. Without a credible baseline, sophistication can be mistaken for comparative value. Without actor-level accounting, benefit can be produced through hidden burden substitution. Without stopping authority, evidence can remain descriptive while continuation becomes automatic.

The improvement mechanism then ceases to remain a means.

It becomes the work.

Glossary

AI Configuration Paradox — The P1–P4 condition in which an AI configuration regime generates substantial human meta-work, materially displaces an evaluative objective or benefit, and persists without bounded learning or an effective correction path.

Configuration Displacement — A condition in which the human and institutional work required to maintain the improvement mechanism materially displaces the evaluative activity or benefit the mechanism was introduced to support.

Configuration Progress Substitution — The substitution of visible configuration activity, architecture sophistication, or internal metrics for evidence of external progress.

Terminal Configuration Failure — Failure of a configuration regime to convert relevant burden or displacement evidence into terminal acceptance, simplification, restriction, suspension, or rollback.

Configuration Equilibrium — A persistent regime in which incentives, dependencies, metrics, and authority relations stabilize continuation even when simplification or exit may be externally preferable.

Configuration-Burden Substitution — Transfer of substantial configuration or correction burden to actors whose labor is hidden or undervalued and whose correction authority is inadequate.

Stopping Authority — Enforceable institutional authority capable of converting a defined signal and threshold into a change in system state.

Human Meta-Work — Human work required to configure, orchestrate, supervise, verify, govern, and repair the AI configuration regime rather than directly perform the evaluative activity; empirical P2 concerns its increment relative to a credible baseline.

Bounded Learning Rationale — An experimental justification with a declared objective, hypotheses, budget, duration, evidence, stopping criteria, and reauthorization rule.

Exploration Lock — Persistent experimentation in which continuing configuration possibilities repeatedly delay transition into stable operation without bounded learning or terminal acceptance.

References

References are separated from the related RATIUM.AI corpus. External sources support historical, conceptual, technical, and empirical claims; RATIUM.AI works provide internal conceptual context only.

Bainbridge, L. (1983). Ironies of automation. Automatica, 19(6), 775–779. https://doi.org/10.1016/0005-1098%2883%2990046-8

Brynjolfsson, E. (1993). The productivity paradox of information technology. Communications of the ACM, 36(12), 66–77. https://doi.org/10.1145/163298.163309

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%2879%2990048-X

Dhanorkar, S., Passi, S., & Vorvoreanu, M. (2026). Human oversight of agentic systems in practice: Examining the oversight work, challenges, and heuristics of developers using software agents. In Proceedings of the 2026 ACM Conference on Fairness, Accountability, and Transparency (pp. 6438–6465). https://doi.org/10.1145/3805689.3812402

Ellram, L. M. (1995). Total cost of ownership: An analysis approach for purchasing. International Journal of Physical Distribution & Logistics Management, 25(8), 4–23. https://doi.org/10.1108/09600039510099928

Endsley, M. R., & Kiris, E. O. (1995). The out-of-the-loop performance problem and level of control in automation. Human Factors, 37(2), 381–394. https://doi.org/10.1518/001872095779064555

Fox, S. E., Shorey, S., Kang, E. Y., Montiel Valle, D., & Rodriguez, E. (2023). Patchwork: The hidden, human labor of AI integration within essential work. Proceedings of the ACM on Human-Computer Interaction, 7(CSCW1), Article 81, 1–20. https://doi.org/10.1145/3579514

Larsen-Ledet, I. (2026). Generative AI at work: Mundane and productive articulation. In Extended Abstracts of the 2026 CHI Conference on Human Factors in Computing Systems (Article 350, pp. 1–6). https://doi.org/10.1145/3772363.3798602

Liang, J. (2026). The novelty bottleneck: A framework for understanding human effort scaling in AI-assisted work. arXiv. https://arxiv.org/abs/2603.27438

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

Merton, R. K. (1940). Bureaucratic structure and personality. Social Forces, 18(4), 560–568. https://doi.org/10.2307/2570634

Moradi, P., Levy, K., & Cheyre, C. (2025). Pseudo-automation: How labor-offsetting technologies reconfigure roles and relationships in frontline retail work. Proceedings of the ACM on Human-Computer Interaction, 9(2), Article CSCW153, 1–21. https://doi.org/10.1145/3711051

National Institute of Standards and Technology. (2023). Artificial Intelligence Risk Management Framework (AI RMF 1.0). NIST AI 100-1. https://doi.org/10.6028/NIST.AI.100-1

National Institute of Standards and Technology. (2024). Artificial Intelligence Risk Management Framework: Generative Artificial Intelligence Profile. NIST AI 600-1. https://doi.org/10.6028/NIST.AI.600-1

National Institute of Standards and Technology. (2026a). AI Agent Standards Initiative. https://www.nist.gov/artificial-intelligence/ai-agent-standards-initiative

Organisation for Economic Co-operation and Development. (2026). The agentic AI landscape and its conceptual foundations. OECD Artificial Intelligence Papers, No. 56. https://doi.org/10.1787/396cf758-en

Parasuraman, R., & Riley, V. (1997). Humans and automation: Use, misuse, disuse, abuse. Human Factors, 39(2), 230–253. https://doi.org/10.1518/001872097778543886

Rao, A. K., Keller, A. J., Kalra, N., Steed, R., Kwegyir-Aggrey, K., Klyman, K., Staheli, D., & Bergman, A. S. (2026). Challenges to the monitoring of deployed AI systems. NIST AI 800-4. https://doi.org/10.6028/NIST.AI.800-4

Schmidt, K., & Bannon, L. (1992). Taking CSCW seriously: Supporting articulation work. Computer Supported Cooperative Work, 1(1–2), 7–40. https://doi.org/10.1007/BF00752449

Sculley, D., Holt, G., Golovin, D., Davydov, E., Phillips, T., Ebner, D., Chaudhary, V., Young, M., Crespo, J.-F., & Dennison, D. (2015). Hidden technical debt in machine learning systems. In Advances in Neural Information Processing Systems 28 (pp. 2503–2511). https://papers.neurips.cc/paper/5656-hidden-technical-debt-in-machine-learning-systems.pdf

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

Simkute, A., Tankelevitch, L., Kewenig, V., Scott, A. E., Sellen, A., & Rintel, S. (2024). Ironies of generative AI: Understanding and mitigating productivity loss in human–AI interactions. arXiv. https://arxiv.org/abs/2402.11364

Staw, B. M. (1976). Knee-deep in the Big Muddy: A study of escalating commitment to a chosen course of action. Organizational Behavior and Human Performance, 16(1), 27–44. https://doi.org/10.1016/0030-5073%2876%2990005-2

Weber, M. (1978). Economy and Society: An Outline of Interpretive Sociology (G. Roth & C. Wittich, Eds.). University of California Press. (Original work published 1922.)

Related RATIUM.AI Research

The following works provide conceptual or architectural context. They are not substitutes for external evidence:

Related Source and Reference Pages

This article belongs to the public essay layer of RATIUM.AI. For readers who want to move from this article into the broader source, technical, and orientation layers of the project, the following pages provide the relevant entry points.

Articles

The articles page gathers the public essay layer of RATIUM.AI, including arguments on stable AI governance, decision-control architecture, visible governance versus real authority, universal reason, technical competence, purpose governance, and the doctoral-scale framing of CEP.