- The paper develops a theory showing that agentic AI can modularize user interfaces and workflows without transferring accountability, especially when verification costs are high and responsibility is difficult to shift.
- It introduces accountability assets, such as evidence trails, review workflows, signoff authority, and governed rules, alongside rule debt—the future governance burden created when formal decision rules move into unmanaged agent instructions.
- Its three boundary strategies—component, integrated, and dual-track—help organizations decide which AI capabilities to outsource while retaining accountability infrastructure for audit, clinical, legal, and procurement decisions.
The core argument
This paper, by Hydari and Muzaffar (2605.23179), develops a conceptual theory of organizational boundaries in agentic AI ecosystems. Its central claim is that the boundary-shifting effects of agentic orchestration are conditional on a dimension that existing theory does not capture: accountability. The authors distinguish between the interface boundary—the location of user interaction, task initiation, and workflow routing—and the accountability boundary—the location of responsibility, evidence, review, signoff, and post hoc defensibility. Agentic orchestrators can move the first without moving the second. This divergence explains why some AI-enabled capabilities resist disaggregation even when their technical interfaces become fully modular.
The paper's motivating puzzle is direct: transaction cost logic predicts that when agentic orchestrators reduce search, invocation, sequencing, schema translation, monitoring, and adaptation costs, market assembly becomes cheaper and vertical disintegration should follow (Coase 1937; Williamson 1985). Yet in audit, clinical decision support, legal commitments, and governed procurement decisions, integration persists despite high technical decomposability. The authors' explanation is that these outputs require evidence, review, signoff, and assignable responsibility, and the assets that provide them cannot be reconstructed at runtime by a general-purpose orchestrator.
Two new constructs
The paper introduces two constructs intended as its primary theoretical contributions.
Accountability assets are defined as complementary assets whose strategic value derives from making AI-supported outputs legitimate, auditable, reviewable, and assignable to a responsible party after the fact. Examples include evidence trails, review workflows, professional signoff authority, escalation procedures, permissions, authoritative repositories, evaluation records, and governed decision rules. The construct extends Teece's complementary-assets framework in a specific direction: traditional complementary assets help commercialize an innovation ex ante; accountability assets make its outputs usable in settings where someone must defend the decision ex post. An output that is technically excellent but indefensible is not commercially equivalent to one that is technically adequate and defensible—a claim with real bite for valuation of vertical AI firms.
The authors are careful to differentiate accountability assets from adjacent constructs: they are a subset of complementary assets but not merely cospecialized assets relabeled; they are distinct from IS control mechanisms (which regulate how work is done) and from IT governance arrangements (which allocate who decides); and they are distinct from relational, reputational, compliance, and dynamic capabilities. They are socio-technical, spanning both IT artifacts (logs, lineage records, permission structures) and organizational arrangements (signoff routines, escalation rights).
Rule debt is defined as the latent governance burden created when organizational decision rules migrate from formal information systems into ungoverned agentic execution environments—informal prompts or agent instructions—producing unmanaged policy artifacts that must eventually be inventoried, versioned, tested, reconciled, and monitored for drift. The construct deliberately builds on technical debt (Cunningham 1992; Sculley et al. 2015) and on IS control theory's distinction between control configuration and control enactment (Wiener et al. 2016). Rule debt is presented as the customer-side consequence of boundary misconfiguration: adopting a component strategy for a capability whose verification cost or responsibility regime requires retained accountability infrastructure. Notably, the costs of misconfiguration fall on the party that retains responsibility—typically the customer—not on the provider or orchestrator.
Boundary strategy: three configurations
The theory organizes boundary choice around two primary dimensions at the level of the AI-enabled capability: verification cost (the cost of determining whether an output is correct, complete, defensible, and usable in accountable action) and responsibility transferability (the degree to which responsibility can be contractually, professionally, or operationally shifted across boundaries). These yield three strategies:
| Regime |
Strategy |
Logic |
| Low verification cost; transferable responsibility |
Component |
Capability sold or invoked as callable function |
| High verification cost and/or non-transferable responsibility |
Integrated |
Accountability assets retained; external computation only within controlled envelope |
| Componentizable edge functions plus accountability-bearing core |
Dual-track |
Decompose by accountability regime, not interface |
Two features of this typology deserve emphasis. First, it is stated at the capability level rather than the firm level, which resolves an otherwise puzzling empirical pattern: the same vertical firm can componentize document extraction while retaining integration around audit signoff. Second, integration does not preclude modularity. An integrated core can expose bounded modular front doors for initiation, input submission, status retrieval, or delivery—the supplementary material formalizes this through interface actions such as submit_case() and retrieve_signed_output()—provided evidence, review, lineage, signoff, and responsibility remain inside the retained boundary. Modularity of access protocol is explicitly distinguished from modularity of responsibility transfer.
Seven propositions
The propositions form a causal chain from technological change to governance cost:
- P1: Greater reductions in interface and assembly costs increase the likelihood of component strategies.
- P2: Rising verification cost weakens the P1 relationship.
- P3: Lower responsibility transferability weakens the P1 relationship independently of verifiability—a lawyer remains responsible for a filed brief even if citation retrieval was perfectly verified.
- P4: Greater accountability-asset cospecialization predicts retention of the accountability boundary, taking dual-track form where edge functions differ from the accountable core.
- P5: Under weak appropriability of the AI capability itself, control over accountability assets becomes a stronger predictor of value capture.
- P6: The greater the share of user goal formulation, invocation history, and feedback observed by the orchestrator rather than the complementor ("intent capture"), the greater the orchestrator's ability to appropriate value through reranking, substitution, or envelopment—even under open protocols such as MCP.
- P7: Pushing low-transferability, high-verification-cost capabilities into component strategies without retained accountability infrastructure accelerates customer-side rule debt accumulation.
P6 is arguably the sharpest platform-theoretic contribution. It modifies readings of open-protocol literature in which technical openness substantially reduces dependence: open protocols reduce proprietary-interface lock-in but do nothing about concentrated intent capture. A complementor tuned to one orchestrator surrenders the contextual data needed to understand why users invoked the tool, what alternatives were considered, and which outputs satisfied them—learning signals that can inform first-party substitution. This extends Boudreau's access-versus-control distinction and Zhu and Liu's envelopment findings to agentic architectures.
Illustrations and their limits
Six structured illustrations—document extraction, legal research retrieval, audit evidence and signoff, clinical decision support, legal drafting versus filed opinion, and procurement scoring versus supplier approval—are used to discipline the logic. The audit case is presented as the clearest failure of assembly-cost-only prediction: despite full technical decomposability of retrieval, anomaly detection, and workpaper drafting, the audit conclusion does not modularize because PCAOB documentation and evidence standards make the evidence trail constitutive of the conclusion's standing. The procurement case (citing Walmart's automated supplier negotiation) extends the framework beyond professional services to organizational decisions generally.
The authors state plainly that these are theoretical illustrations selected for contrast, not empirical tests, and that the propositions have not been subjected to empirical validation. This is the paper's most significant limitation, and the authors acknowledge several others: verification cost and responsibility transferability are treated as exogenous in the static analysis though endogenous over time; orchestrator strategy is taken as given rather than modeled; customer heterogeneity in willingness to pay for accountability is held implicit; and the dynamic migration predictions require longitudinal study. The paper also concedes scope limits: the framework does not extend to commodity exchange or settings where accountability for outputs is institutionally light. To support future testing, the paper provides suggested operationalizations for each construct, including contract analysis for responsibility transferability and IT service management metrics for rule debt.
Theoretical implications
Three implications for IS scholarship are developed. First, the theory specifies a structural condition under which digital innovation theory's decentralization expectation fails: technical decomposability does not entail organizational decomposability when accountability assets are cospecialized and responsibility cannot cross the technical interface. Second, it extends IS control and IT governance theory to settings where decision rules execute outside governed information systems, identifying rule debt as a governance phenomenon and yielding a design implication—registries for agent instructions, versioned rule repositories, test harnesses, behavior logs, lineage records, and exception ownership are governance architecture, not implementation details. Third, it adds accountability-asset cospecialization to the structural determinants of complementor vulnerability in platform ecosystems: complementors cospecialized with accountability assets are harder to envelop because the orchestrator must reproduce the defensibility infrastructure, not just the task output.
Conclusion
The paper offers a disciplined variance theory that redraws the analytical map for agentic AI: the unit of analysis is the capability and its accountability regime, not the firm label or the "headless" strategy as a whole. Its constructs—accountability assets and rule debt—are well-differentiated from adjacent literatures, and its propositions generate falsifiable predictions with stated scope conditions. The principal open question the paper leaves is empirical: whether the predicted boundary migrations—toward components as assurance institutions mature, and back toward integration as rule debt becomes visible—unfold at the layer-by-layer pace the theory implies, and whether providers that supply rule-governance tooling capture value as P7 suggests.