IRAIS 2026 Proceedings

Abstract

Multi-agent AI systems divide work among specialized components that plan, retrieve evidence, make recommendations, use tools, and sometimes change later assignments. These activities can create a dense technical network without creating an organizational delegation network. IS delegation theory defines delegation through task-related rights and responsibilities between a delegator and a proxy (Baird & Maruping, 2021). Recent work shows that delegation can be collective and many-to-many (Stelmaszak et al., 2025), but it does not provide a rule for deciding when several AI assignments belong to one current delegation network. The same technical architecture may therefore be a workflow, a portfolio of separate delegations, or a distributed delegation arrangement, depending on the organizational relations that are recognized around it.

We ask: What constitutes a distributed delegation arrangement when bounded task authority under one human- or organization-originated mandate is entrusted to multiple AI role positions? We develop an analytic, constitutive theory in Gregor's (2006) Type I sense. The focal unit is the current delegation constitution for one integrated task component under an active mandate, an organizational recognition regime, and a point in time. The regime defines which positions, issuers, acts, records, linkage rules, version orders, conflict resolutions, and persistence rules count. An AI component enters this unit as an organizational role only when the regime recognizes an AI-occupied position with a persistent identity, a distinct task boundary, and an attribution rule. A credential, log, model call, or microservice identity does not create such a role by itself (Faulkner & Runde, 2019).

Assignment is the broader relation. Delegation is present only when a principal with current allocation standing confers nonempty bounded task authority together with a complete role-bound performance duty on a distinct role position. Task authority has four modes: analytical, advisory, decisional, and execution. It remains different from technical permission. A complete duty states the expected action or output, completion standard, recipient, and time condition. A human or organizational mandate origin must retain residual standing to amend, recall or terminate, reassign, or accept or reject completion. These conditions separate delegation from appointment without task-specific entrustment, authorization without a performance duty, task allocation without bounded authority, and workflow configuration without a recognized organizational conferral. Formal delegation chains can establish rooted authority provenance and bounded scope, but they do not alone establish a duty-bearing organizational role or network membership (Arop, 2026).

Authority and duty do not need to appear in one physical record. They must belong to the same recognized conferral lineage. One or more records create one delegation relation only when the recognition regime links them to the same mandate, target, task slice, scope, period, and institutional act. A material conflict with no accepted resolution produces an indeterminate result. The mandate also contains a task-integration graph. Task slices belong to one connected component when they share a completion condition or joint output, depend on one another, provide complementary inputs required for completion, or mutually constrain one another. Distributed delegation exists only when at least two distinct AI role positions are current delegation holders in the same connected task component. Several AI holders in disconnected components form a portfolio of independent delegations. Tomašev et al. (2026) describe agent roles and adaptive reallocation, but agent count, service count, or a routing topology is not sufficient for organizational network membership.

These rules yield five mutually exclusive categories for a determinate current mandate: no current AI delegation, single-AI-holder delegation, a portfolio of independent AI delegations, distributed delegation without a current AI-originated reconfiguration, and currently AI-reconfigured distributed delegation. Reconfiguration capacity, current effect, and history remain separate. Reconfiguration status may be not enabled, enabled but unused, currently enacted, or historical only. An arrangement is currently AI-reconfigured only when a valid AI-issued create, amendment, or termination still determines the current delegation constitution of a distributed task component. An unused capability is not current reconfiguration, and an AI-issued event that was later superseded remains historical only. Direct human-to-AI conferrals can therefore create a distributed network when their holders contribute to one integrated task component; AI-to-AI redelegation is not required for network membership.

A supplier-onboarding system illustrates the classification. Components for document intake, screening, recommendation, and record execution may form a fixed workflow when no organizational conferral exists. Exactly one current recognized AI holder under the mandate creates single-holder delegation. Two or more distinct AI roles with valid conferrals form distributed delegation only when their task slices are connected within the same supplier outcome. If a current AI holder uses a covering authorization to issue a valid replacement-role conferral whose effect remains current, the arrangement becomes currently AI-reconfigured. Applying the theory requires policies, role specifications, accepted records, linkage rules, current validity, and evidence of task integration, not architecture diagrams alone. The account changes comparison in three ways: delegation is not routing, a network requires integrated holders rather than multiple labels, and current constitution is not event history. It does not predict performance, trust, control, or accountability outcomes. It provides a basis for comparing organizational arrangements while keeping technical permission, oversight, and answerability visible as separate relations.

References

Arop, J. (2026). Accountability chains: A formal specification for AI agent delegation [Preprint]. ScienceOpen Preprints. https://doi.org/10.14293/PR2199.004019.v1

Baird, A., & Maruping, L. M. (2021). The next generation of research on IS use: A theoretical framework of delegation to and from agentic IS artifacts. MIS Quarterly, 45(1), 315–341. https://doi.org/10.25300/MISQ/2021/15882

Faulkner, P., & Runde, J. (2019). Theorizing the digital object. MIS Quarterly, 43(4), 1279–1302. https://doi.org/10.25300/MISQ/2019/13136

Gregor, S. (2006). The nature of theory in information systems. MIS Quarterly, 30(3), 611–642.

Stelmaszak, M., Möhlmann, M., & Sørensen, C. (2025). When algorithms delegate to humans: Exploring human-algorithm interaction at Uber. MIS Quarterly, 49(1), 305–330. https://doi.org/10.25300/MISQ/2024/17911

Tomašev, N., Franklin, M., & Osindero, S. (2026). Intelligent AI delegation [Preprint]. arXiv. https://doi.org/10.48550/arXiv.2602.11865

Download

Abstract Only

Share

COinS
 
 

To view the content in your browser, please download Adobe Reader or, alternately,
you may Download the file to your hard drive.

NOTE: The latest versions of Adobe Reader do not support viewing PDF files within Firefox on Mac OS and if you are using a modern (Intel) Mac, there is no official plugin for viewing PDF files within the browser window.