How does agency split between humans and AI systems?

Holldack, Banh and Strobel argue that "agentic information systems" (IS) — systems that "exhibit autonomous behavior and advanced cognitive capabilities" rather than functioning as "reactive tools supporting humans" — require a new typology because they question "the long-standing assumption of human agentic primacy in IS research and practice." Taking an entity-focused view of agency, they name four archetypes along a spectrum. In the assisting archetype, "humans possess the agency primacy," with IS artifacts as "reactive tools requiring human initiation and intervention" — their example is generative tools like ChatGPT or Claude that help with text, images, or code but whose "agency remains substantially limited to human-provided tasks and scopes." In the autonomous archetype, the roles invert: "human agents become increasingly dependent on agentic IS artifacts' decision-making capabilities (e.g., autonomous trading agents)," to the point that these artifacts "can be employed to direct other agents (including humans) by prescribing tasks, tracking and evaluating overall progress and even disciplining them through rewards or replacement." In the collaborative archetype, "both entities possessing high degrees of agency" work as "teammates or colleagues" with agency becoming "a dynamically shifting property." In the hybrid archetype, dense interfaces ("AI-based augmented reality glasses or even futuristic brain-computer interfaces") produce "conjoined agency, i.e., a shared capacity to act intentionally," at which point the authors say entity-focused attribution itself breaks down and the phenomenon needs "the theoretical lens of sociomateriality."

The authors build the typology by crossing two lenses: an "entity-focused" perspective that asks which party — human or IS artifact — holds agency, and a "relational" perspective that instead asks what actions a relationship affords or constrains, making agency "a systemic phenomenon rather than an entity-based attribute." The assisting and autonomous archetypes sit at the two poles of the entity-focused view; collaborative and hybrid archetypes emerge once a weak-to-strong relational view is applied, culminating in a "strong relational perspective" under which, in networked multi-agent settings, "multi-agent systems demonstrate high collective agency even when individual entities exhibit low agency," so the system, not any one entity, becomes the right unit of analysis.

This four-archetype split sits alongside other taxonomies the library holds for the same question of where agency sits between human and machine. How do humans and technology share control over organizational routines? names an almost identical concept of "conjoined agency" for the hybrid case, but that taxonomy cross-tabulates two routine components — protocol development and action selection — where this paper instead spectrums a single entity-focused axis from assisting to autonomous before separately layering a relational axis on top. Against Does machine agency exist on a spectrum rather than binary?, the two typologies agree the tension is not binary but disagree on shape: that note's five levels form one continuous degree-of-autonomy axis, while this paper keeps assisting/autonomous as an entity-focused axis and treats collaborative/hybrid as requiring a different, relational axis altogether. The paper's own account of what happens once collective, system-level agency is in play connects to the concern in Can classical agent architectures fix agentic AI accountability? that accountable agentic AI needs formal scaffolding — here the authors make the same point from the relational side, noting the strong relational perspective "does not provide a framework for differentiating how agency is distributed across individual entities, thereby blurring responsibility boundaries."

The excerpt is conceptual, not empirical: the four archetypes are illustrated with brief, named examples (ChatGPT/Claude, unnamed "autonomous trading agents," "generative agents in software engineering," AR glasses or brain-computer interfaces) rather than demonstrated against a dataset of real deployments, and the authors themselves flag that entity-focused attribution "become increasingly inadequate" once systems reach the hybrid archetype. What the excerpt does not establish is how common each archetype is in practice, or what triggers a system's drift from one archetype to another over time; the implication the text supports is narrower — that treating "agentic IS" as a single category obscures real differences in who holds decision rights, and that the differences matter enough to warrant separate design and governance treatment. See Can AI systems delegate tasks back to humans? for the delegation mechanism the paper uses to describe movement between these archetypes.