Can AI systems delegate tasks back to humans?

Holldack, Banh and Strobel argue that once an information system (IS) "can make decisions independently, act in unstructured environments, and even delegate tasks to humans," the traditional IS paradigm — where humans always hold "agency primacy" and delegate downward to tools — becomes "insufficient for conceptualizing the multifaceted interaction and collaboration patterns between human and IS artifacts." They present delegation theory as "the prevailing alternative" lens, defining delegation as "the transfer of rights and responsibilities for making decisions and performing tasks to another agent." The key move is that this transfer is no longer assumed to run one way: "agentic IS artifacts enable extended forms of delegation and even reversed patterns toward humans." Within the most common form, "user-invoked delegation," humans still "transfer rights and responsibilities to IS artifacts or other agents while retaining delegation ownership," and oversight sits on a spectrum, with the most restrictive being "human-in-the-loop," where "a human agent manually approves all decisions or at least those of a critical nature made by the agentic IS artifact prior to execution."

The authors ground delegation in a requirement for "metaknowledge, i.e., a comprehensive understanding of an individual's capabilities, those of other agents, and the overarching objectives" — the reasoning being that "a single agent may lack the resources, capabilities, or knowledge necessary to achieve complex, multi-step goals independently," so productivity and flexibility come from transferring rights and responsibilities to whichever agent, human or artifact, is positioned to act. They note this builds on "HCI's mixed-initiative interaction" but goes further by "specifically foregrounding the transfer of organizational rights and responsibilities" rather than just turn-taking in a task.

This gives a mechanism for the movement the authors describe across their own typology in How does agency split between humans and AI systems? — delegation reversal is how an assisting-archetype system could, by the paper's own logic, start to resemble an autonomous one, since prescribing tasks to humans is delegation running the other direction. Against Can we design AI autonomy separately from capability?, which frames autonomy as a user-chosen design setting across five roles, this paper's human-in-the-loop description names only the most restrictive point on a comparable supervision spectrum, without naming the looser positions the excerpt cuts off before reaching. What UX principles do workplace users want in AI agents? reports that business users weight human control most heavily in practice — a preference this paper's account of reversed delegation would unsettle if agentic IS artifacts increasingly do the delegating. And Can workplace AI risks emerge from interactions alone? models risk as emerging from agent-human interaction rather than component failure, which is the same territory this paper names when it says agentic IS can track and evaluate and even discipline human agents "through rewards or replacement."

The excerpt's description of the supervision spectrum is incomplete as given — it names "human-in-the-loop" as the most restrictive point and breaks off mid-sentence before describing less restrictive positions, so the full shape of that spectrum is not established here. The claim that delegation "even" reverses toward humans is conceptual and illustrated only by pointing back to the paper's own autonomous-archetype examples (autonomous trading agents, agentic IS that discipline human agents), not by new measurement of how often or under what conditions reversal actually happens in deployed systems. The implication the text supports is that delegation, as a framework, needs to track direction as a variable rather than assuming it is always human-to-machine — but how far reversed delegation extends in practice is left open.