How do humans and technology share control over organizational routines?

Murray, Rhymer and Sirmon (Academy of Management Review, 2021) argue that organizational routines are increasingly practiced by "human-nonhuman ensembles" in which the technology itself can hold intentionality, not merely assist a human who holds it. They define conjoined agency as "a shared capacity between humans and nonhumans to exercise intentionality" and split a routine into two components: protocol development ("determine rules and guidelines for what to do") and action selection ("make choices of what to do"). Crossing these two components against which actor — human or technology — controls each yields four forms: conjoined agency with assisting technologies, arresting technologies, augmenting technologies, and automating technologies.

The excerpt works through only the assisting-technology quadrant in full: a technology that "does not have the ability to develop protocols" and "does not have the ability to select actions" is "wielded by humans," illustrated by The New Yorker's editorial team using evaluation software to rank candidate cover stories while the editors themselves apply the ranking procedures and make the final call — both protocol and action stay human. For the other three quadrants the excerpt gives only the responsiveness consequence rather than the full defining criteria: a routine is "highly unresponsive" under arresting technologies, "requiring consensus among many actors before adjustments can occur"; it has "a limited level of responsiveness" under augmenting technologies, "requiring humans to contradict the technology's suggestions"; and it becomes "hyper-responsive" under automating technologies, which drive adjustment through "the technology's real-time determinations." The authors' claim is that where the locus of agency sits — human or technology, for protocol or for action — determines both how much a routine can change at a given moment and how it responds to feedback over time.

This is an organization-theory precursor to framings the library already holds for generative-AI collaboration. BCG's Do humans and AI divide labor differently in knowledge work? uses an almost identical structure — "who selects what needs to be done" and "who identifies how it gets done" as two independent axes — but asks who controls task framing versus method within a single collaboration session, whereas Murray, Rhymer and Sirmon ask which actor holds intentionality over protocol and action as a structural property of a routine's design, independent of any one episode. Rammert's Does machine agency exist on a spectrum rather than binary? levels agency by a single degree-of-autonomy axis; this 2021 paper instead cross-tabulates two routine components, so a "cooperative" machine in Rammert's scheme could be an automating technology in one routine and merely an assisting technology in another, depending on which component it governs. The automating-technology case — hyper-responsiveness via the technology's own real-time determinations — anticipates the concern in Does granting agents more autonomy undermine human oversight?: a routine that adjusts itself in real time is, by this framework's own logic, one where humans sit structurally outside the feedback loop. It also offers vocabulary for the division of labor described in How should AI agents and humans divide research tasks?, which reads as augmenting technology in protocol development (the agent proposes methods) paired with human-held action selection (humans keep final decisions).

The excerpt is a conceptual Academy of Management Review article, not an empirical study: its evidence is a single illustrative case (The New Yorker's cover-story selection) plus citations to prior case studies of non-AI technologies (PowerPoint, CAD software, cardiac surgery equipment, ERP systems), and it predates generative AI and LLM-based agents entirely — the "intentionality" attributed to arresting, augmenting, and automating technologies is theorized, not measured against any named system. The excerpt also withholds the illustrative cases and full defining criteria for the arresting, augmenting, and automating quadrants that it supplies for assisting technologies, so classifying a specific AI deployment requires inferring its quadrant from the responsiveness description alone. The taxonomy is best read as organization-design vocabulary for naming where a technology sits in a routine, not as evidence about what any particular class of AI agent does in practice.