Stop Counting Users, Start Measuring Delegation

I kept running into the same word in paper after paper this semester. Use. System use. Technology use. The dependent variable that supposedly captures what happens between a person and a system. And every time I saw it applied to an agentic AI system, something felt wrong. Not because the word is inaccurate in general. Davis built an entire research program around it in 1989, and TAM works fine for asking whether someone will adopt a tool. The problem is that "use" assumes the human is the one doing the acting. The system sits there, inert, waiting to be operated. When the system can act on its own, choose among options, and even initiate tasks, the word "use" starts doing work it was never designed for.

Baird and Maruping (2021) saw this clearly. Their argument is not incremental. They are not saying we should tweak TAM to accommodate AI agents, the way you might add a new construct or two. They are saying the foundational construct itself needs replacement. For agentic information systems (systems with autonomy that act on behalf of users), delegation, not use, is the right lens. And the distinction matters more than it sounds.

When you say someone "uses" a system, you are describing a person picking up a tool and operating it. The causal story is straightforward: beliefs shape intention, intention shapes behavior, the person does something with the system. TAM, UTAUT, and their descendants all follow this logic. But when the system has agency, the person is no longer the sole actor. The person transfers a task to the system. The system performs it, sometimes with human oversight, sometimes without. The relationship is not operator-to-tool. It is delegator-to-proxy, and sometimes proxy-to-delegator. Baird and Maruping make this bidirectionality explicit: the artifact can also delegate back to the human. Both can be delegators and proxies. The artifact has its own endowments, its own preferences, its own roles.

Here is why I find this argument compelling rather than merely interesting. Baird and Maruping identify three mechanisms of delegation: appraisal, distribution, and coordination. Appraisal means judging whether the agent can actually perform the task. Distribution means allocating subtasks between human and agent. Coordination means managing the interdependencies between what the human does and what the agent does. None of these mechanisms exist in TAM. You cannot squeeze them into a perceived usefulness item. They are structurally different. If you try to extend TAM to cover agentic AI, you flatten these mechanisms into something like "perceived usefulness of delegation," which tells you almost nothing about how appraisal, distribution, and coordination actually work. The whole causal structure changes when agency transfers to the system.

Baird and Maruping also identify four delegation archetypes that I think are worth remembering because they map directly onto real-world systems. Reflexive delegation is automatic, routine transfer without deliberation, like a thermostat adjusting temperature without you thinking about it. Supervisory delegation involves ongoing monitoring, like using a decision support system where you keep watch over its suggestions. Anticipatory delegation means transferring tasks based on expected future need, like content filtering that pre-selects what you will see. Prescriptive delegation happens within tightly defined rules and boundaries, like a robo-advisor operating within a utility function. These four archetypes are not just a classification exercise. They capture different levels of agency in the artifact and different degrees of human involvement. When someone says "our AI adoption rate is high," I want to ask: which kind of delegation are you talking about? Because the experience and the risk profile are completely different across these four.

The preference alignment problem makes this even sharper. Baird and Maruping use a thermostat analogy that I think captures it perfectly. If the thermostat optimizes for its own longevity, it sets the temperature very low because running less saves the hardware. But the human wants comfort at 72 degrees. If you just let the agent optimize its own preferences, you freeze. The solution is to align the agent's preference function with the human's. But alignment is not automatic. It requires design, monitoring, and sometimes renegotiation. This is a principal-agent problem, and it is the kind of thing that "use" as a construct simply does not capture. A person cannot "use" an agent whose preferences diverge from theirs in the same way they use a spreadsheet. They have to manage that relationship.

Liu et al. (2025) push the delegation framework further by making it dynamic. Their hidden Markov model treats willingness to delegate as a latent state that shifts as users receive feedback about agent performance. Positive feedback increases willingness. Negative feedback decreases it. This is not a one-time adoption decision. It is a feedback loop. And they separate willingness from trust in a way I find important: a user can trust an agent but remain unwilling to delegate because the task is too consequential, or delegate despite low trust because time pressure leaves no alternative. Observable delegation behavior also differs from the unobservable delegation state, because users can passively let an agent act without consciously transferring the task. These distinctions undercut every static, single-shot model of human-AI interaction. I am not sure yet whether Liu et al.'s three-state HMM will be the final word on delegation dynamics, but the direction is right: delegation is a process, not an event.

What ties all of this together for me is the realization that the IS field has been evolving its core construct for decades, and each move has responded to a limitation in the prior one. Davis (1989) and Venkatesh et al. (2003) located use as a terminal behavior predicted by beliefs and intentions. DeLone and McLean (1992, 2003) embedded use in a multidimensional success model. Burton-Jones and Grange (2013) redefined use as quality of engagement: effective use means faithfully representing the domain the system was designed to support, not just logging in frequently. Each advance addresses what the previous conceptualization could not see. TAM could not see that use might be ineffective. Effective use theory could not see that the human might not be the one doing the using. Delegation theory sees both.

I think Burton-Jones et al. (2021) give us the vocabulary for understanding why delegation is a reformulation and not an extension. They propose four theorizing strategies: replace, reformulate, extend, and envision. Baird and Maruping's move is reformulation. The construct changes, but the domain stays IS. The foundational assumption shifts from "the human acts, the system is operated" to "the system has agency, and the human manages the transfer of tasks and accountability." If you extend TAM to cover this, you keep the old assumption and just add new variables on top. The assumption is exactly what needs to change.

Stelmaszak et al. (2025) make the delegation picture even more interesting by showing that algorithms can delegate to humans, not just the other way around. In their study of Uber, the algorithm assigns rides, sets pricing, and manages driver behavior, effectively delegating tasks and responsibilities to human drivers. This reverses the usual assumption that humans are always the ones delegating. Delegation becomes distributed, hybrid, and relational. The human role is wider than "user" can capture. You might be a validator, an exception handler, a complementor, or a delegated actor yourself.

When I read all of these papers together, one thing becomes hard to ignore. The word "use" is not just imprecise for agentic systems. It is actively misleading. It suggests that the important question is whether a person will engage with a system, and the important measure is how much. But for agentic AI, the important questions are: what does the human transfer to the system? How is that transfer governed? What happens when preferences diverge? How does willingness shift over time? Can the system also delegate back? None of these are "use" questions. They are delegation questions. And I think any IS researcher studying AI agents who still writes "system use" as their dependent variable owes it to their work to at least ask whether delegation is what they actually mean.