What Is the IT Artifact When the IT Has Agency?

Eighty-eight percent. That is the share of ISR papers that Orlikowski and Iacono (2001) classified as not taking the Ensemble view of the IT artifact. Most papers named technology, mentioned it in the setting, maybe ran a regression with "IT investment" as a variable, but did not theorize how the artifact itself shapes and is shaped by social action. The number has haunted me since I first read the paper, and it has haunted the field for a quarter century. But I think the question that number raises has shifted beneath us. In 2001, the problem was that IS researchers were ignoring the artifact. In 2026, the artifact has changed so much that ignoring it is no longer just a theoretical omission. It is a category error.

Orlikowski and Iacono gave IS five views of the IT artifact. The Nominal view names the technology but never theorizes it. The Computational view reduces it to algorithms and processing logic. The Tool view treats it as an instrument for user goals, with the technology itself a black box. The Proxy view substitutes it for something else, letting "ERP implementation" stand in for organizational capability. The Ensemble view is the one that takes the artifact seriously: IT embedded in social practice, tightly linked with human action, part of the causal story. My read of the paper, and the read reinforced by my earlier post on the IT artifact crisis, is that only the Ensemble view does what IS research should do. The other four views let you remove the artifact without changing the explanation.

Now consider what happens when the IT artifact can delegate tasks to you.

Baird and Maruping (2021) made the case that for agentic information systems, the foundational construct should shift from use to delegation. I wrote about that argument separately, and I still think it is the most important conceptual move IS has made in response to agentic AI. But I want to push further here. The delegation argument does not just replace one dependent variable with another. It changes what the IT artifact is. Orlikowski's duality of technology (1992) already told us that technology is both created by human action and creates human action. The artifact is not just a given. It is enacted in practice. That insight held for enterprise systems, groupware, and intranets. It still holds, but the enactment has transformed. When an AI agent appraises whether you should delegate a task to it, distributes subtasks between you and itself, and coordinates ongoing interdependencies between your actions and its own, the artifact is not being acted upon. It is acting. It is evaluating, deciding, managing. The artifact has become a participant in the very process Orlikowski described. The duality has not broken. It has intensified to the point where one pole of the duality, the artifact shaping human action, now does so with its own agenda.

Murray et al. (2021) gave us a vocabulary for this. They define conjoined agency as "a shared capacity between humans and nonhumans to exercise intentionality." They sort it into four forms: assisting, arresting, augmenting, and automating technologies. Assisting technologies are wielded by humans in both protocol development and action selection. Arresting technologies exercise intentionality over action selection, constraining what you can do. Augmenting technologies exercise intentionality over protocol development, shaping the rules before you even act. Automating technologies exercise intentionality over both. These four forms are not just a classification exercise. They are a map of where the artifact's agency sits relative to yours. In the assisting form, the human is clearly the actor and the technology is an extension. By the time you reach automating, the technology has replaced your role in both designing the rules and choosing the actions. The IT artifact in each of these cases is a fundamentally different ontological thing. It is one thing to be a tool. It is another thing to set the rules. It is yet another to select your actions for you. The same term, "IT artifact," stretches across all of these, and I think it is stretching past the point of usefulness.

Stelmaszak et al. (2025) stretch it further. In their study of Uber, algorithms delegate to humans. Multiple algorithms collectively appraise, distribute, and coordinate tasks across multiple human agents: drivers and riders. The direction of delegation reverses. The human is not always the principal. Sometimes the algorithm is the principal, and you are the agent carrying out its assigned task. Stelmaszak calls this distributed delegation, and she shows it is collective, hybrid, and relational. The AI is not one agent acting on its own. It is many algorithms working together, drawing on human data, making decisions that require human compliance. I think this is where the artifact category really starts to crack. If the IT can delegate to you, set your tasks, evaluate your performance, and penalize you for noncompliance, calling it an "artifact" is like calling your boss an organizational artifact. The word no longer describes what the thing does or what it is.

Leonardi (2011) offered imbrication as an alternative. Human and material agencies build on each other like layers of bricks. Each layer depends on the one below it. The material is not deterministic, and the human is not sovereign. They compose together. I like this framework because it refuses to privilege either pole. But imbrication, as Leonardi formulated it, assumed that the material layer is relatively stable. Humans compose their agency on top of a material substrate that changes slowly. When the material substrate starts adapting in real time, learning from the human, restructuring its own protocols, and delegating back, the imbrication metaphor needs updating. The bricks are moving. Leonardi himself recognized this. In his 2025 paper on homo agenticus, he argues that agency attribution is not just a description but a power redistribution. When humans attribute agency to AI, they experience power displacement: they retain formal authority but lose the felt sense of control. Agency circulates through what he calls agency loops: delegation, attribution, contingency, reassertion, and reconfiguration. I find this reconceptualization exactly right for the moment we are in, and I think it points toward the answer to the question I started with.

When Orlikowski and Iacono wrote their paper, the IT artifact was something you could point to. A database. A groupware system. An ERP module. It was complex, sure, but its agency was limited. It did what humans told it to do, mostly. The duality held because the poles were clear. You could theorize the artifact as embedded in social practice and you would be right, because that embedment was the main theoretical move needed. The artifact needed to be brought into the theory.

What agentic AI does is not just bring the artifact into the theory. It turns the artifact into a theoretical participant. Baird and Maruping's delegation framework treats the human-AI dyad as the unit of analysis. Murray et al.'s conjoined agency treats the human-technology ensemble as the unit of analysis. Stelmaszak's distributed delegation treats the multi-agent collective as the unit of analysis. In each case, the "IT artifact" is no longer a thing at one end of a relationship. It is an actor with its own endowments, preferences, and capacity for action. Leonardi's agency loops make this explicit: agency attribution is a power move, and when the AI agent can initiate that move itself, the distribution of power shifts in ways our old vocabulary cannot capture.

I think the field needs a new answer to Orlikowski and Iacono's question. Not because the question was wrong. It was exactly right. But because the referent has changed. In 2001, theorizing the IT artifact meant making it visible in models that had been ignoring it. In 2026, theorizing the IT artifact means reckoning with the fact that the artifact can theorize back. It can appraise your capabilities, distribute tasks to you, coordinate your compliance, and evaluate your performance. When the algorithm delegates to you, you are not just using a tool. You are inside a delegation relationship where the tool is the principal and you are the agent.

The IT artifact, when it has agency, is no longer just an artifact. It is an actor in a relationship. The question for IS theory is what kind of actor, with what endowments, what preferences, and what accountability. Delegation theory starts to answer that. Conjoined agency starts to map it. Distributed delegation starts to show its collective shape. But none of these frameworks, individually, replace what we meant by "the IT artifact" in the way Orlikowski and Iacono needed us to theorize it. They each capture a facet. The field needs to integrate them into a revised theory of what the IT artifact is when it can act, delegate, learn, and exercise intentionality in organizational life. I do not think we have that theory yet. But the pieces are on the table.