Actor-Network Theory: Why Your ERP System Is an Actor Too

I spent a long time assuming that "the human side" of an IS implementation was the interesting part and the technology was just there in the background doing what it was told. Then I read Callon's 1986 paper on scallops and fishermen, which has nothing to do with information systems on the surface, and something shifted. The argument he was making was that the scallops, the anchoring devices, the larvae collectors, and the fishermen all had to be enrolled into a network for the research project to work. None of them were passive. The scallops either attached to the collectors or they did not. The fishermen either cooperated with the moratorium or they did not. The collectors held the larvae or they failed to. Every entity in the network had to be brought on board, and any one of them could make the whole thing fall apart.

I immediately thought about every ERP implementation story I had ever read.

Actor-Network Theory, which comes primarily from the work of Michel Callon, Bruno Latour, and John Law in the sociology of science and technology, starts from a principle that the IS mainstream still finds uncomfortable. Humans and non-humans should be given equal analytical standing in studying how networks form and function. The technical term for any entity that acts, whether human or non-human, is "actant." A database is an actant. A workflow rule is an actant. An API endpoint is an actant. A printed report that a manager carries into a budget meeting is an actant. Not because these things have consciousness or intentions, but because they act: they shape what is possible, they constrain what humans can do, they make some paths easier and others harder.

The concept that makes this more than a philosophical position is inscription. When designers build a technology, they inscribe assumptions, values, and intended behaviors into the artifact. A form field that requires a numeric vendor code inscribes the assumption that every vendor has a code in the system. A workflow that routes purchase orders above ten thousand dollars to a director for approval inscribes an organizational hierarchy and a threshold judgment. A dashboard that shows only the metrics the designers thought were important inscribes a theory of what counts as performance. The people who built the system are not in the room anymore, but their assumptions are in there, running. Latour called this "delegating" to non-humans: the technology acts on behalf of the designers' intentions, even when those designers have moved on, retired, or been acquired.

This is what makes inscription such a useful concept for IS research. It means that every artifact carries a social history and a normative program. When an organization adopts a system, it is not acquiring a neutral tool. It is adopting a set of inscribed assumptions about how work should be organized, who should approve what, what information matters, and what workflows are legitimate. The mismatch between what is inscribed and what the adopting organization actually does is where most implementation failures live. The ERP system inscribes best practices from the vendor's previous clients, typically large manufacturers or European multinationals. The organization trying to implement it is a regional hospital or a municipal government. The inscribed assumptions do not fit. And now there is a choice: change the organization to match the inscriptions, or customize the system and fight every upgrade cycle for the rest of the system's life.

Translation is the other concept from ANT that I find indispensable for IS research. Translation in ANT means the process by which one actor aligns the interests of others with their own. Callon described four moments of translation: problematization, interessement, enrollment, and mobilization. Problematization is when an actor defines a problem in a way that makes their solution indispensable. A consulting firm that defines the organization's challenge as "fragmented data silos" and then proposes an integrated ERP is doing problematization. Interessement is when that actor gets other actors to accept their framing. Enrollment is when those actors take on the roles the translation requires. Mobilization is when everything holds together and the network acts as a stable whole.

What this gives IS researchers is a vocabulary for studying why implementations succeed or fail that is more granular than "user resistance" or "lack of top management support." Translation can break at any of the four moments. The problem definition never gains traction. The key actors refuse interessement. The enrolled actors abandon their roles under pressure. The mobilized network fragments when the vendor stops supporting the old version. Every breakdown point is visible in the translation sequence. You can map it.

ANT is controversial, and I think the criticism is worth taking seriously. The most common objection is that treating non-humans as actants is either trivially true or anthropomorphically misleading. Of course a database constraint affects what people can do. Calling it an "actant" does not add explanatory power, critics say. It just reframes the obvious. A second objection is that ANT is descriptive rather than predictive. It can tell a rich story about how a network formed, but it does not generate testable hypotheses about why networks form the way they do or what will cause them to stabilize or collapse. Law (1992) himself acknowledged that ANT is a "material semiotics" rather than a causal theory in the positivist sense.

My reading of the IS literature suggests that these criticisms do not make ANT useless. They locate its appropriate use. ANT is not the right lens if you want to test hypotheses about adoption rates or identify variance in user satisfaction scores. It is the right lens if you want to understand how a specific technology became stabilized in a specific organizational setting, what had to be enrolled to make that happen, whose interests were translated and whose were not, and why the network is fragile or resilient. Those are qualitative, process-oriented questions. They call for case study methods, ethnography, and close reading of documents and artifacts. ANT gives that kind of research a theoretical backbone.

The ERP implementation is the canonical IS example because it makes every ANT concept visible. The vendor is an actant, shaping what the system can do and inscribing its own assumptions into the software. The consultants are actants, doing translation work by defining the implementation problem in terms that serve their methodology. The legacy data is an actant, refusing to migrate cleanly and forcing decisions about data quality that nobody planned for. The approval workflow in the old system is an actant, and when the new system's workflow does not match it, people route around the system. The printed workaround that appears six months after go-live is an actant. The shadow spreadsheet that the finance team runs in parallel to "double-check the ERP numbers" is an actant. The network of the implementation includes all of these, and understanding the outcome requires following all of them.

What I find most useful in ANT for my own research thinking is the resistance to predetermined hierarchy. Most IS theories have a default assumption about what matters: user cognition, organizational structure, IT capabilities, institutional pressures. ANT says follow the actors, wherever they lead. If the printed report is doing more work than the dashboard, study the report. If the data entry constraint is shaping manager behavior more than the training program, study the constraint. You cannot decide in advance which actants matter. You have to trace the network as it forms and see what does the work.

The flip side of that flexibility is that ANT requires you to know when to stop. A network has no natural boundary. Every actor is connected to other actors that are connected to other actors. Inscription in a database traces back to a vendor's product decision that traces back to industry standards that trace back to decisions made in the 1980s. Translation work involves consultants who were trained in institutions that were funded by clients who have their own histories. ANT as a method requires choices about scope that the theory itself cannot make for you. That is not a flaw exactly. It is a feature of doing research on genuinely complex sociotechnical phenomena. But it means that an ANT study requires a researcher who knows why they are drawing the boundaries they are drawing, not just one who adopted the vocabulary.