Edge AI Moves Computation Into the Physical World. STS Says That Changes Everything.

I read Trist and Bamforth through Sarker et al.'s sociotechnical axis last week and something bothered me that I had not noticed before. Every application of STS theory I had studied assumed the technical subsystem was somewhere visible, a server room, a software interface, a machine on the factory floor that a person could walk up to and see. Edge AI changes that. When the technical subsystem is a sensor on a hospital IV pump or a chip in a manufacturing robot that makes decisions in milliseconds without consulting anyone, the technical subsystem becomes physically distributed and harder to monitor. And STS theory, which is seventy years old, has something to say about this that most edge AI discussions are missing.

Sarker et al. (2019) classify IS research along a sociotechnical axis where Type IV, genuine sociotechnical interaction where social and technical factors shape each other, is the gold standard but accounts for only about 13% of published IS research. Edge AI is a case where the interaction is not just theoretical but literal. The technical subsystem is a physical device running inference locally. The social subsystem is a clinician, an operator, a driver whose safety depends on that device working correctly. And the two are so tightly coupled that optimizing the technical side without the social side produces exactly the same failure pattern Trist and Emery documented.

The longwall method of coal-getting replaced autonomous mining teams with specialized shifts. The technical optimization was real: continuous production, narrower tasks, machine-paced work. But the social subsystem collapsed because the miners who had coordinated informally and taken collective responsibility now worked in isolation and blamed each other for problems. Productivity fell. Trist and Bamforth (1951) documented this. Trist and Emery formalized it into sociotechnical systems theory, the core claim being that any organization contains a technical subsystem and a social subsystem, and optimizing either one alone degrades the performance of the whole. I wrote about this in detail in the joint optimization post. But the part that keeps me up is this: edge AI is the longwall method happening at the level of individual devices, and the social system might not even know the technical system has acted.

Consider healthcare. Edge AI monitors in intensive care units analyze patient vitals locally and alert clinicians when something goes wrong. The technical optimization is obvious: lower latency, no dependency on hospital network bandwidth, continuous monitoring even if the cloud goes down. But if the alerts are poorly calibrated, if the threshold for escalation is too sensitive or too specific, clinicians develop alarm fatigue. They learn to ignore alerts because most alerts are false. And then a real emergency happens and nobody responds. This is not a technical problem that a better model will solve. This is the longwall method at the device level: the technical subsystem's optimization, sending an alert for every deviation, interacts destructively with the social subsystem's optimization, conserving attention for true emergencies. Lee and See (2004) called this calibration failure: overtrust when clinicians trust alerts that are wrong, undertrust when they ignore alerts that are right. But the framing that matters here is STS. The two subsystems were designed separately. The alarm was designed to maximize sensitivity. The clinician was assumed to adapt. And the result is that performance degrades for both.

Manufacturing is the same story. Predictive maintenance sensors on factory equipment detect vibration patterns that signal impending failure. The edge device shuts down the machine to prevent catastrophic damage. That is technically optimal. But if the shutdown happens without human override, without a window for the operator to decide whether completing the current production run is worth the risk, then the social subsystem is overridden by a technical decision made in milliseconds. Baird and Maruping (2021) argued that for agentic IS systems, delegation replaces use as the central construct. The machine is not a tool anymore. It is a proxy that acts on the human's behalf. And when that proxy acts without the human's awareness, the delegation relationship breaks down because the human cannot appraise, distribute, or coordinate the decisions the machine is making.

The new failure mode edge AI introduces is invisible decisions. Cloud-based AI systems, for all their problems, leave traces. A recommendation appears on a screen. A dashboard shows what the system calculated. The round trip to the cloud creates a delay that makes the decision visible. Someone had to see it, even if only to ignore it. Edge AI processes locally and acts in real time. There is no round trip. There is no moment where the decision surfaces to human awareness. The IV pump adjusts the drip rate. The car applies the brakes. The factory robot stops the line. The decision was made and executed before any human could have seen it. I think this is the most important boundary condition for applying STS theory to edge AI: the theory assumes the social subsystem can at least observe what the technical subsystem is doing. When the technical subsystem acts invisibly, the social subsystem cannot calibrate trust, cannot learn from mistakes, cannot develop the informal coordination that Trist and Bamforth found was load-bearing for system performance.

Ackoff (1971) argued that analyzing a system by decomposing it into parts and optimizing each part independently can degrade overall system performance because system performance depends on interactions. Edge AI is a case where the parts are so tightly coupled that independent optimization is not just harmful but dangerous. The sensor, the algorithm, the network, the clinician, the patient, the hospital protocol, the regulatory environment, every one of these interacts. Optimizing the sensor's battery life without accounting for the clinician's attention budget is the kind of decomposition error Ackoff was talking about. And the failure mode is not lower productivity. It is a patient who codes while the monitor stays silent because the clinician silenced the alerts.

I am not sure edge AI makes STS theory less relevant. I think it does the opposite. It makes the theory more urgent because the technical subsystem is now physically distributed, harder to monitor, and capable of acting without human awareness. The solution path, as Bostrom and Heinen (1977) argued in the first volume of MIS Quarterly, is that system problems are sociotechnical problems. You cannot design the edge device and then add the social system around it. You have to design them together, with the same level of attention to the clinician's decision load as you give to the sensor's latency. That is joint optimization at the device level. And if you skip it, the longwall method has been warning us for seventy years what happens next.