The 2045 Automation Midpoint and What It Means for IS Research Right Now

McKinsey's economic potential report on generative AI includes a number that does not get enough attention: the midpoint for automating half of today's work activities falls around 2045. We are sitting in 2026. That is nineteen years away. The report also estimates that 60 to 70 percent of current work activities are theoretically automatable with existing or near-existing technology, and that the total economic potential of generative AI runs from $2.6 trillion to $4.4 trillion annually across industries. Those are the numbers that get quoted. The 2045 figure is the one that should be slowing people down.

The gap between "theoretically automatable" and "automated" is where the sociotechnical work lives, and it is enormous. Theoretically automatable means technically possible under ideal conditions: the model exists, the data is available, the interface can be built. Actually automating something in an organization means something much harder. It means redesigning the process the task lives in, retraining the people whose roles change, renegotiating the authority structures that determine who has input to what decisions, building the governance mechanisms that catch and correct when the automation fails, and managing the institutional legitimacy questions that arise when machines replace judgment that used to be human. None of that follows the same timeline as model capability.

Trist and Bamforth's (1951) study of the longwall coal mining method is the clearest historical illustration of what happens when you automate the technical tasks without redesigning the social system they are embedded in. The longwall method was technically efficient. It broke coal extraction into three specialized shifts, each doing one narrow task. What it destroyed was the autonomous group structure that had made the prior system resilient, the informal coordination, the mutual support, the shared accountability. The technical optimization was real. The social collapse was also real. Productivity fell below the levels of the less technically sophisticated method because the social subsystem that had been holding the technical system together was gone. Trist and Emery formalized this into sociotechnical systems theory: technical and social subsystems are interdependent, and optimizing one in isolation degrades the performance of the whole.

The 2045 automation midpoint is a sociotechnical prediction, not just a technical one. McKinsey is not saying the models will not be capable by 2045. They may be capable of automating 50% of today's work activities well before then on a narrow technical definition. The 2045 figure reflects the reality that work is embedded in organizational routines, social relationships, institutional arrangements, and physical artifacts in ways that do not yield to pure technical automation. Removing a task from a routine does not automatically restructure the routine. It creates a gap that has to be filled, and filling it requires deliberate organizational design, change management, and often political negotiation about who does what and who holds accountability.

Feldman and Pentland (2003) gave IS researchers the tools to think about this precisely. Organizational routines have an ostensive aspect, the abstract idea of the routine as participants understand it, and a performative aspect, the specific actions taken by specific people in specific circumstances. Automation changes the performative aspect by removing certain tasks from human execution. But the ostensive aspect, the shared understanding of what the routine is for and how it works, does not change automatically. When the performative and ostensive aspects fall out of alignment, which is exactly what happens when you automate a task without redesigning the surrounding routine, you get confusion, workarounds, and failure modes that nobody anticipated because they are artifacts of the gap, not the automation itself.

The IS research contribution here is studying that gap systematically. What organizational conditions allow routines to be successfully redesigned around new automation, rather than awkwardly patched to accommodate it? What governance mechanisms are needed to catch the failure modes that emerge when automated tasks stop performing as expected? How does the social system surrounding a set of automated tasks need to change, and what determines whether those social changes actually happen? These are questions that sit at the intersection of organizational theory, IS implementation research, and technology governance, and they are under-studied relative to the attention given to technical capability.

I find the 60-70% automation figure genuinely alarming when I read it without the 2045 midpoint next to it. With the 2045 midpoint, it reads differently. It says: the technology will get there, but organizations will take nearly two decades to restructure themselves around it, and some will not restructure effectively at all. The organizations that get from 2026 to 2045 with their productivity intact will be the ones that treat the automation of tasks as a sociotechnical redesign problem, not a technology procurement problem. They will redesign routines, not just deploy software. They will govern AI failure as seriously as they govern human error. They will manage the social system changes with as much rigor as the technical changes.

The organizations that treat the 60-70% figure as a roadmap for workforce reduction without the surrounding organizational redesign will discover what the longwall miners already knew. You can eliminate the tasks. If you do not redesign the system those tasks were embedded in, you eliminate the performance too.

The nineteen years between 2026 and 2045 are not empty time. They are the organizational work that converts theoretical automation potential into actual performance change. IS research has the theories to understand that work. The question is whether the field is generating studies at the level of granularity that the organizations doing that work can actually use.