The champagne comes out on go-live day. The vendor team shakes hands, the steering committee thanks everyone for their hard work, and the project closes in the system. Six months later, the warehouse manager is dealing with picker turnover, a robot fleet nobody fully trusts, and a productivity number that still hasn’t matched the business case. Nobody planned for that phase, because officially, it doesn’t exist.
The day we call it finished
There is a particular ritual to automation projects. Months, sometimes years, of planning, vendor selection, layout redesigns, and testing culminate in a single date. The system goes live. The project team disbands. The case study gets written. And in almost every organization, this is the moment everyone agrees the work is done.
It isn’t. Research on organizational implementation, most clearly laid out in Morten Hertzum’s work on information systems adoption, describes implementation as a three-phase process: preparation, going live, and what he calls the long, improvisational process of design in use. That third phase is not an afterthought. It is where the system either earns its place in daily operations or quietly becomes the thing everyone routes around.
What “going live” actually feels like
Go live is rarely the smooth moment depicted in the project plan. It typically comes with a temporary drop in productivity, a spike in errors, and a workforce that has to abandon habits built over years in favor of procedures that don’t yet feel natural. None of this is a sign of a failed implementation. It is what going live looks like, every time, for every system. The mistake is treating it as the finish line instead of the starting gun for the phase that actually determines whether the investment pays off.
The cost nobody puts in the business case
This is where the research gets uncomfortable. A study from WU Vienna, using long-term panel data from Germany, one of the most automated economies in the world, found that higher exposure to industrial robots is associated with a measurable decline in workers’ mental well-being. The effect runs through two channels: rising fear of job loss, and something subtler, a reduced sense that the work still matters. That second effect shows up even among workers whose jobs are in no danger at all.
A Swedish research project on algorithmic management, AMOSH, found something just as concrete. Workers under heavy algorithmic oversight, meaning systems that direct, monitor, and pace their work in real time, reported significantly higher rates of psychological distress, occupational accidents, headaches, and musculoskeletal pain. The pattern was strongest among drivers, but warehouse workers were not exempt. This is not an argument against automation. It is evidence that how a system directs people matters as much as what the system does.
The robots don’t create the problem. The silence after go-live does.
What actually protects people, and performance
The encouraging part of this research is that it points to something fixable. Studies out of the Technical University of Munich and the University of Cologne on human-robot collaboration in warehouses found that motivation and performance both depend on the same three things: feedback, transparency, and autonomy. Workers in hybrid environments compare themselves to the machines next to them, and they have real preferences about which tasks they keep and which they hand off. Ignore that, and you get quiet resistance dressed up as low productivity. Build it into how the operation runs, and the same automation produces better outcomes for both the business and the people running it.
Related research on order-picking operations found that giving workers a degree of autonomy in how they execute automated workflows improved both job satisfaction and well-being, without sacrificing the efficiency gains the automation was bought to deliver. The conclusion researchers keep arriving at independently is the same one Hertzum reaches from a completely different angle: technology should assist the operator, not replace their judgment, and the system that does this well outperforms the one that doesn’t, even on pure throughput numbers.
The nuance worth keeping
It would be convenient to conclude that fear of job loss is the whole story, and that reassurance solves it. The evidence doesn’t support that. Research on AI acceptance has found that job replacement anxiety, while common, is not actually a strong predictor of whether workers accept new automation. What predicts acceptance is closer to what the Munich and Cologne researchers found: whether people feel they understand the system, have some say in how it’s used, and get something back from it. Anxiety and acceptance can coexist. Leaders who treat “calming people down” as the goal are solving the wrong problem.
What this means for the next project
If the research says anything plainly, it’s this: the human side of an automation project cannot be a workstream that closes when the technical one does. It needs an owner who is still in the building six months after go-live, when the vendor has moved on and the steering committee has stopped meeting. That owner’s job isn’t morale management. It’s making sure the system keeps explaining itself to the people running it, that there’s still a channel for feedback that goes somewhere, and that the operators retain enough control over their own work to feel like operators rather than components, the same governance question explored from the other direction in The Trust Problem: How Much Autonomy Should You Give an AI Agent?
The project plan ends at go live because that’s where the Gantt chart ends. The actual implementation doesn’t. Organizations that understand the difference are the ones whose automation investments still look good two years later.
