The friction described here is not a workflow problem. It is a structural one. When data scientists and machine learning engineers operate as separate camps, with a ticketing system standing in for shared context, the person who knows the model best ends up debugging production issues they were never supposed to own. This experience is not an outlier. It is the natural outcome of treating deployment and monitoring as a handoff rather than a partnership.
The practical cost is measurable in time and trust. Every time a data scientist has to trace a serving error because the MLE lacks model context, that is a cycle of context-building that should have been unnecessary. And when the fix requires a change to the underlying stack, the delay is not days. It is months, because roadmaps move at the speed of organizational alignment, not technical necessity. The request is not for more documentation or a better dashboard. They are asking for a seat at the table where serving decisions are made. That is a reasonable request, and it is the one that most large organizations fail to honor.
What worked in the small company was not magic. It was proximity. When the author handed off a trained model and stayed available for feature-level advice, the MLEs had what they needed to build and maintain serving pipelines without constant escalation. The division of labor was clean because the context was shared. No one was debugging someone else's blind spot. That model scaled in the place where it mattered most: the relationship between the people who build and the people who operate.
The takeaway is not that client-provider setups are inherently broken. It is that they break when they remove context instead of managing it. If you are a data scientist constantly fighting production fires, the fix is not to become an MLE. It is to push for a structure where the people serving your model are embedded enough to understand it. If that means pushing back on a rigid pod structure or demanding shared ownership of the serving path, do it. Because every month you spend debugging someone else's lack of context is a month you are not spending on the work only you can do.