A hypothetical example would be one person spending a week or two setting up an extensive in-house Spark cluster to solve a problem that a second person in one day realizes can be solved on one machine using some clever shell scripting tricks. The former person knows a lot and may have been fully capable of the solution the second came up with, but they followed the wrong guiding principles in analyzing the problem and arrived at a solution of considerable complexity.

An alternate formulation of this would involve the second person above arriving at a new job, observing that they're using Spark on an expensive cluster to regularly perform a computation, and noticing that actually they could do the whole calculation on a single node using simpler tools.

Designing a delta-sigma modulator.

Team engineers: we need a high-speed design with advanced adders and someone who can do the difficult static timing analysis.

Expert: No, you need someone who actually understands VLSI design who can interleave time and space in an algorithm. At that point, you can basically use a single step up from stupid-simple ripple carry adders. And if you use non-overlapped clock generators you don't even have to do static timing analysis. The biggest wins are in architecture, not implementation.

Yeah, 33MHz design in 125nm means VLSI like it's 1999.

There's Juicero https://blog.bolt.io/juicero/