It has been a little while since I have played Power Grid, but I will take a stab at this.
In the early and mid-game the decisions that have to be made are fairly straightforward and tactical like improving power plant efficiency/capacity and expanding into reasonably priced cities. Power plant prices are generally based on their expected monetary output for the medium length duration.
Towards the end of the game, players need to more seriously plan out the how they will get to the last turn of the game by optimizing power plants, carefully selecting which cities to expand into, and budgeting and/or hoarding resources to ensure production. The decisions are more subtle and tend to interleave with the decisions of the other players in ways that can be difficult to model.
For example, buying a power plant in the end game usually has very little to do with how much money it will make. Instead, making sure you can power the target number of cities in paramount and, in some cases, making sure a rival cannot power the target number of cities is also a factor. Because of the way the power plant deck is stacked, there is a lot of information that human players can remember that a robot player cannot easily track, like the odds that the next power plant to appear will meet the needs of a player currently bidding.
Another example would be in how to expand a network in the late game. Sometimes there is more value in spending extra money to limit the options of another player even though a cheaper city is available.
I believe these are the types of changes in focus late in the game that make a simple procedural robot difficult to create. The economic engine building and expansion phase of the game is just straightforward, but how that engine is leveraged to create a victory is tricky.