I won’t get into the quality and/or religious aspects of using an autorouter here, but this appears to be one of the key areas where commercial offerings distinguish themselves.
In my case, I need to route a 40-pin bus operating at 200 Mhz so I probably should have tight length matching. The KiCAD “Tune Length of a Single Track” command is working great, but doing 40 lines by hand is very time consuming.
I also don’t believe that KiCAD should integrate an autorouter, because that’s a lot of work and you need experience and large budgets to get good results. I tried https://github.com/freerouting/freerouting and it can route all the connections, but then I afterwards need to manually move the wires into position to create space for KiCAD’s meandering. The source code of freerouting looks pretty messy and like data model and GUI are all mixed up, so enhancing that one won’t be easy.
But what I believe could be done with relatively little effort is to create a development platform to make it as easy as possible for others to build a great autorouter.
For that, one would need a command-line tool that can read a KiCAD PCB - or better yet a Specctra DSN - and convert that into a black/white bitmap and a JSON file with a list of routing targets.
The advantage would be that the autorouting problem is converted into image-based pathfinding. And for the latter, game developers have been building creative GPU-accelerated solutions for a while now:
https://nullprogram.com/blog/2014/06/22/
Plus having the routing problem as an image means one could “train an AI” to solve it. I put that in quotes because actually one would just let a network memorize good solutions from training examples. So the “artificial intelligence” would be more like “find the closest example board and copy&paste the routing from there”.