I was thinking about this and you wouldn’t even need AI, just some good image processing code. There’s a preview of the finished image in the corner, and you can determine the number of rows and columns (it’s a simple columnar jigsaw): only the holes and bumps vary. So if you can set the zoom of the main jigsaw space to match the preview size you could just do some kind of image matching to work out with good confidence exactly where each piece lives, for most jigsaws.

I’ve got no recent experience in image visualisation code and don’t have time to learn, but somebody might be able to do this. No AI required!