I think he said that make solved only a small, relatively easy part of the problem, and he didn't want to have to work around the limitations of make and the syntax of makefiles. He also mentions that different make implementations have different idiosyncrasies that he'd have to work around. Also, I think he wanted to preserve more information about the dependency tree between the various steps, instead of having multiple tools each making their own half-assed (or completely manual) attempts at deducing that information.
